openclaw/docs/plugins/architecture.md

summary	read_when	title	sidebarTitle
Plugin internals: capability model, ownership, contracts, load pipeline, and runtime helpers	Building or debugging native OpenClaw plugins Understanding the plugin capability model or ownership boundaries Working on the plugin load pipeline or registry Implementing provider runtime hooks or channel plugins	Plugin internals	Architecture

This is the deep architecture reference for the OpenClaw plugin system. For practical guides, start with one of the focused pages below.

End-user guide for adding, enabling, and troubleshooting plugins. First-plugin tutorial with the smallest working manifest. Build a messaging channel plugin. Build a model provider plugin. Import map and registration API reference.

Public capability model

Capabilities are the public native plugin model inside OpenClaw. Native plugins can register one or more capability types:

Capability	Registration method	Example plugins
Text inference	api.registerProvider(...)	anthropic, openai
CLI inference backend	api.registerCliBackend(...)	anthropic, openai
Embeddings	api.registerEmbeddingProvider(...)	Provider-owned vector plugins
Speech	api.registerSpeechProvider(...)	elevenlabs, microsoft
Realtime transcription	api.registerRealtimeTranscriptionProvider(...)	openai
Realtime voice	api.registerRealtimeVoiceProvider(...)	google, openai
Media understanding	api.registerMediaUnderstandingProvider(...)	google, openai
Transcripts source	api.registerTranscriptSourceProvider(...)	discord, google-meet, teams-meetings, zoom-meetings
Image generation	api.registerImageGenerationProvider(...)	fal, google, openai
Music generation	api.registerMusicGenerationProvider(...)	fal, google, minimax
Video generation	api.registerVideoGenerationProvider(...)	fal, google, qwen
Web fetch	api.registerWebFetchProvider(...)	firecrawl
Web search	api.registerWebSearchProvider(...)	brave, firecrawl, google
Channel / messaging	api.registerChannel(...)	matrix, msteams
Gateway discovery	api.registerGatewayDiscoveryService(...)	bonjour
Migration	api.registerMigrationProvider(...)	migrate-claude, migrate-hermes

A plugin that registers only hooks is **hook-only**. Plugins with tools, commands, background services, or routes but no capabilities are **non-capability** plugins. Both patterns remain supported; gateway discovery is an explicit capability listed above.

External compatibility stance

The capability model is landed in core and used by bundled/native plugins today, but external plugin compatibility still needs a tighter bar than "it is exported, therefore it is frozen."

Plugin situation	Guidance
Existing external plugins	Keep hook-based integrations working; this is the compatibility baseline.
New bundled/native plugins	Prefer explicit capability registration over vendor-specific reach-ins or new hook-only designs.
External plugins adopting capability registration	Allowed, but treat capability-specific helper surfaces as evolving unless docs mark them stable.

Capability registration is the intended direction. Legacy hooks remain the safest no-breakage path for external plugins during the transition. Exported helper subpaths are not all equal — prefer narrow documented contracts over incidental helper exports.

Plugin shapes

OpenClaw classifies every loaded plugin into a shape based on its actual registration behavior (not just static metadata):

Registers exactly one capability type (for example a provider-only plugin like `arcee` or `chutes`). Registers multiple capability types (for example `openai` owns text inference, speech, media understanding, and image generation). Registers only hooks (typed or custom), no capabilities, tools, commands, or services. Registers tools, commands, services, or routes but no capabilities.

Use openclaw plugins inspect <id> to see a plugin's shape and capability breakdown. See CLI reference for details.

Compatibility signals

openclaw doctor, openclaw plugins inspect <id>, openclaw status --all, and openclaw plugins doctor surface these compatibility notices:

Signal	Meaning
config valid	Config parses fine and plugins resolve
hook-only (info)	Plugin registers only hooks; a supported path, but not migrated to capability registration yet
deprecated memory-embedding API (warn)	Non-bundled plugin uses the old memory-specific embedding provider API instead of registerEmbeddingProvider
hard error	Config is invalid or plugin failed to load

None of the advisory/warn signals break your plugin today. These signals also appear in openclaw status --all and openclaw plugins doctor.

Architecture overview

OpenClaw's plugin system has four layers:

OpenClaw finds candidate plugins from configured paths, workspace roots, global plugin roots, and bundled plugins. Discovery reads native `openclaw.plugin.json` manifests plus supported bundle manifests first. Core decides whether a discovered plugin is enabled, disabled, blocked, or selected for an exclusive slot such as memory. Native OpenClaw plugins are loaded in-process and register capabilities into a central registry. Managed instances load JavaScript through Node and compile TypeScript source when needed. Compatible bundles are normalized into registry records without importing runtime code. The rest of OpenClaw reads the registry to expose tools, channels, provider setup, hooks, HTTP routes, CLI commands, and services.

For plugin CLI specifically, root command discovery is split in two phases:

• parse-time metadata comes from registerCli(..., { descriptors: [...] })
• the real plugin CLI module can stay lazy and register on first invocation

That keeps plugin-owned CLI code inside the plugin while still letting OpenClaw reserve root command names before parsing.

The important design boundary:

• manifest/config validation should work from manifest/schema metadata without executing plugin code
• native capability discovery may load trusted plugin entry code to build a non-activating registry snapshot
• native runtime behavior comes from the plugin module's register(api) path with api.registrationMode === "full"

That split lets OpenClaw validate config, explain missing/disabled plugins, and build UI/schema hints before the full runtime is active.

Failed registrations remain visible in plugin diagnostics after their contributions are rolled back. Those records do not enter execution scopes or block healthy plugins and core context-engine admission; the loader still owns their cleanup.

Plugin metadata snapshot and lookup table

One PluginCache starts on the first plugin metadata access, including CLI preflight before Gateway startup, and fills progressively as metadata and artifacts are needed. Gateway startup retains that owner and builds its immutable PluginMetadataSnapshot. The snapshot includes plugin metadata from all configured agent workspaces, including disabled plugins, with source precedence and workspace provenance preserved. It stores the installed plugin index, manifest registry, manifest diagnostics, owner maps, and a plugin id normalizer. Package contents and lazily loaded module exports belong to other typed views of the same cache, not the snapshot itself.

Plugin-aware config validation, startup auto-enable, and Gateway plugin bootstrap consume that snapshot instead of rebuilding manifest/index metadata independently. PluginLookUpTable is derived from the same snapshot and adds the startup plugin plan for the current runtime config.

Channel setup catalogs retain the requested workspace and load-path scope, including raw plugin shadows, so trust filtering can select the appropriate installed alternative.

After startup, runtime readers reuse that inventory without filesystem discovery, manifest rereads, or freshness checks. Narrow plugin selections are in-memory views of the same inventory. Changing an account or an agent's run workspace does not invalidate it. Explicit plugin lifecycle operations prepare a new inventory for installs, updates, removals, source or manifest edits, and discovery-root changes before publishing it to the running Gateway.

Plugin reload reconciles config watcher events after asynchronous metadata preparation. An unchanged source event does not cancel the operation; newer writes or changed config, install records, or source ownership still supersede it.

Legacy session-key migration selects plugins that declare that capability before checking channel presence. Owners already eligible under migration policy do not need a channel-presence probe. Scoped selections probe persisted credentials only for their channel owners, so unrelated authentication modules stay unloaded during Doctor repairs. This credential scope does not limit environment-based presence signals: configured channels with missing plugins still produce installation and recovery hints.

Model-id normalization policies are prepared with each snapshot or narrowed view. Model selection, catalogs, and runtime normalization carry that view forward instead of rebuilding policies from its plugin list. An empty view remains authoritative and cannot inherit policies from a broader process snapshot.

Fleet model-runtime preparation captures one immutable config and authored-source view per build. Plugin argument handling, activation fingerprints, and agent lookups reuse those captured facts across agents. Dynamic model hooks still receive each agent's directory, workspace, and model registry. Preparation yields to the event loop between agents so Gateway requests can run during a large fleet build; config refresh creates a new capture. Existing installations need no configuration changes or migration.

The snapshot and lookup table keep repeated startup decisions on the fast path:

• channel ownership
• startup plugin planning
• startup plugin ids
• provider and CLI backend ownership
• setup provider, command alias, model catalog provider, and manifest contract ownership
• plugin config schema and channel config schema validation
• startup auto-enable decisions

Startup and hot replacement share one prepared registry publisher and the same inventory across all configured agent workspaces. Reload preserves workspace provenance so an unchanged linked plugin is not replaced when another plugin changes. Replacement retains unchanged plugin instances and validates candidate metadata before draining affected services and channels. Reordering object keys in equivalent metadata or settings does not replace a registration; changed values, ordered lists, and explicit reload requests still do. It stops and disposes the previous registration before registering its replacement, then publishes runtime methods and metadata together. Connected clients refresh their plugin capabilities after publication. If replacement fails before publication and cleanup succeeds, recovery registers captured previous code and configuration with fresh resource ownership. A failure after publication reports the committed generation. Plugin runtime imports remain lazy; retaining metadata does not activate every discovered plugin.

Durable final channel replies can use the admitting Gateway's current registry after an unrelated reload only when their exact channel registration is retained. The handoff also requires unchanged channel settings, shared channel defaults, and owning-plugin settings, plus channel-owned validation that preserves the admitted sender. Telegram checks its resolved bot credential and pins it for the final send; changed token-file contents, environment tokens, and SecretRef values cannot select another bot. Channels without sender preparation, new or replaced channel registrations, changed settings, and closing Gateways remain blocked. This never falls back to another Gateway or retries a send that may already have reached the provider.

Replacement reserves the affected instance even when agent turns or unfinished cleanup retain it. The prepared-model replacement gate holds new runs while already admitted runs finish using their original callbacks. New top-level retained work cannot acquire the old instance; already admitted consumers can still derive work needed to finish their runs. Detailed readiness and logs report the retained-work count and drain deadline; the final RPC receipt reports application and any drain notices. Reload from the instance's own active callback still fails during preparation to avoid waiting on itself. Idle prepared publications do not block replacement.

While admitted work drains, ordinary model catalog, auth-status, and chat metadata reads continue using the active publication. Catalog refresh work, downloaded catalog adoption, and new execution still wait. Published facts retire when resource replacement begins; an admitted-work timeout keeps those facts available without rebuilding them. An unfinished startup publication still follows its existing cancellation and recovery path.

Retained work and in-flight calls share a 60-second pre-stop budget. Retained runs finish before ordinary call admission closes. Existing calls then settle while published metadata remains readable. Sidecars release their capability consumers before the remaining finite consumers drain; memory teardown retains cleanup authority for its exact retiring provider instances until the raw cleanup settles. Replacement stops services and channels only after that handoff. Idle service and channel custody stops later with its owners. If work does not finish, the reload fails once, resumes admission, and clears the reload status; the previous plugin generation keeps serving. The deadline does not cancel agent runs or permit disposal of unfinished writes. Retry openclaw plugins reload <id> after that work finishes, or explicitly select --wait to wait until admitted work settles. The explicit wait belongs to its requesting connection: Ctrl+C or disconnect cancels the pre-publication wait and runs the same rollback. Cancellation never disposes admitted work, bypasses cleanup, or reverses a committed publication. Service shutdown and recovery retain their bounded deadlines. Successful publication logs the applied replacement and emits plugins.changed.

A provider or harness plugin load failure remains recorded in its runtime generation. It makes that plugin unavailable without superseding the generation or blocking models that use healthy plugins. Inspect the failing owner with openclaw plugins inspect <id> --runtime --json. Use openclaw doctor --fix for supported installation repairs, or fix the reported problem in plugin code, then request plugins.reload through the admin Gateway API to load the repaired plugin.

Read-only model validation, effective tool inventory, and isolated model probes acquire their own registrations when they need executable provider or harness hooks. Concurrent callers share the prepared generation, and its lifecycle disposers run after the final borrower and any unfinished preparation or catalog work settle. Cancellation does not close a registration while its callback is still running. Process shutdown revokes these registry views before joining their remaining work and disposal. Catalog reads that need only metadata do not acquire these executable registrations. Effective tool inventory prepares only configured and session-selected model facts, including captured catalogs from enabled providers; it does not refresh the full model catalog. Session selections do not change the configured model picker.

Each plugin service startup attempt owns one cleanup operation, including failed starts. Hot replacement observes candidate startup and service cleanup with five-second deadlines. Candidate startup failure rejects the replacement. Replacing a loaded plugin requires successful cleanup before another registration can acquire its resources; pending cleanup or a cleanup error can therefore reject replacement and prevent automatic recovery. A pending startup retains its resources until it finishes and its one stop operation settles. Plugin removal can report deferred cleanup; Gateway shutdown joins that work before releasing the plugin's resources. Disposal stops new registered calls while physical cleanup finishes. Service cleanup is not invoked a second time merely because an observer timed out. If a command catalog refresh also stopped unchanged channels, failed replacement resumes those healthy registrations while the failed plugin remains fenced.

A replacement that fails before publication automatically tries to restore the previous code and configuration. If its channels have already stopped, recovery observes pending service cleanup and admitted work once, for up to 60 seconds. Its observation has an independent Gateway-owned cancellation signal, so a disconnected reload requester cannot interrupt recovery. Only service-stop observer timeouts qualify for this wait; channel-stop failures, rejected service cleanup, and failed candidate cleanup still prevent recovery. The wait joins the original service stop promises, including services whose shared stop deadline was already exhausted. After cleanup and work settle successfully, recovery disposes the old registration, registers its captured code with fresh resource ownership, and restarts its channels. Manually stopped accounts stay stopped. A timeout never grants another registration ownership of an unfinished write's resources, and retries do not repeat service stop or resource disposal. Uncommitted failures also attempt to republish the prepared model runtime against the previous config, even when plugin recovery fails or cannot safely run, unless the Gateway is shutting down.

For plugin-only replacements that fail after publication, the Gateway attempts to republish model catalogs and reply dispatch against the committed configuration and plugin generation. The original plugin failure remains visible, including failed channel readiness. If model publication also fails, the operation reports both failures and asks the operator to retry the plugin reload or restart the Gateway. Shutdown or Gateway replacement cancels this recovery; mixed config changes with unfinished runtime work retain their existing restart recovery.

When recovery cannot safely finish, the operation settles as failed and releases its channel reload pauses. Channel health reads retain captured account facts without invoking unavailable plugin code; healthy registrations can restart normally. Detailed readiness reports failing: ["plugin-reload"] with the affected plugin IDs and an actionable recovery reason, instead of leaving an indefinite "reloading" state. Retry openclaw plugins reload <id> after admitted work and pending cleanup settle, or restart the Gateway. Permanent cleanup failures still prevent another registration from acquiring those resources.

A later reload can retry after pending cleanup completes successfully, without capturing code from the disposed instance again. If draining fails after services or channels stop but before disposal starts, the quiesced registration retains its original loader for the next reload's recovery capture; ordinary plugin calls remain closed when recovery fails. Rejected replacements and failed recovery registrations retain the same cleanup barrier. Recovery preserves existing error records as diagnostics without executing their code. It never substitutes current package files for an already retired registration whose captured source has been released.

Missing captured source files produce a replacement warning instead of preventing a fresh plugin load. Recovery snapshots must contain every previously captured input, including companion files. If replacement then fails, recovery restores only plugins with available captured code; it does not substitute changed installed files for a missing snapshot. Existing call-drain and resource-cleanup requirements still apply.

Gateway shutdown also joins actual harness, MCP, LSP, embedding, and media cleanup after their initial grace periods. When clearing the active registry, plugin host cleanup can advance to later hooks after a timeout, but registry resets and shared database closure wait for its actual completion. These waits preserve resources for cleanup; they do not restore a retired plugin's runtime authority.

Shutdown closes admission before waiting for config reloads to settle. Those reloads can still defer cleanup held by existing consumers. Final Gateway close releases those consumers, joins retained cleanup, and reports its failures before another Gateway can start.

Gateway config reload and startup metadata persistence use process-bound plugin lifecycle leases. After a forced stop, the next process can reclaim a lease whose recorded same-host process identity is proven dead. Package installation leases retain their expiry protection because installer subprocesses can outlive their parent. A contended lease logs its holder and observed expiry while waiting; older leases without process identity must still expire before startup can safely refresh the plugin index. This uses the existing state schema and requires no update migration.

Executable CLI cleanup reports each disposer that exceeds five seconds and proceeds with later cleanup without canceling the pending work. On macOS with Node's system CA support enabled, automatic exit after command completion waits for this pending cleanup to finish. Explicit command exit requests and the update exit watchdog retain their bounded behavior.

Standalone plugin and Codex supervision MCP stdio services retain their discovered registrations through accepted tool work, harness cleanup, and nested SDK provider lookups. Terminal shutdown cancels and joins handlers before releasing these registrations and awaiting their resource disposers. Transport-close and registration-disposal failures reach the serving caller. Programmatic servers created from supplied tools leave those resources with the caller; closing and reconnecting the same server does not dispose them.

Hot registry publication does not wait for retired host cleanup; terminal shutdown also joins cleanup already started by earlier registry replacements before resetting shared state. Activation and rollback apply only to their captured registry version. An activation superseded by a lifecycle callback reports an error.

The cache rule is documented in Plugin architecture internals: Gateway retains one cache generation, while explicit management operations use isolated generations of the same cache. There are no wall-clock TTLs for Gateway metadata.

Install, update, registry refresh, and doctor flows may read fresh package metadata to validate their changes. A management snapshot or installed-index write alone does not replace the running Gateway's inventory: the Gateway lifecycle owner must prepare and publish it. Runtime flows use their selected snapshot or lookup table instead of falling back to cold management paths.

Runtime instance and source lifetime

Inference verification fingerprints the artifact selected by the runtime's source/build preference, including explicit bundled source overrides. Loading an unchanged plugin preserves that proof; changing its selected runtime files invalidates it.

A managed runtime instance owns its module results, registered callables, and runtime-store slots. Non-bundled instances also own a captured source artifact. Package plugins capture their package inputs when the instance is created. Standalone files capture their entry and statically known inputs without copying the surrounding workspace. Bundled runtime and setup modules, including TypeScript source entries, share the host's code identity; each inventory still owns its registered callbacks and cleanup. Loading edited bundled code requires a Gateway restart; rebuild first when the installation loads compiled output. Conditional package aliases retain their package metadata, and native Node conditions, including module-sync, select the target from that captured metadata. Source inspection uses the same synchronous-module condition without evaluating plugin code. Captured source retains the difference between authored imports and require calls, so Bun's compiler resolution previews do not acquire a deferred dependency before its first call. Missing selected targets remain absent for that captured generation. Legacy packages without an exports map also admit their existing main or index entry without executing unselected code. Native entries reuse the recorded admission below. The selected package's remaining body is captured before execution. Dependency links retain existing nested installation locations. Dependencies installed beside a package remain siblings in the capture, including optional platform packages whose native assets are read through relative filesystem paths. Other ancestor dependencies link at the captured package root. Capture does not add node_modules beside individual source files, so native-addon loaders can still locate their package root and its build assets.

Native artifacts are admitted with their owning package directory, preserving the binary's package-relative path and declared companion library dependencies. Artifacts without an admitted package root retain their containing directory. Installer-owned directories use hardlinks or an existing retained-directory reference; files inspected by plugin safety checks keep independent copies. Mutable source trees retain one private directory snapshot per admitted identity, preserving old binary and companion bytes through in-place edits. Files in this namespace are prepared at admission; module execution remains on demand. Registrations share admission facts without sharing their runtime authority. Private Doctor inspections keep their native admission facts separate from the operator's state. Their temporary captures never become deferred writes to the installed index after inspection ends; ordinary deferred writes retain their original state directory. When native packages share a dependency, admission reconciles identities only for its own hardlinks, even when the filesystem's ctime has not advanced. Recorded digests are checked against installed bytes before promotion, including companions previously captured as independent copies. Unchanged companions remain valid during Doctor and reload; source content checks still reject edits. When file symlinks are unavailable, a generation can use hardlinks only if its directory preserves every captured companion and the selected host SDK. Otherwise that plugin reports a load error asking for file symlink support; the update continues with the existing plugin-failure warning behavior. Within a capture, admission checks each immutable namespace and companion-directory mapping once. Preparing more modules reuses those facts and checks newly admitted placements. Replacement captures, host selection, and recovery copies validate again, so Doctor and Gateway preparation avoid repeated walks without reusing another capture's verdict. The existing installed-index SQLite payload records directory membership, device, inode, mode, size, mtime, and ctime identities, SHA-256 digests, and the initial generation receipt. Unchanged warm startup reuses those facts. Added, removed, or changed companions require admission again. If a recorded capture directory is missing, fresh admission reads the installed package without promoting the missing capture's identities. This also applies to post-update Doctor with older updaters. Ctime-only uncertainty is resolved with a bounded rehash, including ordinary companion files whose inodes another capture retains or releases. Legacy reload receipts keep their framed raw-byte value, so a changed receipt still requires streaming its native payloads. Read-only inspection never publishes its native admissions during later settlement. Deferred publication retains the original database authority and state directory. Doctor's private checks keep captured code in profile-qualified temporary storage, so removing their database snapshots cannot remove an image still loaded in the process. An entirely missing retired capture can be rebuilt from its installed source; missing files in a retained or partially present namespace remain load errors. Identity reuse cannot detect an edit that preserves every recorded identity field. Source code outside an admitted native namespace is captured and verified separately. Published native captures survive ordinary scratch cleanup. Doctor maintenance removes unreferenced captures while preserving installed-index references, warm generations, and live owners. System-temp fallback captures are scoped to their state directory; captures with unknown ownership are preserved.

Each captured generation links the selected host openclaw package so Workers and child processes started from its modules can resolve the host SDK. This link does not depend on the main thread's module hooks and is recreated during recovery. Imports of resolved SDK file URLs and absolute paths keep the same host identity; they do not create a selective copy of the host package or its runtime chunks. Deferred SDK imports and import.meta.resolve() retain the generation's selected source or built host even after the active plugin cache changes. Snapshot cleanup and update source inspection do not descend through these links into the host package.

After the existing runtime, setup, or executable-discovery checks admit an entry, its instance captures imported shared files and dependency modules on demand. Relative, absolute, and file-URL imports use captured files; TypeScript dependency entries compile in their own package scope. Metadata and install inspection remain confined and do not acquire executable shared inputs.

Executable loading can follow a shared-module link by capturing its selected module separately. Cold source snapshots still reject links outside the plugin root, so deferred install batches and install-digest settlement require those inputs to be packaged as dependencies. Linked non-module resources outside the plugin root are not captured by this module-loading path.

A first-demand import() or require() can observe later source edits; captured metadata and entry bytes remain unchanged. The initial source digest covers the creation-time capture; later inputs extend explicit source-current checks without changing that digest. Invalid optional package metadata fails only when selected. Module acquisition uses the instance's current admission, and disposal closes further capture. Runtime and setup retirement use the existing five-second instance shutdown budget. If calls, retained consumers, or cleanup exceed that budget, logical retirement returns a forced-retirement diagnostic with the still-running call and consumer counts. Ordinary invocation authority closes and late successful results are refused. Physical cleanup continues asynchronously: captured files and module resolvers remain until calls, consumers, and cleanup tails actually settle. An explicitly retained consumer keeps its admitted turn and cleanup authority until its host closes or releases it. Its callbacks and late results are refused after release. Shared writable state stays owned until its cleanup finishes; late cleanup failures remain failures of the resource handoff. Web provider descriptors keep their registration identity; runtime projections bind their factories and returned tools to the selected plugin instance. Synchronous source inspection and failed capture still clean up before returning.

Default source captures live under <stateDir>/tmp/plugin-captures/<instanceId>/captures/, with a random instance ID and one owner.sqlite token holding its native lifetime lease. Gateway metadata and its source captures retain the same process-local instance; a concurrent CLI process owns a separate instance. Releasing one capture cannot retire another capture or a still-running metadata owner. Gateway metadata supplies its scheduler for hourly cleanup; executable CLI commands own maintenance through their invocation scope. Capturing source alone does not create a timer. Cleanup does not retain the first command's invocation context. The managed tmp/plugin-captures subtree is excluded from source snapshots when the state directory is inside a plugin's source directory. Recovery can still load a preserved source package from within that subtree.

Executable CLI commands retire their plugin inventory through the existing invocation resource scope on success and failure. Inventory adopted by Gateway publication remains with Gateway metadata retirement. At process exit, the capture owner synchronously retires any remaining instance that holds this process's custody, including explicit exits and CLI-handled signals. Forced termination still relies on startup reclamation. Plugin cleanup owns this capture subtree; reclamation removes captured payload before retiring the token so a partial deletion remains retryable. The token closes before its directory is removed, including on Windows.

A native library mapped in the current process retains its capture and token through process exit, even after its JavaScript module cache entry is removed. Cleanup records retained-by-loaded-module once for that capture lifetime instead of trying to unlink a loaded Windows image. Unchanged native package identities continue to share the retained payload across reloads. Native-load attempts also retain their capture when initialization throws: the native image can remain mapped after the initialization error.

Before runtime plugin loading, startup attempts receipt-aware cleanup under exclusive maintenance ownership. It can reclaim a retired, unlocked instance immediately, including unpublished native payloads retained until the previous process exited. Published native payloads still referenced by the installed index remain available. If another process holds state ownership, this opportunistic cleanup silently skips without asking the operator to stop a healthy Gateway. Other maintenance or cleanup failures produce a warning and startup continues; observational reads leave captures untouched.

Hourly cleanup also inspects this owned subtree. An instance becomes eligible after one hour. Age alone never authorizes removal: for token-bearing instances, cleanup must also acquire the existing token's exclusive native lease to prove released custody. This works after process termination or reboot without PID or boot-namespace records and preserves the same cleanup contract for shipped owner.sqlite markers. Cleanup rechecks directory and token identity before removal. Live leases, unreadable entries, symlinks, and invalid tokens preserve files. An aged instance left without a token by interrupted allocation or older partial cleanup is reclaimed after a successful rename probe. Allocation creates the token before creating payload, and disposal removes payload before its token. The token belongs to its capture instance; captures do not create a global coordination database. Removal remains asynchronous and advisory. This subtree is excluded from state backups because its captured package bytes are reconstructible.

Metadata retention does not create directories until a source capture is needed. If the state directory cannot accept captures, loading falls back to an isolated system-temporary instance and reports a warning. Normal disposal still removes that instance; automatic cleanup does not scan unrelated system-temporary roots. There is no total disk quota, and an active instance may legitimately exceed the one-hour cleanup grace period. If its payload directory is removed while the instance still holds custody, the next capture recreates that directory under the same lease. This does not restore previously deleted captured files or recreate a missing ownership directory.

Startup and hourly cleanup also reclaim tokenless openclaw-plugin-build-* and openclaw-model-catalog-* roots in the selected state's temporary directory and the current system temporary directory. Roots must be older than one hour and have no custody token. On macOS and Linux, cleanup rechecks that each legacy root belongs to the current UID immediately before its rename, preserving other users' captures even in privileged runs. Windows has no equivalent UID check, so privileged Windows cleanup keeps the age and rename-probe rules below. A complete process census that finds another OpenClaw producer preserves legacy roots. When the census is unavailable, including on Windows, cleanup uses age and a rename probe instead; sharing violations leave locked roots for a later cycle. This is best-effort cleanup of reconstructible legacy scratch, not proof that an older producer has stopped using it.

Older openclaw-plugin-build-* directories in the system temporary directory have no owner record proving whether their producer is still alive. Doctor reports tokenless openclaw-plugin-build-* and openclaw-model-catalog-* roots under the state temporary directory, ~/.openclaw/tmp even when another state directory is selected, the current system temporary directory, /tmp on POSIX hosts, and recorded managed-service TMPDIR locations. It deduplicates directory aliases and reports each capture's path and regular-file size without following links inside captures. Catalog roots include all nested openclaw-plugin-build-* trees in their reported size and removal receipt.

openclaw doctor --fix reclaims these legacy roots only while Doctor holds Gateway maintenance and a complete host process census finds no other OpenClaw producer. The rule rechecks both conditions before each removal and prints a receipt listing the paths removed and their sizes. A live sibling, unavailable census, or missing maintenance authority leaves the captures in place with an explanatory message. Captures created or changed during the current process and token-bearing captures remain untouched. Linux and macOS preserve native argument boundaries when inspecting processes. macOS can also identify native system services under another user by their kernel executable path and valid Apple platform signature; unavailable arguments for other live processes keep cleanup blocked with a reason. On hosts without a complete process census (including Windows and recognized container environments), Doctor reports legacy captures but skips their removal. For a container sharing the host's temporary directory, run maintenance on the host after stopping its OpenClaw containers. Doctor's maintenance repair remains separate from runtime reclamation: its broader inventory includes old service temporary locations the current runtime does not use. Modern captures retain their custody-token cleanup.

Configured Gateway agents share one model-catalog worker per plugin-inventory lifetime. Agent and authentication facts belong to each task; plugin registrations and captured source remain with the shared inventory. Standalone hosts that supply their own environment retain an isolated catalog worker for that environment. Provider-discovery entries use the exact selected runtime instance's captured source when it is already loaded, so discovery does not create a second copy of the same plugin package. Standalone discovery keeps its own setup lifetime. Each worker retains the current plugin registration context for each loader workspace, shared by agents with matching configuration, environment, and plugin inventory. Alternating unchanged workspaces reuse their captured source; replacing one workspace does not evict another. Node retains native ESM module graphs until worker retirement even after their capture files are removed, so actual source or configuration revisions can still retain module memory during that lifetime. Agent credentials and configured model facts travel with each request; catalog jobs do not rebuild the agent workspace. Discovery reuses the registrations already acquired by that context. The first catalog request prepares registrations for the agent's known configured and credential providers together; only the requested providers run catalog hooks. Newly observed owners extend that context without discarding earlier owners. Replacement releases them after admitted work settles. Successfully disposed registrations leave their plugin caches.

Catalog observation is passive. Inventory requests can ask the catalog owner to renew expired providers while returning its accepted rows. Chat metadata and session projections only observe publication, so a refresh cannot schedule itself through its own notifications. Selected native-model discovery has an independent acquisition owner and does not wait for provider inventory renewal. Both owners merge their results with the latest accepted counterpart before publication. Catalog workers use a 512 MiB V8 old-generation limit rather than inheriting the Gateway's default heap budget. Explicit process-wide heap flags override this limit; native and external allocations are outside it.

Catalog and authentication refresh tasks carry the host's prepared Claw consent provenance. Worker config reconstruction and provider imports consume these facts without opening or copying the shared state database. Host config publication and Doctor retain their existing provenance refresh and artifact-preserving inspection paths; stored data, schemas, and update/rollback behavior are unchanged.

Credential persistence publishes fresh shared-store ownership before credential discovery. Login and explicit auth refresh join the credential owner's publication instead of creating another catalog generation for the same change.

Model-catalog workers keep their captured plugin files in a worker-owned directory under the same managed capture instance, with custody retained by their producer. The parent removes any remaining captures after that worker exits, including cancellation and crashes. Files remain available while the worker is running, and retiring one worker does not remove another generation's captures. If the whole Gateway is killed, the existing hourly cleanup reclaims the abandoned instance only when it can acquire the existing token's exclusive native lease. The same rule covers shipped SQLite markers and works after a reboot. Age alone never releases captures. Cancellation releases compute capacity after the worker exits; terminal shutdown also waits for file cleanup. Failed file removal is reported as a cleanup warning.

Loading metadata alone does not execute every plugin, and registration remains synchronous. Synchronously loaded TypeScript entries and their synchronous TypeScript imports retain Jiti's CommonJS compilation behavior, including .mts and .mtsx entries. Node evaluates the captured output. Keep top-level await out of synchronous entrypoints; start asynchronous work through lifecycle callbacks or a later dynamic import.

Dynamic TypeScript imports preserve asynchronous CommonJS execution, including top-level await and module.exports, while source loaded from native JavaScript follows Node's module format. The first evaluation fixes that mode for the instance; resolving a module alone does not evaluate it. Source import.meta.resolve retains Jiti's optional parent URL and resolution options, including custom conditions and try. The one-argument resolver uses the source's directory and package scope. Entries loaded from captured source retain evaluation failures for their instance instead of retrying through another loader. Core-shipped JavaScript and libraries loaded outside a captured plugin instance keep their existing native/Jiti loading behavior.

The host Plugin SDK always stays on the host's native module graph, including when Jiti compiles a plugin entry. SDK aliases use canonical filesystem paths so symlinked checkouts cannot create another host owner. The running host selects source or built SDK modules; a plugin's file extension does not select a second SDK graph. Source hosts need a native TypeScript loader such as the repository's tooling preload. An SDK that cannot load natively fails instead of being evaluated again by the plugin transformer. Plugin reloads may create new instances of the plugin's private code, while existing and replacement plugins share host SDK identity and authority.

Captured packages also retain a link to the selected host installation so child workers and processes can import its public SDK. These separate isolates use the installation's normal package exports; they do not inherit the parent's source aliases or authority. Capture disposal removes the link, never the host package.

Managed TypeScript filename metadata (import.meta.url, import.meta.filename, import.meta.dirname, __filename, and __dirname) identifies the captured source so relative asset reads stay within that generation. Node executes compiled JavaScript from a separate directory; its module URLs and CommonJS cache keys can differ from the source filenames.

Node owns plugin resolution through Module.registerHooks; Bun keeps its native/Jiti loader and Bun.plugin resolver even when Module.registerHooks exists.

Bun uses its native/Jiti loader with a separate captured source artifact for each managed instance. Reload prepares fresh TypeScript entries and helpers while existing consumers retain their old instance. Disposal removes that instance's captured cache records and files without evicting its replacement or the host SDK. Native imports and Jiti imports retain their respective package conditions.

Bun needs local package import/export targets to exist before native resolution. Selective captures therefore acquire existing files matched by those declarations, including conditional branches and wildcard targets, before evaluation. Unselected source remains raw bytes; its code and TypeScript configuration are not evaluated. This can read more files at startup than Node's demand-driven capture. Other deferred imports still acquire source on first use through the instance's current admission; already prepared modules need no new acquisition.

When using Jiti's TypeScript path settings, keep the original tsconfig files and configuration dependencies available while the plugin is active. Loaded modules retain their selected path mappings; previously unvisited modules may read those configuration files on first use. Newly loaded instances select the current path settings.

Registry retirement revokes managed execution separately from physical resource release. An acquired inspection can release its execution authority while a borrower still holds the underlying registration resources; the last physical claim owns their disposal, including a cleanup work scope that remains usable after the releasing request has ended. Bare SDK provider results retain their own instance consumer, so their callbacks remain usable until the owning SDK host closes. That host joins admitted callback work before releasing consumers and resources; releasing the inspection still prevents new borrows. Gateway shutdown keeps shared dependencies until the owners that still need them have joined. These ownership rules do not make native plugins a sandbox or automatically close plugin-created resources. See Plugin lifecycle and cleanup for the plugin author's cleanup contract.

An admitted agent turn keeps its original context engine through accepted commit and engine disposal. Reload can report deferred cleanup while that turn finishes. Starting engine disposal closes its normal callbacks immediately; the engine's cleanup remains owned until it settles.

Activation planning

Activation planning is part of the control plane. Callers can ask which plugins are relevant to a concrete command, provider, channel, route, agent harness, or capability before loading broader runtime registries.

The planner keeps current manifest behavior compatible:

• activation.* fields are explicit planner hints
• providers, channels, commandAliases, setup.providers, contracts.tools, and hooks remain manifest ownership fallback
• the ids-only planner API stays available for existing callers
• the plan API reports reason labels so diagnostics can distinguish explicit hints from ownership fallback

Do not treat `activation` as a lifecycle hook or a replacement for `register(...)`. It is metadata used to narrow loading. Prefer ownership fields when they already describe the relationship; use `activation` only for extra planner hints.

Channel plugins and the shared message tool

Channel plugins do not need to register a separate send/edit/react tool for normal chat actions. OpenClaw keeps one shared message tool in core, and channel plugins own the channel-specific discovery and execution behind it.

The current boundary is:

• core owns the shared message tool host, prompt wiring, session/thread bookkeeping, and execution dispatch
• channel plugins own scoped action discovery, capability discovery, and any channel-specific schema fragments
• channel plugins own provider-specific session conversation grammar, such as how conversation ids encode thread ids or inherit from parent conversations
• channel plugins execute the final action through their action adapter

For channel plugins, the SDK surface is ChannelMessageActionAdapter.describeMessageTool(...). That unified discovery call lets a plugin return its visible actions, capabilities, and schema contributions together so those pieces do not drift apart.

Message action names use a deliberately closed, core-owned vocabulary so every transport can render every action. Plugins add action names through a core PR; runtime registration is intentionally unsupported.

When a channel-specific message-tool param carries a media source such as a local path or remote media URL, the plugin should also return mediaSourceParams from describeMessageTool(...). Core uses that explicit list to apply sandbox path normalization and outbound media-access hints without hardcoding plugin-owned param names. Prefer action-scoped maps there, not one channel-wide flat list, so a profile-only media param does not get normalized on unrelated actions like send.

Core passes runtime scope into that discovery step. Important fields include:

• accountId
• currentChannelId
• chatType (direct, group, or channel when the inbound route establishes it)
• currentThreadTs
• currentMessageId
• sessionKey
• sessionId
• agentId
• trusted inbound requesterSenderId

That matters for context-sensitive plugins. A channel can hide or expose message actions based on the active account, current room/thread/message, authoritative conversation type, or trusted requester identity without hardcoding channel-specific branches in the core message tool. Treat chatType as discovery scope supplied by the current inbound route, not something to infer again from an opaque channel id; it is absent when that route did not establish the conversation type.

This is why embedded-runner routing changes are still plugin work: the runner is responsible for forwarding the current chat/session identity into the plugin discovery boundary so the shared message tool exposes the right channel-owned surface for the current turn.

For channel-owned execution helpers, channel plugins should keep the execution runtime inside their own plugin modules. Core no longer owns the Discord, Slack, Telegram, or WhatsApp message-action runtimes under src/agents/tools. We do not publish separate plugin-sdk/*-action-runtime subpaths, and those plugins should import their own local runtime code directly from their plugin-owned modules.

The same boundary applies to provider-named SDK seams in general: core should not import channel-specific convenience barrels for Discord, Signal, Slack, WhatsApp, or similar plugins. If core needs a behavior, either consume the bundled plugin's own api.ts / runtime-api.ts barrel or promote the need into a narrow generic capability in the shared SDK.

Bundled plugins follow the same rule. A bundled plugin's runtime-api.ts should not re-export its own branded openclaw/plugin-sdk/<plugin-id> facade. Those branded facades remain compatibility shims for external plugins and older consumers, but bundled plugins should use local exports plus narrow generic SDK subpaths such as openclaw/plugin-sdk/channel-policy, openclaw/plugin-sdk/runtime-store, or openclaw/plugin-sdk/webhook-ingress. New code should not add plugin-id-specific SDK facades unless the compatibility boundary for an existing external ecosystem requires it.

For polls specifically, there are two execution paths:

• outbound.sendPoll is the shared baseline for channels that fit the common poll model
• actions.handleAction("poll") is the preferred path for channel-specific poll semantics or extra poll parameters

Core now defers shared poll parsing until after plugin poll dispatch declines the action, so plugin-owned poll handlers can accept channel-specific poll fields without being blocked by the generic poll parser first.

See Plugin architecture internals for the full startup sequence.

Capability ownership model

OpenClaw treats a native plugin as the ownership boundary for a company or a feature, not as a grab bag of unrelated integrations.

That means:

• a company plugin should usually own all of that company's OpenClaw-facing surfaces
• a feature plugin should usually own the full feature surface it introduces
• channels should consume shared core capabilities instead of re-implementing provider behavior ad hoc

`google` owns text inference, CLI backend, embeddings, speech, realtime voice, media understanding, image/music/video generation, and web search. `openai` owns text inference, embeddings, speech, realtime transcription, realtime voice, media understanding, image generation. `minimax` owns text inference plus media understanding, speech, image/music/video generation, and web search. `arcee` and `chutes` own text inference only; `microsoft` owns speech only. A vendor plugin can stay this narrow until it needs to cover more of that vendor's surface. `voice-call` owns call transport, tools, CLI, routes, and Twilio media-stream bridging, but consumes shared speech, realtime transcription, and realtime voice capabilities instead of importing vendor plugins directly.

The intended end state is:

• a vendor's OpenClaw-facing surface lives in one plugin even if it spans text models, speech, images, and video
• other vendors can do the same for their own surface area
• channels do not care which vendor plugin owns the provider; they consume the shared capability contract exposed by core

This is the key distinction:

• plugin = ownership boundary
• capability = core contract that multiple plugins can implement or consume

So if OpenClaw adds a new domain such as video, the first question is not "which provider should hardcode video handling?" The first question is "what is the core video capability contract?" Once that contract exists, vendor plugins can register against it and channel/feature plugins can consume it.

If the capability does not exist yet, the right move is usually:

Define the missing capability in core. Expose it through the plugin API/runtime in a typed way. Wire channels/features against that capability. Let vendor plugins register implementations.

This keeps ownership explicit while avoiding core behavior that depends on a single vendor or a one-off plugin-specific code path.

Capability layering

Use this mental model when deciding where code belongs:

Shared orchestration, policy, fallback, config merge rules, delivery semantics, and typed contracts. Vendor-specific APIs, auth, model catalogs, speech synthesis, image generation, video backends, usage endpoints. Discord/Slack/voice-call/etc. integration that consumes core capabilities and presents them on a surface.

For example, TTS follows this shape:

• core owns reply-time TTS policy, fallback order, prefs, and channel delivery
• elevenlabs, google, microsoft, and openai own synthesis implementations
• voice-call consumes the telephony TTS runtime helper

That same pattern should be preferred for future capabilities.

Multi-capability company plugin example

A company plugin should feel cohesive from the outside. If OpenClaw has shared contracts for models, speech, realtime transcription, realtime voice, media understanding, image generation, video generation, web fetch, and web search, a vendor can own all of its surfaces in one place:

import { definePluginEntry } from "openclaw/plugin-sdk/plugin-entry";
import { exampleAiMedia } from "./exampleai-media.js";

export default definePluginEntry({
  id: "exampleai",
  name: "ExampleAI",
  description: "ExampleAI models and media capabilities.",
  register(api) {
    api.registerProvider({
      id: "exampleai",
      // auth/model catalog/runtime hooks
    });

    api.registerSpeechProvider({
      id: "exampleai",
      // vendor speech config — implement the SpeechProviderPlugin interface directly
    });

    api.registerMediaUnderstandingProvider({
      id: "exampleai",
      capabilities: ["image", "audio", "video"],
      describeImage: (req) => exampleAiMedia.describeImage(req),
      transcribeAudio: (req) => exampleAiMedia.transcribeAudio(req),
      describeVideo: (req) => exampleAiMedia.describeVideo(req),
    });

    api.registerWebSearchProvider({
      id: "exampleai-search",
      createTool() {
        // Return the vendor-owned web search tool.
      },
    });
  },
});

What matters is not the exact helper names. The shape matters:

• one plugin owns the vendor surface
• core still owns the capability contracts
• provider request translation and HTTP helpers stay in the vendor plugin
• channels and feature plugins consume api.runtime.* helpers, not vendor code
• contract tests can assert that the plugin registered the capabilities it claims to own

Capability example: video understanding

OpenClaw already treats image/audio/video understanding as one shared capability. The same ownership model applies there:

Core defines the media-understanding contract. Vendor plugins register `describeImage`, `transcribeAudio`, and `describeVideo` as applicable. Channels and feature plugins consume the shared core behavior instead of wiring directly to vendor code.

That avoids baking one provider's video assumptions into core. The plugin owns the vendor surface; core owns the capability contract and fallback behavior.

Video generation already uses that same sequence: core owns the typed capability contract and runtime helper, and vendor plugins register api.registerVideoGenerationProvider(...) implementations against it.

Need a concrete rollout checklist? See Adding capabilities.

Contracts and enforcement

The plugin API surface is intentionally typed and centralized in OpenClawPluginApi. That contract defines the supported registration points and the runtime helpers a plugin may rely on.

Why this matters:

• plugin authors get one stable internal standard
• core can reject duplicate ownership such as two plugins registering the same provider id
• startup can surface actionable diagnostics for malformed registration
• contract tests can enforce bundled-plugin ownership and prevent silent drift

There are two layers of enforcement:

The plugin registry validates registrations as plugins load. Examples: duplicate provider ids, duplicate speech provider ids, and malformed registrations produce plugin diagnostics instead of undefined behavior. Bundled plugins are captured in contract registries during test runs so OpenClaw can assert ownership explicitly. Today this is used for model providers, speech providers, web search providers, and bundled registration ownership.

The practical effect is that OpenClaw knows, up front, which plugin owns which surface. That lets core and channels compose seamlessly because ownership is declared, typed, and testable rather than implicit.

What belongs in a contract

- typed - small - capability-specific - owned by core - reusable by multiple plugins - consumable by channels/features without vendor knowledge - vendor-specific policy hidden in core - one-off plugin escape hatches that bypass the registry - channel code reaching straight into a vendor implementation - ad hoc runtime objects that are not part of `OpenClawPluginApi` or `api.runtime`

When in doubt, raise the abstraction level: define the capability first, then let plugins plug into it.

Skill previews

The Control UI previews a declared plugin skill without installing or executing it. Opening a preview reads its file inventory and only the entry SKILL.md body. Selecting another file reads that body on demand; navigation never prefetches sibling contents. The file tree remains available while a selected file loads, and failed reads can be retried in place.

Catalog reads stay pinned to the selected package version and validate the inventory’s paths, sizes, and SHA-256 hashes. Installed reads stay inside the resolved plugin root, reject unsafe links, and reject a changed plugin version. Both paths retain the file-count, tree-depth, per-file, and aggregate bundle limits. The aggregate limit applies to the inventory, not selection order.

Loaded bodies and pending reads belong to one open preview and Gateway connection. Closing, reopening, navigating away, or reconnecting retires that cache. A late file response can populate its own cache entry but cannot change the selected file. New selected-file requests revalidate the inventory; only already loaded bodies are reused. Installed files edited in place become visible on reopening.

Execution model

Native OpenClaw plugins run in-process with the Gateway. They are not sandboxed. A loaded native plugin has the same process-level trust boundary as core code.

Native plugin implications: a plugin can register tools, network handlers, hooks, and services; a plugin bug can crash or destabilize the gateway; and a malicious native plugin is equivalent to arbitrary code execution inside the OpenClaw process.

Compatible bundles are safer by default because OpenClaw currently treats them as metadata/content packs. In current releases, that mostly means bundled skills.

Use allowlists and explicit install/load paths for non-bundled plugins. Treat workspace plugins as development-time code, not production defaults.

For bundled workspace package names, keep the plugin id anchored in the npm name: @openclaw/<id> by default, or an approved typed suffix such as -provider, -plugin, -speech, -sandbox, or -media-understanding when the package intentionally exposes a narrower plugin role.

**Trust note:** `plugins.allow` permits **plugin ids** to load; it does not verify source provenance or choose which same-id copy loads. An auto-discovered workspace plugin does not shadow a bundled plugin merely because that id is enabled or allowlisted.

For intentional local overrides, use plugins.load.paths to select the plugin path. Tracked global installs can also override ordinary bundled copies. On source installs, plugins built with the host retain priority over tracked globals, including when OPENCLAW_DEV_SOURCE_ROOT is unset. Matching package versions alone do not prove that a registry plugin matches a source build's SDK. See Discovery precedence for the full order.

A configured path or install record pointing to the host’s own bundled plugin tree retains bundled provenance, including source and compiled entries; a different local copy does not inherit trust from its name or allowlist entry. Checkout runners supply the development selector automatically, including for compiled plugins. See development debugging.

Bundled-plugin trust is resolved from the source snapshot — the manifest and code on disk at load time — rather than from install metadata. A corrupted or substituted install record cannot silently widen a bundled plugin's trust surface beyond what the actual source claims.

Export boundary

OpenClaw exports capabilities, not implementation convenience.

Keep capability registration public. Trim non-contract helper exports:

• bundled-plugin-specific helper subpaths
• runtime plumbing subpaths not intended as public API
• vendor-specific convenience helpers
• setup/onboarding helpers that are implementation details

Reserved bundled-plugin helper subpaths have been retired from the generated SDK export map. Keep owner-specific helpers inside the owning plugin package; promote only reusable host behavior to generic SDK contracts such as plugin-sdk/gateway-runtime, plugin-sdk/security-runtime, and injected plugin API capabilities.

Internals and reference

For the load pipeline, registry model, provider runtime hooks, Gateway HTTP routes, message tool schemas, channel target resolution, provider catalogs, context engine plugins, and the guide to adding a new capability, see Plugin architecture internals.

Related

• Building plugins
• Plugin manifest
• Plugin SDK setup
• Context engines
• Plugin Runtime - the api.runtime helpers plugins call