# Mistral Vibe Integration

brainclaw integrates with **Mistral Vibe** through MCP tools, the universal `.agents/skills/brainclaw/SKILL.md`, and a `.live.md` companion file. Mistral Vibe is a Tier B agent — MCP + skills + CLI spawn, but no hooks (BeforePrompt is a feature request, not yet implemented) and no native rules file equivalent to `CLAUDE.md`.

Mistral Vibe is open-source (Apache 2.0), Paris-based, and supports on-premise / EU-hosted deployment — making it a strategic option for teams that need GDPR compliance and freedom from US CLOUD Act exposure.

## Auto-setup

`brainclaw init` detects Mistral Vibe (via `~/.vibe/` directory or `VIBE_HOME` env var) and writes:

- `.vibe/config.toml` — TOML MCP config registering brainclaw as a stdio server
- `AGENTS.md` — shared instruction file (also read by Codex and OpenCode)
- `.agents/skills/brainclaw/SKILL.md` — universal skill auto-discovered by Vibe
- `.vibe/live.md` (opt-in via `--include-live`) — refreshed on session-end / handoff with current plans, claims, traps, candidates, handoffs

Manual regeneration:

```bash
brainclaw export --format agents-md --write
brainclaw export --format agents-md --write --include-live   # add the live companion
```

## MCP configuration

Mistral Vibe reads MCP servers from `.vibe/config.toml` (project, prioritaire) or `~/.vibe/config.toml` (user fallback). Format is **TOML** — Mistral Vibe is the only Tier B agent in brainclaw's catalog using TOML rather than JSON/YAML.

```toml
[[mcp_servers]]
name = "brainclaw"
transport = "stdio"
command = "npx"
args = ["-y", "brainclaw@latest", "mcp"]
```

Other transports supported by Mistral Vibe : `http`, `sse` (Streamable-HTTP). brainclaw uses `stdio` by default, the simplest and most secure local path.

The brainclaw writer is **idempotent** — if `.vibe/config.toml` already declares a `[[mcp_servers]]` block whose `name = "brainclaw"`, the writer leaves it alone (preserves any user customization of `command`, `args`, etc.). Other `[[mcp_servers]]` entries (e.g., Serena, custom MCP servers) are also preserved on regeneration.

## Permissions

Mistral Vibe supports per-tool permission control in `.vibe/config.toml`:

```toml
[tools.bash]
permission = "always"     # auto-approve

[mcp_servers.brainclaw]
enabled_tools = ["bclaw_work", "bclaw_context", "bclaw_find", "bclaw_get"]
```

Patterns supported: exact names, glob (`bash*`), regex (`re:^bclaw_.*$`). brainclaw does not yet emit a curated `enabled_tools` list automatically — configure manually for headless workflows.

For interactive sessions, toggle global auto-approve with `Shift+Tab` or pass `--auto-approve` on the CLI.

## Instruction files

- **`AGENTS.md`** — shared instruction file (read by Mistral Vibe, Codex, OpenCode). Static brainclaw protocol + active constraints. Generated by `brainclaw export --format agents-md --write`.
- **`.agents/skills/brainclaw/SKILL.md`** — universal skill, auto-discovered by Mistral Vibe alongside `.vibe/skills/`. Provides slash-command access (`/brainclaw`).
- **`.vibe/live.md`** (opt-in) — live companion regenerated on session-end and handoff with plans, claims, traps, candidates, handoffs. Critical for Tier B agents like Vibe that lack pre-prompt hooks: this is the parity backstop so the agent sees current coordination state without an MCP round-trip on every prompt.

Mistral Vibe **does not have a native rules file** equivalent to `CLAUDE.md` — there is no documented mechanism to inject project-wide instructions outside `AGENTS.md` and skills. brainclaw works within that constraint.

## Headless invocation

Mistral Vibe is CLI-spawnable for parallel lanes and dispatched workflows. The canonical invoke template is:

```bash
vibe --prompt "{prompt}" --auto-approve --max-turns 5
```

- `--prompt` runs the task non-interactively from a single string.
- `--auto-approve` disables interactive approval on file edits and bash exec.
- `--max-turns N` caps the agent loop. brainclaw uses 5 by default; raise for larger tasks.
- `--max-price <USD>` (optional) caps cost before interruption.

For long prompts, `stdin_pipe` is supported (`echo "prompt" | vibe`) and is the fallback when `inline_arg` exceeds 8000 chars.

Output is text-formatted on stdout. brainclaw parses completion via the brief-ack file (same handshake pattern as Codex and other CLI-spawned workers).

## Capability profile

| Field | Value |
|-------|-------|
| Tier | B |
| MCP | yes (TOML config, project + user) |
| Hooks | no (BeforePrompt is feature request #531, not implemented) |
| Auto-approve | per-call (`--auto-approve` flag, or `Shift+Tab` interactive toggle) |
| Skills | yes (`.vibe/skills/` + universal `.agents/skills/`) |
| CLI spawnable | yes |
| Max concurrent tasks | 2 (conservative, see caveats) |
| Workflow model | task-based (one-shot `--prompt` exec) |
| MCP config scope | both (project + user fallback) |
| Prompt delivery | `inline_arg` (preferred, max 8000 chars), `stdin_pipe` (fallback) |
| Native rules file | no (relies on `AGENTS.md` + skill universel) |
| Live companion | `.vibe/live.md` (opt-in via `--include-live`) |

## Sovereignty & compliance

Mistral Vibe is a strategic option when EU/FR data sovereignty matters:

- **Mistral AI is Paris-based** — fully subject to GDPR, **not** subject to the US CLOUD Act.
- **Vibe CLI is open-source** (Apache 2.0) on [github.com/mistralai/mistral-vibe](https://github.com/mistralai/mistral-vibe) — auditable end-to-end.
- **Open-weight models** (Mistral Medium 3.5, Codestral, Devstral) deployable on-premise via Hugging Face — full data control, no third-party API calls when self-hosted.
- **EU-hosted Mistral cloud** available with a Data Processing Agreement (DPA).

For teams that have ruled out US-hosted agents on compliance grounds (regulated sectors, public sector, defense), this is often the only viable Tier B option.

## Caveats

- **Windows + Git Bash unsupported.** Mistral Vibe spawns its child shells via PowerShell or `cmd.exe` on Windows and cannot reuse environment variables sourced inside Git Bash. Tracked upstream as [issue #135](https://github.com/mistralai/mistral-vibe/issues/135). Workaround: run `vibe` from PowerShell or WSL.
- **No pre-prompt hooks.** Vibe has an experimental post-agent-turn hook system (v2.9.0+) but no `BeforePrompt` / `UserPromptSubmit` equivalent. Live coordination state must reach the agent via MCP calls or the `.vibe/live.md` companion regenerated on session-end / handoff — there is no real-time push mechanism.
- **No `CLAUDE.md` equivalent.** Mistral Vibe does not auto-load a free-form rules file at session start; integration relies on `AGENTS.md` + skills.
- **CLI freezes documented** on the current version (early v2.x). For production parallel dispatch, `max_concurrent_tasks` is capped at 2 conservatively.
- **Rate-limit messages are vague** (upstream issue) — when running heavy parallel workloads, expect generic `rate limit exceeded` errors without specific guidance.
- **Skill discovery overlap.** Vibe auto-discovers both `.vibe/skills/` and `.agents/skills/`. brainclaw writes only to `.agents/skills/brainclaw/SKILL.md` to avoid duplicate definitions; if a project already had a `.vibe/skills/brainclaw/` folder, remove it manually after this integration to avoid two competing definitions.

## See also

- [docs/integrations/overview.md](overview.md) — full agent matrix and integration tier model
- [docs/integrations/mcp.md](mcp.md) — recommended dynamic runtime path
- [Mistral Vibe official documentation](https://docs.mistral.ai/mistral-vibe/introduction)
- [Mistral Vibe GitHub repository](https://github.com/mistralai/mistral-vibe) (Apache 2.0)