mistral-mcp

> Mistral, wherever you run it.

> MCP server for the full Mistral AI API — chat, OCR, audio (Voxtral), code (Codestral), vision, agents, batch, durable workflows — against Mistral Cloud or your own infrastructure.

> Plug into Claude Code, Cursor, Zed, Windsurf, or Claude Desktop in one command.

>

> _Version française : README.fr.md_

What this is

`mistral-mcp` exposes the full Mistral AI API as a set of MCP tools, resources, and prompts. An MCP client (Claude Code, Cursor, etc.) can call `mistral_ocr` to extract text from a PDF, `voxtral_transcribe` to transcribe a meeting recording, or `workflow_execute` to start a durable multi-step process — all without leaving the agent loop.

Unique to Mistral and not available from other MCP servers:

• `mistral_ocr` — Mistral Document AI: structured text + bbox annotations from any PDF or image
• `voxtral_transcribe` — Voxtral: transcription with optional speaker diarization
• `codestral_fim` — Codestral fill-in-the-middle (FIM) for inline code completion
• `workflow_*` (6 tools) — Temporal-backed durable execution: what is deployed and runnable, what is running, human-in-the-loop signals, and graceful or forced stop
• French-optimized models (`mistral-large-latest`, `mistral-medium-latest`) and curated French prompts

What this server does not expose: fine-tuning, user management, non-FR/EN prompts.

Why this matters for European teams

`mistral-mcp` is designed for teams that want to use Mistral capabilities inside MCP clients (Claude Code, Cursor, Zed, Windsurf, Claude Desktop) while keeping control over deployment, API keys, cache behavior, and tool exposure.

This can be useful for European organisations evaluating AI stacks under GDPR, DORA, sector-specific constraints (HDS, EBA), or internal sovereignty requirements.

What this project provides:

• self-hosted MCP server, no mandatory SaaS proxy
• bring-your-own Mistral API key (BYOK) — Mistral states API data is not used to train its models
• `MISTRAL_BASE_URL` routes every call to your own OpenAI-compatible endpoint (vLLM, TGI, LiteLLM, an internal gateway) — no traffic to `api.mistral.ai`
• lean `core` profile and focused `metier-docs` profile to limit tool exposure
• `process_document` cache configurable per-call and via `MISTRAL_MCP_CACHE_DIR`, with a retention window (`MISTRAL_MCP_CACHE_TTL_HOURS`, default 7 days, `0` to disable) after which entries are deleted, not merely bypassed
• ID document cache bypass enabled by default, even when `kind:"auto"` resolves to `id_document`
• Streamable HTTP + bearer auth path for controlled / on-prem deployments
• French-first prompts and skills (meeting minutes, legal summary, invoice reminder, commit message, email reply)
• free Experiment tier on the Mistral side covers most evaluation needs (~1B tokens/month)

What this project does NOT claim:

• it is not a GDPR, DORA, HDS, or ISO certification, and it does not replace a DPIA, vendor review, security audit, or legal assessment
• Mistral provider terms, data residency, subprocessor list, retention settings, and incident handling must be reviewed separately at mistral.ai/terms and legal.mistral.ai
• this repo is community-maintained, not an official Mistral integration; nothing here modifies Mistral's contractual data terms with you

In practice, `mistral-mcp` reduces the integration surface you have to assess. It does not replace the legal/compliance work itself.

Quick start

Claude Code (recommended — auto-installs, prompts for API key, ships 11 skills):

/plugin install mistral-mcp@swih-plugins

Cursor / Zed / Windsurf / Claude Desktop — add to your MCP settings JSON:

{
  "mcpServers": {
    "mistral": {
      "command": "npx",
      "args": ["-y", "mistral-mcp@latest"],
      "env": { "MISTRAL_API_KEY": "your_key_here" }
    }
  }
}

Manual Claude Code registration:

claude mcp add mistral -- npx -y mistral-mcp@latest

Profiles

`MISTRAL_MCP_PROFILE` controls how many tools are exposed (default: `core`).

> `full` is accepted as a deprecated alias of `admin` for backward compatibility.

MISTRAL_MCP_PROFILE=admin npx mistral-mcp

Read `mistral://capabilities` from any client to see which tool families are on,

which are off, and why — no need to diff this table against your deployment.

Tools

Core profile (16 tools — always available)

Documents vertical (`MISTRAL_MCP_PROFILE=metier-docs`)

Admin profile only (+28 tools, set `MISTRAL_MCP_PROFILE=admin`)

Resources

Prompts

Curated prompts with structured arguments and MCP completion support:

Claude Code skills (11)

Install via the `swih-plugins` marketplace to get these namespaced skills:

Routing

• `/mistral-mcp:mistral-router` — picks the right Mistral model + tool for any task

Code

• `/mistral-mcp:codestral-review` — fetches the current diff, runs a focused review

French workflows

• `/mistral-mcp:french-commit-message` — Conventional Commits message in French
• `/mistral-mcp:french-meeting-minutes` — audio or text → structured French minutes
• `/mistral-mcp:french-invoice-reminder` — B2B dunning letter with controlled tone

Document & audio processing

• `/mistral-mcp:contract-analyzer` — OCR → risk-rated clause extraction (JSON)
• `/mistral-mcp:pdf-invoice-extractor` — OCR → structured invoice fields for reconciliation
• `/mistral-mcp:audio-dispatch` — transcribe + diarize → per-speaker action plan

Human-in-the-loop workflows

• `/mistral-mcp:contract-review-workflow` — durable contract review with approval gates
• `/mistral-mcp:compliance-audit-workflow` — multi-step audit with mid-run findings + decisions
• `/mistral-mcp:research-pipeline-workflow` — hypothesis-driven research with amendment injection

Install

# Run directly (no global install)
npx mistral-mcp

# Global install
npm install -g mistral-mcp && mistral-mcp

# Docker
docker build -t mistral-mcp .
docker run -i --rm -e MISTRAL_API_KEY=your_key mistral-mcp

# From source
git clone https://github.com/Swih/mistral-mcp.git
cd mistral-mcp && npm install && npm run build
node dist/index.js

Document ingestion, evaluated

`process_document` ships with a corpus and a harness, because "handles

heterogeneous PDFs" is a claim, and a claim without a measurement is marketing.

npm run fixtures:generate   # rebuild the corpus from source (no key needed)
npm run eval:docs           # score it against real OCR (needs MISTRAL_API_KEY)

The corpus is eight synthetic documents chosen for the cases that actually

break ingestion pipelines, not for the ones that flatter them: a rotated

landscape scan (`/Rotate 90`), ruled line-item tables, side-by-side address

columns, a blank page in the middle of a document, mixed FR/EN, French accents

and the euro sign, and one near-empty page. Ground truth for each document —

expected kind, page count, and the strings that must survive OCR — lives in

`test/fixtures/corpus.json`.

Everything in it is invented: fictional companies, fictional people, fictional

identifiers. No real PII is in this repo, and none should be added — the

corpus is only useful if it can be published.

`npm run eval:docs` reports, per document, whether `kind: "auto"` classified it

correctly, whether the required fields survived, and the OCR confidence. It

then derives a `minOcrConfidence` from the run: the midpoint between the worst

document that extracted cleanly and the best document marked low-signal. When

those two overlap, it says no threshold is defensible rather than inventing

one.

The shipped default of `0.3` is a conservative starting point, not a

measured value. Run the harness on your own documents and set the number that

run justifies.

Observability

Every tool call emits one JSON line on stderr, and the caller's W3C trace

context follows the request all the way to the inference endpoint.

{"ts":"2026-08-28T09:14:02.117Z","kind":"tool_call","tool":"mistral_ocr","outcome":"ok","duration_ms":1840,"trace_id":"4bf92f3577b34da6a3ce929d0e0e4736","span_id":"00f067aa0ba902b7"}

• Trace continuity. `traceparent`, `tracestate` and `baggage` arrive in the

MCP request's `_meta` and are stamped onto the outgoing HTTP call, so your

collector joins the MCP span to the Mistral (or vLLM) span it caused instead

of showing two unrelated traces. A malformed header is ignored, never fatal.

• No payloads, ever. A line says what ran, how long it took and whether it

failed. Prompts, documents, transcripts, arguments and model output never

appear — those are your records, and this process has no business copying

them into a log it does not own. `test/stdio/observability.test.ts` asserts

that negative directly against the built binary.

• Nothing to enable, and one thing to disable. It is on by default because

an audit trail you have to discover is one you will not have when you need

it. `MISTRAL_MCP_AUDIT=off` silences it. stderr is used because stdout

carries JSON-RPC, and because MCP's own `logging` capability is deprecated

in 2026-07-28 in favour of stderr and OpenTelemetry.

The instrumentation wraps `registerTool` rather than each handler, so a tool

cannot be left out of the trail without being left out of the server.

Self-hosted inference

Point `MISTRAL_BASE_URL` at any OpenAI-compatible endpoint — vLLM, TGI, LiteLLM,

an internal token factory — and every request goes there instead of

`api.mistral.ai`:

MISTRAL_BASE_URL=http://vllm.internal:8000/v1 MISTRAL_DEFAULT_MODEL=my-org/mistral-small-3.2 npx mistral-mcp

Two things change when the endpoint is not Mistral's:

1. The profile becomes `self-hosted`. Only the five tools such an endpoint

can actually serve stay registered — `mistral_chat`, `mistral_chat_stream`,

`mistral_embed`, `mistral_tool_call`, `mistral_vision`. OCR, Voxtral, Files,

Batch and Workflows are Mistral-platform endpoints; advertising them against

vLLM would only produce 404s the calling model has to guess its way out of.

Set `MISTRAL_MCP_PROFILE` explicitly if your gateway does proxy the full API.

2. Model ids are no longer checked against a list. Any non-empty string is

forwarded as-is, because your endpoint's identifiers are yours.

`mistral://capabilities` reports the active endpoint, the profile, whether it was

inferred, and the reason each unavailable family is off.

Compose and Kubernetes manifests, plus the full environment reference, are in

`deploy/README.md`.

Protocol

The server speaks MCP 2026-07-28 and the 2025-era handshake, from the same

tool registrations, on the same endpoint. That matters because practically

every client shipping today still opens with the 2025 handshake: upgrading the

server does not ask anyone to upgrade their client.

`test/stdio/protocol-eras.test.ts` drives the built binary with a real 1.30.x

client and a real 2026-07-28 client and asserts both see the same tools — the

compatibility claim above is a test, not a promise.

Built on `@modelcontextprotocol/server` 2.x. Sampling and elicitation tools are

not exposed: sampling is deprecated in 2026-07-28, and the multi-round-trip

replacement is a client capability this server has no use for.

Transport

HTTP env vars: `MCP_HTTP_HOST`, `MCP_HTTP_PORT`, `MCP_HTTP_PATH`, `MCP_HTTP_TOKEN` (bearer auth), `MCP_HTTP_ALLOWED_ORIGINS`.

HTTP serving is stateless per request in both protocol eras, so `MCP_HTTP_STATELESS` no longer does anything and was removed in 0.10.0. Setting it is harmless.

`/healthz` is public and does not touch the MCP server.

Use as a Mistral Connector (beta)

`mistral-mcp` ships the Streamable HTTP transport and bearer auth that Mistral Connectors require. Deployment guides for Cloudflare Tunnel, Fly.io, and Cloud Run are in `deploy/connector-public.md`.

curl -X POST https://api.mistral.ai/v1/connectors \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -d '{"name":"mistral_self","server":"https://your-deploy/mcp","visibility":"private"}'

> Mistral Connectors expose tools only today. Resources and prompts remain available via local clients.

Comparison with other Mistral MCP servers

`mistral-mcp` differentiates by combining OCR, Voxtral diarization, Codestral FIM, and Temporal-backed Workflows in one server, with French-first prompts and a Claude Code plugin marketplace.

Development

npm run dev      # tsx watch
npm run build    # tsc → dist/
npm run lint     # tsc --noEmit
npm test         # 190+ tests (unit + contract + stdio e2e + live API)
npm run inspector

Test pyramid: unit → contract → stdio e2e → live API (requires `MISTRAL_API_KEY`).

License

MIT — Copyright Dayan Decamp