Skip to content

Mistral Vibe

Vibe Code is Mistral’s coding assistant for terminal and editor workflows. The CLI can read files, edit code, and run commands. A compatible self-hosted inference endpoint lets you use open-weight models on a server you control. The model server and the machine executing coding tools can be different machines.

The provider/runtime analysis was refreshed on 2026-10-19 for Vibe 3.0.0; CLI flags and session handling were rechecked on 2026-10-21 against 3.0.1. Check vibe --version and vibe --help before copying automation examples; defaults have changed between releases. See upstream releases.

Install the CLI with a package manager, then confirm the installed version:

Terminal window
uv tool install mistral-vibe
vibe --version
vibe --help

For repeatable unattended work, record the version and review upgrades before changing it. An installed CLI does not prove that its configured provider can generate a response.

Flags to know: approvals and saved sessions

Section titled “Flags to know: approvals and saved sessions”

Checked against Vibe 3.0.1 on 2026-10-21:

TaskCommand
Auto-approve tool callsvibe --yolo (alias of vibe --auto-approve)
Continue the latest session in this directoryvibe --continue (or vibe -c)
Choose a saved sessionvibe --resume
Resume a specific sessionvibe --resume SESSION_ID
Fork the current conversation/branch inside the interactive session
Run a prompt and exitvibe --prompt "Your task" (or -p)
Continue with a headless follow-upvibe --continue --prompt "Continue the task"
Choose working directory/agentvibe --workdir /path/to/repo --agent accept-edits
Bound a scripted run--time-limit SECONDS, --max-turns N, --max-tokens N
Keep structured output and private evidence--output json, --output-dir /path/to/private-artifacts
Check flags/versionvibe --help; vibe --version
Terminal window
vibe --continue --yolo
vibe --resume SESSION_ID --yolo

--yolo changes approval handling; it does not create an OS sandbox or grant project trust. --trust is a separate, invocation-scoped choice. Current administrator policy can disable auto-approval. Saved sessions retain their previous agent; passing another --agent while resuming does not necessarily replace it. Check the active profile and tools before continuing. Resume needs saved session logs; a conversation branch does not isolate files.

This installed CLI has no --model or --fork option. Select models through the interactive /model command or reviewed active_model configuration; use /branch for a conversation fork. See official CLI/session guidance, headless details, and the cross-agent quick reference.

Configuration lives in ~/.vibe/config.toml; trusted projects can supply .vibe/config.toml. Review project configuration, hooks, and skills before granting trust. Use the official configuration reference to check fields against your version.

Keep the literal credential in one private location, such as ~/.vibe/.env, and refer to its environment-variable name from provider configuration. Do not duplicate the value in TOML, shell startup files, and project files. Create or edit the credential file with your editor rather than putting a key in shell history or command arguments.

Terminal window
mkdir -p ~/.vibe
chmod 700 ~/.vibe
# Create ~/.vibe/.env in your editor, then:
chmod 600 ~/.vibe/.env

A provider definition uses a variable name, not a key value:

[[providers]]
name = "mistral"
api_base = "https://api.mistral.ai/v1"
api_key_env_var = "MISTRAL_API_KEY"
api_style = "openai"
backend = "mistral"

Choose a currently available model with /config. Confirm the exact model ID and capabilities in your provider’s inventory. A selectable preset can still be unavailable to the account or unable to complete a request.

An HTTP 402 response containing API budget exhausted is a billing or allowance failure. Changing model aliases, creating another key in the same organization, or increasing a request’s token limit does not restore that budget.

Current Mistral plans include usage shared across Studio, the API, and Vibe Code. Older advice that every web subscription excludes API usage is no longer a reliable general rule. Check the organization and workspace belonging to the key actually used, its allowance, prepaid balance, pay-as-you-go setting, and spending limits. A paid subscription is not a promise of unlimited inference. See subscriptions.

Mistral documents that continuing beyond included usage requires both pay-as-you-go and positive prepaid credits; spending limits still apply. Requests can fail with 402 when the balance is exhausted. Restoring paid capacity is an account decision, not a configuration repair. See billing.

For a fallback independent of this balance, use a separately funded provider or a self-hosted endpoint. Routing another alias through the same exhausted account does not provide independence.

This illustrative preset connects to an SSH-forwarded server endpoint on loopback. Replace the model name with the exact name advertised by that server and set context to its actual allocation. The server performs inference; this example does not install or run model weights on the client.

active_model = "private-worker"
enable_telemetry = false
[[providers]]
name = "private-inference"
api_base = "http://127.0.0.1:8080/v1"
api_key_env_var = "PRIVATE_INFERENCE_API_KEY"
api_style = "openai"
backend = "generic"
[[models]]
name = "served-model-id"
provider = "private-inference"
alias = "private-worker"
max_context_length = 8192
max_output_tokens = 2048
thinking = "off"
thinking_levels = ["off"]
supports_images = false
input_price = 0.0
cached_input_price = 0.0
output_price = 0.0

Set PRIVATE_INFERENCE_API_KEY through your private credential store when the server requires authentication. Zero token prices here describe a server without a provider token bill; electricity, hardware, storage, and administration still cost money. A zero-priced preset also makes a CLI cost limit ineffective at limiting compute.

Keep the endpoint on loopback behind a tunnel or on a restricted private network. Authenticate shared endpoints and restrict client access. An OpenAI-compatible API must also correctly implement the model’s chat template and tool-call format; compatibility is not a quality guarantee. See Mistral’s offline-model guide.

For each preset record the exact served model/version, provider, allocated context, output limit, image support, supported reasoning settings, and any token prices with a verification date. A model’s published maximum context can greatly exceed the runtime’s configured context. Vibe’s preset must reflect the smaller usable allocation.

Reasoning levels are model-specific. Mistral’s reasoning documentation distinguishes off/high models from models such as GLM 5.3 with low/high/max. Do not advertise every CLI reasoning level for every model. With generic backends, verify that the server actually receives and honors the requested setting; a client label alone does not disable thinking.

Keep hosted prices current if using cost estimates. Refresh capabilities from provider inventory, but do not treat an inventory response as authoritative pricing unless the provider includes billing information. Avoid undated price tables or assumptions about -latest aliases.

Vibe 3.0 requires the Unified Harness runtime; it does not silently fall back to a legacy agent loop. Its active model route forwards configured temperature, top_p, and max_output_tokens. Check the versioned runtime mapping, rather than inferring behavior from an older adapter still present in the package.

For the generic OpenAI chat adapter, thinking = "off" omits reasoning_effort; it does not send enable_thinking=false in chat_template_kwargs. Configure and verify the endpoint’s explicit thinking-off mode when the model supports it. Other thinking labels are forwarded as effort values: Vibe’s max must not be assumed equivalent to a model’s xhigh. See the generic payload builder and CLI level definitions.

The ordinary model schema does not expose top_k, min_p, repetition/presence penalties, or sampling seeds. Matching a benchmark that uses these fields requires verified per-model endpoint defaults or an explicit translating proxy/adapter. Inspect sanitized request fields and effective server settings before claiming equivalent sampling.

The generic chat adapter preserves reasoning_content history, or maps it to the configured provider field; it does not generally discard that history. A model template’s separate preserve_thinking control still needs verification. See the versioned message conversion.

MCP is optional. Vibe already has file and shell tools; add external capabilities only when the task needs them. Start with one server, inspect its tools using /mcp, and test the exact permission policy. Server-prefixed tool names and transport fields are described in Mistral’s MCP documentation.

Two useful additions are:

ServerPurposeSuggested boundary
Playwright MCPCheck browser behavior and narrow viewportsPinned package and browser versions, isolated test profile, explicit target allowlist; no personal browser cookies
Context7Retrieve version-specific library documentationRead-only documentation queries; do not send secrets or private source

Pin and lock a reviewed Playwright package in a dedicated tools directory. Configure its installed executable rather than resolving @latest each run. An illustrative stdio configuration is:

[[mcp_servers]]
name = "playwright"
transport = "stdio"
command = "/absolute/path/to/tools/node_modules/.bin/playwright-mcp"
args = ["--headless", "--isolated"]
startup_timeout_sec = 30
tool_timeout_sec = 60

Install the browser revision required by that pinned package. Confirm both a desktop and a narrow viewport when evaluating responsive changes. An isolated browser profile separates session storage; it does not establish a network or operating-system sandbox.

Context7’s documented remote endpoint can be configured separately:

[[mcp_servers]]
name = "context7"
transport = "streamable-http"
url = "https://mcp.context7.com/mcp"
api_key_env = "CONTEXT7_API_KEY"
api_key_header = "Authorization"
api_key_format = "Bearer {token}"

Use an environment reference for any required key. Verify current authentication and quota requirements with Context7. Remote documentation calls require network access, so disable them during offline tests. Keep browser, documentation, repository, and deployment capabilities separate; a worker seldom needs all of them.

Skills provide reusable instructions; agents define operating profiles. Read downloaded instructions before enabling them. Start interactive work with ask or accept-edits, keep global permission bypass disabled, and explicitly restrict enabled tools for unattended jobs.

Vibe 3.0.1 programmatic mode uses default_agent; approval-requiring tools are denied when nobody can answer. Pass the intended --agent and, when appropriate, --auto-approve explicitly for a fresh scripted run. A resumed session retains its saved agent; verify it rather than assuming a new flag replaced the old profile. See official agents documentation.

AGENTS.md — Your Project’s AI Rules File

Section titled “AGENTS.md — Your Project’s AI Rules File”

Put repository commands, owned paths, acceptance criteria, and stop conditions in project instructions. Give a worker the exact baseline, the smallest necessary file set, and executable checks. Instructions guide behavior; enforcement belongs in tool permissions, credential boundaries, filesystem isolation, and network policy.

Select a model that fits the server’s memory at the intended quantization and context. Total weights matter even for mixture-of-experts models with relatively few active parameters. Mistral recommends Devstral Small 2 for coding and documents substantially greater memory needs for Mistral Small 4. See server-only model screening for a qualification workflow.

Separate inference from code execution. The model service should receive prompts and return text/tool proposals; a constrained controller decides what, if anything, executes. A successful tool-call JSON probe does not establish safe autonomous operation.

Claude Code + Vibe Orchestration (Best of Both Worlds)

Section titled “Claude Code + Vibe Orchestration (Best of Both Worlds)”

A stronger model can plan work, write a narrow contract, and review a less expensive server worker’s result. Delegate bounded transformations, reconnaissance, fixtures, or documentation with objective checks. Retain independent review for security, identity, money, migrations, deployment, and state transitions.

Run workers in separate worktrees and a constrained execution environment, with an externally enforced wall deadline, turn limits, per-response token caps, and cumulative usage/spend guards. Token and price guards react to reported usage and may overshoot; retain provider/account spending controls. Inspect the actual diff and tests before accepting the result. See headless execution.

cc-vibe — Using Claude Code with the Mistral API

Section titled “cc-vibe — Using Claude Code with the Mistral API”

A compatibility proxy can translate API formats. Its alias names do not establish model identity: an alias named sonnet or opus can route to a Mistral model. Document the upstream model and account behind every alias. Determine whether Vibe uses the proxy by inspecting its selected provider configuration; proxy health is relevant only when that is the configured route.

For a maintained setup path, use LiteLLM’s gateway quickstart and Claude Code integration guide. The gateway exposes Anthropic Messages at /v1/messages; this provides an API format, not an Anthropic model identity.

  1. Configure the gateway with an explicit model alias mapped to the intended upstream provider/model. Keep the upstream credential on the gateway and issue a narrowly scoped virtual key to the client.
  2. In an isolated client session, set ANTHROPIC_BASE_URL to the gateway root, ANTHROPIC_AUTH_TOKEN from the private virtual-key store, and ANTHROPIC_MODEL to that exact configured alias. Avoid a conflicting ANTHROPIC_API_KEY. Do not append /v1/messages to the base URL or paste keys into shell history.
  3. Send a harmless prompt, then verify the gateway’s usage record shows the expected alias, upstream route, and virtual key. Test tool-call behavior separately before allowing repository edits. A healthy gateway or a model name in the picker does not establish that routing works.

These are generic setup and verification steps; they do not attest that a particular proxy or model has been qualified.

Shared instructions and mixed-model alternatives

Section titled “Shared instructions and mixed-model alternatives”

Native OpenCode V2 can reuse portable Claude skill files and assign different models to child agents, but it needs explicit handling of Claude rules and profiles. Claude Code with a Go gateway retains Claude’s customization system. See shared rules, skills, and agents and OpenCode model orchestration and caching.

Recent changes

  1. docs(wiki): add coding agent approval and session flag references @sac
  2. docs(wiki): audit reliability study and record OpenCode qualification @sac
  3. docs(wiki): refresh OpenCode V2 orchestration and shared customization @sac
  4. docs(wiki): clarify Vibe limits and report bounded model screens sac
  5. Refresh Vibe setup and private server model guidance sac