feat: LM Studio Integration

[rugvedS07]

Summary

• Problem: Users worldwide rely on LM Studio for running open models on their own hardware. Connecting LM Studio with OpenClaw is currently hard to do. This PR fixes it
• Why it matters: Makes it easier for users to use local models in OpenClaw using LM Studio
• What changed: Adds LM Studio in the onboarding, loads the default embedding model when required. Users can configure LM Studio to use models for inference and embedding generation.
• What did NOT change (scope boundary): Only affects changes regarding LM Studio and adds new functionality.

Change Type (select all)

• Bug fix
• Feature
• Refactor
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes For God's sake provide LM Studio in onboarding! #40107
• Closes 功能建议：原生支持 LM Studio 本地模型配置 #32898
• Closes [Bug]: #49890
• Related #

User-visible / Behavior Changes

---

---

---

Security Impact (required)

• New permissions/capabilities? No
• Secrets/tokens handling changed? Yes - stores lmstudio:default API-key auth profile / resolves LM_API_TOKEN (placeholder allowed for local no-auth setups), while real auth is enforced by the external LM Studio server.
• New/changed network calls? Yes - uses LM Studio's API, /api/v1/load and /api/v1/models on the base URL the user has configured
• Command/tool execution surface changed? No
• Data access scope changed? No
• If any Yes, explain risk + mitigation:
  Requests are made to users LM Studio instance (localhost in most cases) and the API token is used for the said instance only.

Repro + Verification

Environment

• OS: Linux
• Runtime/container: Node 22
• Model/provider: LM Studio
• Integration/channel (if any): NA
• Relevant config (redacted): Default for LM Studio

Steps

1. Run openclaw onboard
2. Select LM Studio as provider and complete on-boarding
3. To use LM Studio as an embeddings provider, use openclaw config set agents.defaults.memorySearch.provider lmstudio

Expected

• LM Studio provider is configured and populates all the models available while setup
• Uses LM Studio for LLM inference and embedding generation
• Stores the LM_API_TOKEN in the auth profile

Actual

• Works as above

Evidence

Attach at least one:

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Tests were added in

• src/agents/lmstudio-models.test.ts
• src/agents/lmstudio-runtime.test.ts
• src/agents/memory-search.test.ts
• src/agents/model-auth-markers.test.ts
• src/agents/model-auth.test.ts
• src/agents/models-config.providers.normalize-keys.test.ts
• src/commands/doctor-memory-search.test.ts
• src/commands/lmstudio-setup.test.ts
• src/commands/onboard-non-interactive.provider-auth.test.ts
• src/memory/embeddings-lmstudio.test.ts
• src/memory/manager.mistral-provider.test.ts
• src/plugins/bundled-provider-auth-env-vars.test.ts
• src/plugins/contracts/discovery.contract.test.ts

Human Verification (required)

What you personally verified (not just CI), and how:

• Verified scenarios: Using LM Studio with OpenClaw, chatting with the model, using tools and verifying if embeddings are generated when memory is updated
• Edge cases checked: LM Studio server is off/unreachable, JIT setting in LM Studio is off
• What you did not verify: NA

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Failure Recovery (if this breaks)

• How to disable/revert this change quickly: Reverting commit should remove the LM Studio as a provider
• Files/config to restore: None, LM Studio provider will not be picked up
• Known bad symptoms reviewers should watch for: NA

Risks and Mitigations

None

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 26149a30b5

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[greptile-apps]

Greptile Summary

This PR integrates LM Studio as a first-class provider in OpenClaw, covering onboarding, runtime model discovery, and embedding generation. The implementation is well-structured and thorough, following established patterns from the Ollama and Mistral integrations already in the codebase.

Key changes:

• New extensions/lmstudio/ plugin registering the provider for auth, discovery, and dynamic model preparation
• src/agents/lmstudio-models.ts — model-wire parsing, URL normalization (idempotent /v1 stripping), and JIT model loading via /api/v1/models/load
• src/agents/lmstudio-runtime.ts — auth header construction with lmstudio-local placeholder support for no-auth local setups; API key resolution chaining profile store → env → placeholder
• src/memory/embeddings-lmstudio.ts — embedding provider with 120-second warmup and graceful warmup-failure handling
• src/commands/lmstudio-setup.ts — interactive and non-interactive setup flows, plus provider discovery merging explicit config with live-discovered models
• Adds lmstudio to the EmbeddingProviderId union and the Zod memory-search schema, completing the end-to-end embedding path

Minor observations noted inline:

• The module-level cachedDynamicModels in the plugin is never invalidated within a process lifetime; if models change in LM Studio between sessions the stale entries persist until restart
• resolveEmbeddingTimeout uses remote (60s query, 2-min batch) timeouts for LM Studio, consistent with Ollama but potentially tight for large models on slower hardware after an unexpected model eviction — the initial 120-second warmup covers first-load but not re-loads

Confidence Score: 5/5

• Safe to merge — the implementation is well-tested, backward-compatible, and follows established provider patterns throughout the codebase
• The PR adds a substantial but self-contained feature with 13 dedicated test files, careful URL normalization (idempotent), proper SSRF guard usage in the embedding path, and a solid auth marker system for no-auth local setups. All changes are additive (no modifications to existing provider logic), and the PR description explicitly confirms backward compatibility. The two inline comments are non-blocking style suggestions.
• No files require special attention for merge; the two inline P2 notes on src/memory/manager-embedding-ops.ts and extensions/lmstudio/index.ts can be addressed in follow-up work

Comments Outside Diff (1)

1. src/memory/manager-embedding-ops.ts, line 621-627 (link)

   LM Studio uses remote timeouts despite being a local provider

   resolveEmbeddingTimeout classifies only "local" (node-llama-cpp) for the extended 5-minute query / 10-minute batch timeouts. LM Studio is treated as a remote provider (60-second query, 2-minute batch), which is consistent with how Ollama is handled.

   The initial warmup in createLmstudioEmbeddingProvider uses a generous 120-second timeout, so the first model load is covered. However, if LM Studio evicts the embedding model under memory pressure during a session, subsequent embedding calls go through the 60-second remote timeout, which could expire on slower hardware with large models before LM Studio can reload and infer.

   Consider whether lmstudio (and ollama) should use a longer timeout tier — or at minimum document this as a known limitation in the provider docs.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: src/memory/manager-embedding-ops.ts
   Line: 621-627
   
   Comment:
   **LM Studio uses remote timeouts despite being a local provider**
   
   `resolveEmbeddingTimeout` classifies only `"local"` (node-llama-cpp) for the extended 5-minute query / 10-minute batch timeouts. LM Studio is treated as a remote provider (60-second query, 2-minute batch), which is consistent with how Ollama is handled.
   
   The initial warmup in `createLmstudioEmbeddingProvider` uses a generous 120-second timeout, so the first model load is covered. However, if LM Studio evicts the embedding model under memory pressure during a session, subsequent embedding calls go through the 60-second remote timeout, which could expire on slower hardware with large models before LM Studio can reload and infer.
   
   Consider whether `lmstudio` (and `ollama`) should use a longer timeout tier — or at minimum document this as a known limitation in the provider docs.
   
   
   
   How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: extensions/lmstudio/index.ts
Line: 16

Comment:
**Stale dynamic model cache**

`cachedDynamicModels` is a module-level `Map` that is populated by `prepareDynamicModel` but never invalidated. If the user loads or unloads models in LM Studio after the cache was last populated, `resolveDynamicModel` will return stale (or missing) entries until the process is restarted.

This is consistent with how other providers handle dynamic models in this codebase, but it's worth considering whether a TTL or explicit invalidation signal on provider config changes would improve the experience for LM Studio users who frequently swap models between sessions.

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: src/memory/manager-embedding-ops.ts
Line: 621-627

Comment:
**LM Studio uses remote timeouts despite being a local provider**

`resolveEmbeddingTimeout` classifies only `"local"` (node-llama-cpp) for the extended 5-minute query / 10-minute batch timeouts. LM Studio is treated as a remote provider (60-second query, 2-minute batch), which is consistent with how Ollama is handled.

The initial warmup in `createLmstudioEmbeddingProvider` uses a generous 120-second timeout, so the first model load is covered. However, if LM Studio evicts the embedding model under memory pressure during a session, subsequent embedding calls go through the 60-second remote timeout, which could expire on slower hardware with large models before LM Studio can reload and infer.

Consider whether `lmstudio` (and `ollama`) should use a longer timeout tier — or at minimum document this as a known limitation in the provider docs.

```suggestion
  private resolveEmbeddingTimeout(kind: "query" | "batch"): number {
    const isLocal =
      this.provider?.id === "local" ||
      this.provider?.id === "lmstudio" ||
      this.provider?.id === "ollama";
    if (kind === "query") {
      return isLocal ? EMBEDDING_QUERY_TIMEOUT_LOCAL_MS : EMBEDDING_QUERY_TIMEOUT_REMOTE_MS;
    }
    return isLocal ? EMBEDDING_BATCH_TIMEOUT_LOCAL_MS : EMBEDDING_BATCH_TIMEOUT_REMOTE_MS;
  }
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "Feat: LM Studio Integration" | Re-trigger Greptile}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 70050f6f6f

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4f9ceea5d0

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[SeananTQ]

😀##太好了！
我太需要这个功能了，因为我是LM Studio的用户，它的界面非常直观，便于我在调试OpenClaw的时候可以快速看后台信息以及各种模型的加载情况。

但今天我更新了OpenClaw 2026.3.23 (ccfeecb)版本，我发现了一个BUG。这个BUG具体就是OpenClaw 拿不到LM Studio的上下文使用情况，具体表现是当我输入/status 之后：

🦞 OpenClaw 2026.3.23 (ccfeecb)
🧠 Model: lmstudio/qwen/qwen3.5-35b-a3b · 🔑 api-key (models.json)
📚 Context: 0/131k (0%) · 🧹 Compactions: 0
🧵 Session: agent:main:feishu:direct:ou_47c32081487e7930ea2b8cd7fcf7aa9f • updated just now
⚙️ Runtime: direct · Think: off
🪢 Queue: collect (depth 0)

其中 📚 Context: 0/131k (0%) · 🧹 Compactions: 0 是一个明显的错误

然后我又在tui里试了一下/status ：

 - agent:main:feishu:direct:ou_47c32081487e7930ea2b8cd7fcf7aa9f [direct] | 2h ago | model
 qwen/qwen3.5-35b-a3b | tokens ?/131k | flags: system, id:b3dccaa9-cf57-4652-a375-555caeaa94e8

其中| tokens ?/131k | 是一个明显的错误。

因为我是一个小白用户，不知道该如何解决这个问题，也不知道怎么提交我发现的问题，我通过搜索引擎找到了这个地方，把我所知道的东西讲出来，希望有人能看到。

[rugvedS07]

@SeananTQ thank you for flagging this, it will be fixed once this PR is merged!

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b1e0cf903a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb42863623

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb42863623

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb42863623

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3c973a698c

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 99da19495f

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 48bc6245f3

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0d63251970

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[frankekn]

Thanks for the PR. I did a clean-worktree review and local validation against a real LM Studio install. This is not merge-ready yet.

Blockers:

1. pnpm check and pnpm build are both failing on the current branch.

2. The LM Studio setup path is currently breaking the provider config type contract. In extensions/lmstudio/src/setup.ts, the header patch is built as Record<string, string | undefined>, but ModelProviderConfig.headers is typed as Record<string, SecretInput>. That mismatch is one of the reasons the branch is red.

3. The branch also introduces a larger memory-core compile break. The current changes in extensions/memory-core/src/memory/* are not type-clean right now, so this PR is not only adding LM Studio support, it is also leaving the repo in a non-buildable state.

4. I ran local LM Studio validation with a real server and model. Discovery, non-interactive onboarding, config writing, gateway startup, and model preload all worked. But the full live gateway lane still failed on tool probing for lmstudio/google/gemma-4-26b-a4b with tool probe missing nonce. So at the moment this proves the basic wiring is partially working, but it does not prove the LM Studio integration is end-to-end ready for the OpenClaw gateway live suite.

Please update the PR so that:

• pnpm check passes
• pnpm build passes
• the memory-core compile regressions are removed from this branch
• LM Studio is validated with a live gateway run that passes the tool path, not only discovery/onboarding

Once those are addressed, I can re-review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7ed05cd3e9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Suppress profile API key when Authorization header is set

resolveLmstudioRuntimeApiKey returns any resolved profile/env key even when an Authorization header is configured, because the header-only check is only used in the fallback path. In header-auth LM Studio/proxy setups with a stale lmstudio:default profile key, this causes downstream request assembly to replace the configured header with Bearer <stale-key> (for example via resolveLmstudioRequestContext in dynamic model discovery), leading to avoidable 401s and failed model resolution.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Require http(s) when normalizing LM Studio base URLs

This path accepts any URL protocol and returns it as normalized base URL, but inputs like localhost:1234 are parsed by URL as a custom scheme (localhost:), not an HTTP host. Because setup validation only checks non-empty input, users can enter common host:port values without http://, which then normalize to non-fetchable endpoints and cause discovery/setup failures even when LM Studio is running.

Useful? React with 👍 / 👎.