Use Codex and Claude Code

mistral.rs can back coding-agent clients through the compatibility APIs those clients already speak.

Use default for a single-model server. With multi-model serving, use a model id exactly as it appears in GET /v1/models.

Start mistral.rs

Start a model that is tuned for tool use and code:

mistralrs serve -p 1234 -m Qwen/Qwen3-Coder-Next

Check that the server is reachable:

curl http://localhost:1234/v1/models

For Codex and Claude Code, let the coding agent own normal file edits, shell commands, and repository inspection. Enable mistral.rs server-side agent tools only when you deliberately want web search or Python code execution to happen inside the mistral.rs server as well.

Codex

Codex uses the OpenAI Responses wire API for custom providers; wire_api = "responses" is the only supported value. Put provider configuration in your user-level ~/.codex/config.toml. Per the Codex configuration reference, model_provider and model_providers are ignored when they appear in project-local .codex/config.toml files.

model = "default"
model_provider = "mistralrs"
model_context_window = 32768
model_reasoning_summary = "none"
model_supports_reasoning_summaries = false

[model_providers.mistralrs]
name = "mistral.rs"
base_url = "http://localhost:1234/v1"
wire_api = "responses"
request_max_retries = 1
stream_max_retries = 0
stream_idle_timeout_ms = 300000

With model_provider set at the top level, plain codex uses the local server. To keep mistral.rs as a switchable alternative instead, move the model and model_provider lines into a profile file (~/.codex/mistralrs.config.toml) and launch with codex --profile mistralrs.

If a reverse proxy enforces authentication, add env_key = "MISTRALRS_API_KEY" under [model_providers.mistralrs] and export that variable before launching Codex. The local mistral.rs server itself does not validate API keys.

Codex tool calls arrive through /v1/responses as Responses function tools. mistral.rs routes them through the same tool-calling path used by Chat Completions. For Responses endpoint behavior, see OpenAI compatibility.

Claude Code

Claude Code should use the Anthropic-compatible root URL, without /v1. Claude Code appends /v1/messages and /v1/messages/count_tokens itself.

For a persistent local setup, add this to ~/.claude/settings.json or to a project-local .claude/settings.local.json:

{
  "model": "sonnet",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:1234",
    "ANTHROPIC_API_KEY": "not-used",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "default",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "default",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "default",
    "ANTHROPIC_CUSTOM_MODEL_OPTION": "default",
    "ANTHROPIC_CUSTOM_MODEL_OPTION_NAME": "mistral.rs default",
    "ANTHROPIC_CUSTOM_MODEL_OPTION_DESCRIPTION": "Local model served by mistral.rs",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Or set the same values for one shell session:

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_API_KEY=not-used
export ANTHROPIC_MODEL=sonnet
export ANTHROPIC_DEFAULT_SONNET_MODEL=default
export ANTHROPIC_DEFAULT_OPUS_MODEL=default
export ANTHROPIC_DEFAULT_HAIKU_MODEL=default
export ANTHROPIC_CUSTOM_MODEL_OPTION=default
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
claude

Map the Sonnet, Opus, and Haiku defaults to default so the sonnet setting and Claude Code's background and planning calls all use the single loaded mistral.rs model. If you serve several models, map each Claude Code default to the local model id you want for that role.

Every variable above is documented in the Claude Code environment variables reference. CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC turns off auto-update checks, telemetry, and error reporting so the client stays local.

Claude Code client tools arrive as Anthropic tool definitions and later tool_result content blocks. mistral.rs translates these to its internal tool format. For direct Anthropic examples, see the Anthropic Messages API guide.

Server-side agent tools

The coding clients already provide their own editing and shell tools. Server-side mistral.rs tools are separate:

Server-side web search and code execution require the corresponding mistral.rs agentic runtime flags at server startup. Use them when you want the model server to perform web search or Python execution independently of the coding client's own terminal tools.

Examples

Copyable config snippets live in examples/server/:

Troubleshooting

Upstream references

• Codex configuration reference
• Claude Code environment variables
• Claude Code LLM gateway configuration