Proof of concept: LM Studio + Claude Code CLI

A router on http://localhost:5090 that answers OpenAI-style chat requests:

Prompt	Goes to
easy (chit-chat, summarize, rewrite, translate, quick facts)	LM Studio (free, local)
hard (design, debugging, analysis, step-by-step reasoning), or anything complex	Claude Code CLI, on your own Claude subscription
private (customer records, salaries, medical history)	LM Studio only. If LM Studio is down, the request fails rather than going to Claude.

If Claude is unavailable (not signed in, or out of its session limit), hard prompts fall back to LM Studio and are flagged as degraded. If LM Studio is down, easy prompts go to Claude instead.

Prerequisites

1. .NET 10 SDK. Check with dotnet --version.
2. LM Studio with a chat model loaded and the local server started (Developer tab, or lms server start). The default port is 1234.
3. Claude Code, installed and signed in:

   claude          # then /login with your Claude account, then exit
   claude auth status

   On Windows, use the native installer, so that claude.exe is on PATH. The router never runs commands through cmd.exe, so an npm claude.cmd shim won't work; the preflight will tell you if that's what it found.

Run the self-test

From the repository root:

dotnet run --project samples/LmStudioClaudePoc -- --demo

The demo does the following:

1. Preflight. No Claude usage. It:
   • finds LM Studio and picks its first loaded chat model, unless you configured one;
   • finds claude;
   • checks that it is signed in, and with what (subscription or API key).
2. Calls each backend directly with "Reply with exactly one word: pong". A broken Claude setup shows its real error here instead of silently falling back.
3. Sends five prompts with "model": "auto" and shows which backend answered each one.
4. Streams a reply from Claude and reports the time to the first token.

It exits with code 0 when everything works. It uses about four short Claude requests from your subscription.

The output looks like this; the replies and timings will be your models':

Preflight
---------
[ OK ] LM Studio: http://localhost:1234, using 'qwen/qwen3-8b' (auto-selected from: qwen/qwen3-8b)
[ OK ] Claude CLI: 2.1.300 (Claude Code) at /usr/local/bin/claude, signed in with your Claude subscription

1. Each backend, called directly ("model": "lmstudio" / "claude")
[ OK ] lmstudio    0.4s  pong
[ OK ] claude      6.1s  pong

2. Automatic routing ("model": "auto")
kind     tier reason                answered by try   time  reply
easy     1    SemanticMatch         lmstudio    1     1.2s  I'm doing well, thanks for asking! …
easy     1    SemanticMatch         lmstudio    1     0.9s  The client meeting was moved to Friday …
hard     3    SemanticMatch         claude      1     9.8s  Event sourcing gives you a full audit …
hard     3    SemanticMatch         claude      1    11.2s  Calling .Result blocks the thread that …
private  1    SemanticMatch         lmstudio    1     0.7s  I don't have access to customer records …

3. Streaming from Claude ("stream": true)
[ OK ] first token after 4.9s, 23 chunks in 5.6s: 1 2 3 4 5 6 7 8 9 10

Proof of concept OK. Point any OpenAI-compatible client at http://localhost:5090/v1 with model "auto".

Claude is slower to answer than LM Studio, because every request starts the claude process.

Use it

dotnet run --project samples/LmStudioClaudePoc

The dashboard is at http://localhost:5090/router: live activity, the models' state, and every setting editable with hot reload (saved to semantic-router.json next to the app).

From Claude Code

Keep using Claude Code as before, pointed at the router:

ANTHROPIC_BASE_URL=http://localhost:5090 ANTHROPIC_AUTH_TOKEN=semantic-router claude --model semantic-router

--model semantic-router is the router's own name for auto, so Claude Code shows that the router picks the model. Every turn is routed on your latest message: easy and private ones go to LM Studio, hard ones to Claude. The claude-cli connector has ToolCalling: Prompted, so Claude Code's tool-bearing turns can land on your subscription too: the router describes Claude Code's tools to the inner claude -p run, parses the tool call it answers with, and hands it back to your Claude Code, which runs the tool. If your LM Studio model supports function calling (e.g. qwen3-coder), add "tools" to the lmstudio model's capabilities and easy agentic turns stay local. See Connect your tools.

Then send requests (poc.http has more examples):

curl -s http://localhost:5090/v1/chat/completions -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Compare event sourcing and CRUD for a billing system"}]}' -i

Invoke-RestMethod http://localhost:5090/v1/chat/completions -Method Post -ContentType 'application/json' `
  -Body '{"model":"auto","messages":[{"role":"user","content":"hello!"}]}'

Any tool that lets you set an OpenAI base URL will work. Use base URL http://localhost:5090/v1, any API key, and one of these models:

model	Effect
auto, semantic-router	the router decides
lmstudio	force LM Studio
claude	force Claude
auto:best	the most capable model available
claude-sonnet-…, claude-opus-… (what Claude Code sends)	mapped by Middleware:ModelAliases: *haiku* → auto:cheapest, every other Claude name → auto (Claude Code sends its default Opus name whether or not you chose Opus)
opus-pinned (claude --model opus-pinned)	family:opus: the catalog's Opus model, or else the most capable (here claude)

Every response carries routing headers:

Header	Meaning
X-Prompt-Route	the matched route (local-easy, claude-hard, private, agent-housekeeping, or none)
X-Prompt-Route-Tier / -Reason	what the prompt needed, and why
X-Prompt-Route-Model / -Attempt	who answered, and on which try
X-Prompt-Route-Degraded: true	answered by a less capable model because the better one was unavailable

The console logs each decision too. To see why a prompt would go where it goes without calling any model:

curl -s http://localhost:5090/semantic-router/classify -H "Content-Type: application/json" -d '{"prompt":"why does my async code deadlock?"}'

Tune it

Everything is in appsettings.json:

• LM Studio model. Set SemanticRouter:Models:0:Model to an id from curl http://localhost:1234/v1/models, or leave it empty to auto-select. You can also pass --SemanticRouter:Models:0:Model=<id> on the command line.
• Claude model. Change SemanticRouter:Models:1:Model to sonnet, opus, haiku or a full model name.
• What counts as easy or hard. Add example prompts to the local-easy and claude-hard routes; more varied examples route better.
  • The proof of concept uses built-in lexical embeddings, so matching is driven by shared words.
  • For meaning-based matching, load an embedding model in LM Studio and set Embeddings to { "Provider": "OpenAICompatible", "Endpoint": "http://localhost:1234", "Model": "<embedding id>" }. Raise DefaultSimilarityThreshold to around 0.6, then check with /semantic-router/classify.
• Complexity escalation. ComplexityTiers: [{ "MinScore": 0.6, "Tier": 3 }] sends any sufficiently complex prompt to Claude, even if it matched an easy route. Raise MinScore to use Claude less, or lower it to use Claude more.
• Unmatched prompts. FallbackTier: 1 sends prompts that match no route to LM Studio. Set it to 3 to send them to Claude.

Troubleshooting

Symptom	Fix
[FAIL] LM Studio: not reachable	Start LM Studio's server. If you changed its port, update Connectors:lmstudio:BaseAddress.
[FAIL] LM Studio: … requires authentication	"Require Authentication" is on in LM Studio's server settings. Copy the token there and run dotnet user-secrets set "SemanticRouter:Connectors:lmstudio:ApiKey" "<token>" (or export LM_API_TOKEN=<token>). Set SemanticRouter:Embeddings:ApiKey too if LM Studio serves your embeddings.
[FAIL] LM Studio: … rejected the API token	The configured token does not match LM Studio's. Copy it again from the server settings.
[FAIL] LM Studio: model '…' is not available	Use one of the ids it lists, or clear Models:0:Model.
[FAIL] Claude CLI: 'claude' not found	Install Claude Code, or set Connectors:claude-cli:Executable to the full path (on Windows, claude.exe).
[FAIL] Claude CLI: … not signed in	Run claude, then /login.
[WARN] … signed in with an API key	Requests will be billed to that key. Run /logout and /login with your Claude account to use your subscription.
[FAIL] claude … unknown option '…'	Your Claude Code version doesn't accept a flag the preset uses (they were checked against 2.1.293). Update Claude Code (claude update), or tell us the version.
[FAIL] claude … You've hit your session limit	Expected when you're out of quota. Hard prompts fall back to LM Studio (marked *) for 15 minutes, then Claude is tried again.
Hard prompts show lmstudio*	Claude failed or is cooling down; the console log shows why.

Use --skip-preflight to start without the checks.

What it does with your Claude login

Nothing directly. The router runs your installed claude binary, which uses its own login. The router never reads, stores or forwards credentials. It also:

• removes ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL from the CLI's environment, so your subscription is what gets used;
• runs every request in an empty temporary folder, with Claude Code's tools disabled;
• only listens on localhost, and only answers local callers through the CLI.

See Subscription CLI terms for why it is built this way and how that relates to Anthropic's terms.