--- title: Run Claude Code against another model (OpenRouter, local) nav_title: Other models layout: default parent: English nav_order: 16 description: Point Claude Code at an Anthropic-compatible backend such as OpenRouter and run a session on a model other than Claude. --- # Using another model via OpenRouter {: .no_toc } - TOC {:toc} Claude Code can talk to any **Anthropic-compatible** backend. MulmoTerminal reads those backends from its config, their **keys from the environment the server was started with**, and lets you choose a model per session — so Kimi, DeepSeek, GLM, Gemini, GPT or Grok run in the same terminal you already use. This page walks through **OpenRouter** end to end, from the first setup to adding models of your own. Moonshot, a company LiteLLM gateway, or any other Anthropic-compatible backend follows the same steps. --- ## The shape of it Configuration lives in **three places**, each for a reason: | What | Where | Why there | |---|---|---| | **The backend** (URL, name of the key variable) | `providers` in `~/.mulmoterminal/config.json` | shared by the whole app. **Nothing ships built in — without this, there is nothing to choose** | | **The API key** | the server's environment (`.env`) | **that config file is served to the browser and the phone** — a key must never be in it | | **The default model** | a project's `.mulmoterminal.json` | it differs per project | On top of that, you can **pick per session at launch** — without rewriting the default. --- ## 1. Get an OpenRouter key 1. Create an account at [openrouter.ai](https://openrouter.ai) 2. Issue an API key under [Keys](https://openrouter.ai/settings/keys) (`sk-or-…`) 3. Add credit — usage-based, and some models below are $0.08 per million input tokens **Also check** your [privacy settings](https://openrouter.ai/settings/privacy). If they exclude every provider serving a model, that model answers `404 No endpoints available`. The rows marked *unreachable* in the table below hit exactly that — **not a defect in the model**. --- ## 2. Register the backend (**required**) **Without this, there is no MODEL select at all.** No backend ships with the app. The 27 measured models are built in, but what they carry is only the model **id and its measurements**. **Where to send requests (`baseUrl`) and which environment variable holds the key (`tokenEnv`) are not in them**, so until a provider is registered there is nowhere to send anything. Measured: | `providers` | Models offered | MODEL select | |---|---|---| | unset (fresh install) | **0** | hidden | | registered (`id` = `openrouter`) | **27** | shown | ### Add it in one command {: .no_toc } Adds the provider without disturbing the rest of your config — it backs the file up to `config.json.bak` first, and running it twice does not duplicate the entry. ```bash node -e ' const fs = require("fs"), os = require("os"), path = require("path"); const file = path.join(os.homedir(), ".mulmoterminal", "config.json"); const config = fs.existsSync(file) ? JSON.parse(fs.readFileSync(file, "utf8")) : {}; if (fs.existsSync(file)) fs.copyFileSync(file, file + ".bak"); config.providers = [ ...(config.providers ?? []).filter((p) => p.id !== "openrouter"), { id: "openrouter", label: "OpenRouter", baseUrl: "https://openrouter.ai/api", tokenEnv: "OPENROUTER_API_KEY", maxOutputTokens: 16000 }, ]; fs.mkdirSync(path.dirname(file), { recursive: true }); fs.writeFileSync(file, JSON.stringify(config, null, 2) + "\n"); console.log("added openrouter to " + file); ' ``` ### Or write it by hand {: .no_toc } If `~/.mulmoterminal/config.json` does **not** exist yet, paste this as the whole file: ```json { "providers": [ { "id": "openrouter", "label": "OpenRouter", "baseUrl": "https://openrouter.ai/api", "tokenEnv": "OPENROUTER_API_KEY", "maxOutputTokens": 16000 } ] } ``` **If the file already exists, add the `providers` key inside the existing `{ … }` — do not replace the file.** Replacing it drops your `cwdPresets`, header buttons and everything else. ```json { "cwdPresets": [ ... keep yours ... ], "chips": [ ... keep yours ... ], "providers": [ { "id": "openrouter", "label": "OpenRouter", "baseUrl": "https://openrouter.ai/api", "tokenEnv": "OPENROUTER_API_KEY", "maxOutputTokens": 16000 } ] } ``` ### Or ask an LLM {: .no_toc } In any directory, tell a Claude session `add OpenRouter to my mulmoterminal config`. The bundled `mulmoterminal-config` skill keeps your existing settings and writes a valid entry — **without your key in it**. **Under the id `openrouter`, do not list models here.** Registering a provider with this `id` is enough — the [27 measured models](#verified) appear in the picker on their own. You only list models to add ones that are **not** in that list (→ [Adding models](#add-models)). **Under any other `id`, `models` is required.** The measured list is OpenRouter's alone, matched by that exact id, so a backend registered as `deepseek`, `moonshot` or an in-house gateway starts with none — and **a backend with no models is not offered in the picker**, because a session named on a provider without a model refuses to start. List the ids you want to run (→ [Adding models](#add-models)). | Key | Meaning | |---|---| | `id` | the name other settings refer to; letters, digits and `. _ : / - ~`. **The built-in 27 models appear only when `id` is exactly `openrouter`** — any other id (Moonshot, an in-house LiteLLM, …) gets no presets, so [list models by hand](#add-models) | | `label` | what the picker shows | | `baseUrl` | **no trailing `/v1`** — see below | | `tokenEnv` | the **name** of the environment variable holding the key, never the key | | `maxOutputTokens` | optional; defaults to 16000 | | `models` | **required unless `id` is `openrouter`** — that is the only id with presets. Under `openrouter`, only to add models **not** in the built-in list (→ [Adding models](#add-models)) | ### Never end `baseUrl` in `/v1` {: .no_toc } Claude Code appends `/v1/messages` itself, so `https://openrouter.ai/api/v1` produces `…/v1/v1/messages` and every request 404s. MulmoTerminal rejects it before launching and says why. ### Do not lower `maxOutputTokens` {: .no_toc } Starved of output headroom, a thinking model spends its whole budget thinking and returns **empty** visible text — which reads as a hung session. Keep it at 16000 or above. --- ## 3. Put the key in the server's environment In the shell that starts MulmoTerminal, or a `.env` beside it: ```bash OPENROUTER_API_KEY=sk-or-… ``` **Never put it in the config file.** `config.json` is served to the browser and the phone through `GET /api/config` — that is exactly why `tokenEnv` is a variable *name*. **Restart the server** afterwards; the environment is read at startup. A provider whose key cannot be resolved **refuses to start the session**. Quietly falling back to Anthropic would send that session's prompts to a backend the directory never selected. --- ## 4. Use it ### Choose at launch Once at least one provider is usable, the empty cell's launch form grows a **MODEL** select. - The choice applies to **that session only** - It does not rewrite `.mulmoterminal.json` - Choosing nothing keeps the directory's default Once the session runs, the **badge on the header's first row** shows the running model and how full its context is — `Kimi K2.7 Code · ctx 12%`. If you don't see it, check that `ctx` is in your `chips`. ### Make it a project default In that project's `.mulmoterminal.json`: ```json { "provider": "openrouter", "model": "moonshotai/kimi-k2.7-code" } ``` Omit `provider` and give only `model` to pick a different model **on Anthropic itself**. Ids may contain letters, digits and `. _ : / - ~` — the same rule as a provider's `id`, so OpenRouter's "always the latest" aliases such as `~anthropic/claude-opus-latest` work too. A value shaped differently (whitespace, a leading dash) makes sessions in that directory **refuse to start**, rather than quietly running on some other model. The directory's other settings still load. A model id may also end in **`[1m]`** — Claude Code's own way of asking for the 1M context window, on an alias or a full name alike: ```json { "model": "claude-opus-5[1m]" } ``` The suffix is passed to `claude --model` exactly as you wrote it; Claude Code reads it and strips it before the id reaches the backend, so nothing here interprets it. It is the only bracketed suffix that exists — `[2m]`, `[200k]` and friends are refused — and it belongs at the very end of the id. See [Extended context](https://code.claude.com/docs/en/model-config#extended-context) for which models it actually changes anything for. A **provider** `id` may not carry it. ### What happens on resume {: .no_toc } - A model **picked at launch** is kept when that session resumes (for as long as the server runs) - A **`.mulmoterminal.json` default** is live like every other field there — edit it and the next launch uses it --- ## 5. Adding models {#add-models} Models outside the built-in presets work too. ### List them in the config {: .no_toc } ```json { "providers": [ { "id": "openrouter", "label": "OpenRouter", "baseUrl": "https://openrouter.ai/api", "tokenEnv": "OPENROUTER_API_KEY", "models": ["qwen/qwen3-coder", "inception/mercury-coder"] } ] } ``` They appear in the picker marked **`not tested`** — meaning we have not checked them, which is more honest than implying we had. Model ids are the `vendor/model` strings shown on each model's page at [openrouter.ai/models](https://openrouter.ai/models). ### Or ask an LLM to configure it {: .no_toc } If you would rather not edit JSON, ask a Claude session in any directory: > set up OpenRouter in my mulmoterminal config The bundled `mulmoterminal-config` skill knows this file's shape and the tested model list, and writes a valid config — **without ever writing your key into it**. ### Measure whether a model actually works {: .no_toc } **Answering a prompt and driving Claude Code are different things.** Measured here, some models replied fluently and never called a tool once — one printed `Read(file_path=…)` as prose instead of using it. So a probe that **cannot** be completed by talking — read one file, write another — runs through the real spawn path: ```bash yarn tsx scripts/model-trials.ts --provider openrouter --trials 3 qwen/qwen3-coder ``` ``` 3/3 16s qwen/qwen3-coder ``` Two models flipped between pass and fail across runs, so this records a **ratio, not a verdict**. The numbers in the built-in presets (`common/modelPresets.ts`) were measured the same way. --- ## Measured models {#verified} As of 2026-07-22, from a single OpenRouter account. **Passed** is how many attempts completed the tool-using probe above; **median** is over the attempts that passed; price is per million input / output tokens. | Model id | Name | Passed | Median | Context | Price (in/out) | |---|---|---|---|---|---| | `nvidia/nemotron-3-super-120b-a12b` | Nemotron 3 Super 120B | 3/3 | 18s | 1M | $0.08 / $0.45 | | `qwen/qwen3-235b-a22b-2507` | Qwen3 235B A22B | 3/3 | 16s | 262k | $0.09 / $0.55 | | `minimax/minimax-m2.7` | MiniMax M2.7 | 3/3 | 16s | 205k | $0.25 / $1 | | `deepseek/deepseek-v3.2` | DeepSeek V3.2 | 3/3 | 42s | 164k | $0.269 / $0.4 | | `minimax/minimax-m3` | MiniMax M3 | 3/3 | 14s | 1M | $0.3 / $1.2 | | `deepseek/deepseek-v4-pro` | DeepSeek V4 Pro | 3/3 | 20s | 1M | $0.435 / $0.87 | | `deepseek/deepseek-v4-flash` | DeepSeek V4 Flash | 3/4 | 26s | 1M | $0.094 / $0.188 | | `openai/gpt-oss-120b` | GPT-OSS 120B | 3/4 | 18s | 131k | $0.037 / $0.17 | | `moonshotai/kimi-k2-0905` | Kimi K2 0905 | 3/3 | 18s | 262k | $0.6 / $2.5 | | `moonshotai/kimi-k2-thinking` | Kimi K2 Thinking | 3/3 | 20s | 262k | $0.6 / $2.5 | | `moonshotai/kimi-k2.6` | Kimi K2.6 | 3/3 | 46s | 262k | $0.684 / $3.42 | | `z-ai/glm-5.2` | GLM 5.2 | 3/3 | 21s | 1M | $0.819 / $2.574 | | `moonshotai/kimi-k2.7-code` | Kimi K2.7 Code | 3/3 | 14s | 262k | $0.82 / $3.75 | | `moonshotai/kimi-k3` | Kimi K3 | 3/3 | 29s | 1M | $3 / $15 | | `tencent/hy3` | Tencent Hy3 | 3/3 | 17s | 262k | $0.14 / $0.58 | | `nvidia/nemotron-3-ultra-550b-a55b` | Nemotron 3 Ultra 550B | 3/3 | 13s | 512k | $0.6 / $3.6 | | `google/gemini-3.5-flash-lite` | Gemini 3.5 Flash-Lite | 3/3 | 11s | 1M | $0.3 / $2.5 | | `amazon/nova-2-lite-v1` | Nova 2 Lite | 2/3 | 20s | 1M | $0.3 / $2.5 | | `openai/gpt-5.6-luna` | GPT-5.6 Luna | 3/3 | 27s | 1M | $1 / $6 | | `openai/gpt-5.6-luna-pro` | GPT-5.6 Luna Pro | 3/3 | 69s | 1M | $1 / $6 | | `google/gemini-3.6-flash` | Gemini 3.6 Flash | 3/3 | 16s | 1M | $1.5 / $7.5 | | `x-ai/grok-4.5` | Grok 4.5 | 3/3 | 13s | 500k | $2 / $6 | | `openai/gpt-5.6-terra-pro` | GPT-5.6 Terra Pro | 3/3 | 38s | 1M | $2.5 / $15 | | `meta-llama/llama-4-maverick` | Llama 4 Maverick | **0/4** | — | 1M | $0.2 / $0.8 | | `qwen/qwen3.7-plus` | Qwen3.7 Plus | unreachable * | — | 1M | $0.32 / $1.28 | | `mistralai/mistral-medium-3-5` | Mistral Medium 3.5 | unreachable * | — | 262k | $1.5 / $7.5 | | `mistralai/devstral-2512` | Devstral 2512 | unreachable * | — | 262k | $0.4 / $2 | \* **unreachable** — the account used for the measurement had every serving provider excluded by its [privacy settings](https://openrouter.ai/settings/privacy). **Not a defect in the model**; another account may run it fine. **`meta-llama/llama-4-maverick` passed 0 of 4.** It connects and replies, but never calls a tool. It stays in the list rather than being dropped, so "what about that one?" is answered with a measurement instead of silence. If you are unsure where to start, try **Kimi K2.7 Code** (fast, coding-oriented) or **Nemotron 3 Super 120B** (cheap). --- ## When it doesn't work | Symptom | Cause and fix | |---|---| | Refuses to start, saying `OPENROUTER_API_KEY` is needed | the key isn't in the server's environment — add it to `.env` and restart | | Every request 404s | `baseUrl` ends in `/v1` | | `404 No endpoints available …` | that model is excluded by your [privacy settings](https://openrouter.ai/settings/privacy) | | Empty replies, looks hung | `maxOutputTokens` too low — keep it at 16000+ | | Talks but never uses tools | the model's own limitation — check the table or `model-trials.ts` | | No MODEL select in the launch form | no provider is registered (step 2), its key is missing, or it has no models to pick; the link beside the MODEL label — "Needs attention" / "Use another model…" — names what's missing | | A registered backend is missing from the MODEL list | it has no models: presets come with the id `openrouter` and no other, so [list its `models`](#add-models) | | `models` is in the config and the picker still offers none | every id in it was refused — a model id is letters, digits and `. _ : / - ~` (plus an optional trailing `[1m]`), so an object like `{"id": "…"}` or a value with a space is dropped. The server's log names what it dropped | | No model in the header | `ctx` is not in your `chips` | --- ## What protects the key - The key is **never stored in a config file** — only the name of the variable is - It reaches the session through a **0600 file**, not a command-line argument, so `ps` doesn't expose it to other users on the machine - An unresolvable key **refuses the launch**, so prompts never reach an unintended backend - `ANTHROPIC_API_KEY` is **removed** from a provider session's environment — left in place it would outrank the auth token ## When the model is reached by a COMMAND, not a URL {#custom-agents} Everything above assumes a backend you can name with a base URL and a token. If instead your model is reached by **running something** — `ollama launch claude --model … --`, a wrapper script, a second Claude Code install — that is a **custom agent** rather than a provider: it puts your own command in the Agent Picker and hands it Claude Code's own arguments, so the cell is still a real session (resume, cost, context, GUI tools). → [Custom agents](config.html#custom-agents) ## Running a local LLM (claude-ollama) To run Claude Code **fully locally against an Ollama model** — no cloud — use the bundled `claude-ollama` command (`npx -p mulmoterminal claude-ollama qwen3:4b`). See **[Local models with claude-ollama](claude-ollama.html)** for the details. --- ← [back to Configuration](config.html) / [English guide index](index.html)