HarnessMatch
Menu

Coding agent

Mistral Vibe

Open multi-provider CLI and IDE agent with permissions, delegation, and rewind.

Status
Active
License
Apache-2.0
Evidence
Documented + code-verifiable, 26 sources
Product record checked
2026-07-27
Compare harnessesInspect sources

At a glance

Mistral's Apache-2.0 coding harness combines an interactive CLI and ACP editor integration with local or hosted models, delegated subagents, MCP, skills, policy hooks, structured automation, durable sessions, and message-level file rewind.

Good choice if

  • Vibe coders who want approvals, a plan mode, and file rewind without giving up a capable agent
  • Mistral-plan users or teams that want an open, inspectable terminal and ACP editor agent
  • Developers running local or OpenAI-compatible models through a configurable provider layer
  • Automated runs with turn, token, and price budgets plus structured output
  • Teams enforcing policy through per-tool permissions and inherited lifecycle hooks

Check before choosing

  • The local CLI executes on the host: permissions and Git worktrees are controls, not an OS sandbox; Vibe Code Web's managed sandbox is a separate surface
  • Anonymous telemetry, update checks, and auto-update are enabled by default; a fully offline setup requires disabling them and any networked tools
  • Official docs currently lag the v2.22.0 code on programmatic approval defaults and MCP OAuth, so pinned source and changelog take precedence
See 1 more considerations
  • There is no built-in browser automation; web search and fetch tools are not a browser agent

Why it qualifies as a coding harness

This confirms category fit, not product quality. Every required criterion links back to first-party evidence.

Qualifies4 of 4 required criteria evidenced

  • Adaptive agent loop

    Documented

    The system repeatedly observes results and chooses the next action instead of following a fixed one-pass graph.

    EvidenceVibe overview

  • Repository tool execution

    Documented

    The system can use tools to inspect and change a repository or its execution environment.

    EvidenceWork with the CLI

  • Task-aware context management

    Documented

    The runtime assembles, updates, compacts, retrieves, or persists task-relevant context while work proceeds.

    EvidencePinned local-session architecture

  • Model-independent runtime control

    Documented

    Permissions, budgets, interruption, policy, or stop controls operate outside the model's own text generation.

    EvidenceSafety, approvals, and permissions

Membership establishes category fit only. It does not score quality, safety, autonomy, model capability, or benchmark performance. · Read the membership rule.

How it works under the hood

Seven mechanisms mapped from first-party records. These labels describe what the harness provides, not how intelligent its model is.

7/7layers documented
  • Execution & isolationWorkspace isolationDocumented mechanism, not a performance score.
  • Tooling & integrationsExtensible toolsDocumented mechanism, not a performance score.
  • Context & stateManaged contextDocumented mechanism, not a performance score.
  • Lifecycle & recoveryCheckpoint/rewindDocumented mechanism, not a performance score.
  • ObservabilityStructured tracesDocumented mechanism, not a performance score.
  • VerificationTool-assistedDocumented mechanism, not a performance score.
  • Governance & permissionsPolicy controlsDocumented mechanism, not a performance score.

Public code audit

3/5public artifacts present
Security policy
Not found
CI workflow
Present at inspected commit
Automated tests
Present at inspected commit
Evaluation assets
Not found
Contributor documentation
Present at inspected commit

The v2.22.0 tree contains 387 Python test files and CI but no repository security policy or dedicated evaluation suite. Product docs still contradict shipped code on programmatic approval defaults and MCP OAuth; no product benchmark is inferred from engineering tests.

Inspect commit 89350a4064ca, checked 2026-07-27

Measured configurations

No benchmark run passes the full metadata admission policy for this harness yet. Missing data is not scored as zero.

Benchmark policy and all runs

Capability support

Documented first-class product support, checked against the sources below.

External tools (MCP)
DocumentedProduct-supported MCP integrationSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Local models
DocumentedLocal or self-hosted model pathSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Agent parallelism
DocumentedDelegated or parallel agent workflowSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Runs without an open UI
DocumentedNon-interactive or automation surfaceSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Browser control
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Isolated execution
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Undo file changes
DocumentedProduct-supported file or session rollbackSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.

Primary evidence

Each capability claim is tied to a first-party record and a verification date.

Product record checked 2026-07-27
Pinned Mistral Vibe source treeAudited v2.22.0 CLI implementation, package layout, tests, documentation, and release-aligned source stateOfficial repositoryPinned Apache-2.0 licenseLicense text for the open CLI codebase, kept separate from model and hosted-service termsOfficial repositoryMistral Vibe v2.22.0 releaseVersioned release provenance for the exact source commit used by this auditOfficial repositoryInstall and setupSupported operating systems, Python requirement, Mistral-plan access, API keys, offline operation, and setup flowOfficial docsWork with the CLIInteractive and programmatic modes, tool review, trust, structured output, budgets, sessions, and cost-estimate caveatOfficial docsSafety, approvals, and permissionsLayered agent policy, trusted-folder boundary, per-tool rules, outside-directory confirmation, and auto-approve riskOfficial docsAgents and subagentsBuilt-in approval profiles, plan mode, delegated explore agent, custom subagents, and AGENTS.md discoveryOfficial docsLifecycle hooksPre-tool, post-tool, and post-agent policy hooks, strict failure behavior, structured payloads, and subagent inheritanceOfficial docs
View 18 additional sources
Additional first-party evidence18 sources
MCP serversMCP transports, static authentication, tool naming, permissions, and the documentation's stale OAuth limitationOfficial docsSkillsReusable skill format, discovery paths, enable and disable rules, and user-invocable commandsOfficial docsOffline and local modelsvLLM, llama.cpp, LM Studio, Ollama, generic OpenAI-compatible providers, hardware guidance, and offline network controlsOfficial docsCLI configurationProvider presets, OpenRouter example, tools, hooks, configuration precedence, and default-on telemetry and updatesOfficial docsCLI and editor surfacesCLI, VS Code, ACP-compatible editors, and the boundary between local execution and remote web sessionsOfficial docsVibe Code Web sandbox boundarySeparate single-tenant web sandbox, outbound-network posture, credential scope, ephemerality, and missing network allowlistsOfficial docsPinned shipped defaultsDefault provider and model catalog, local llama.cpp provider, compaction, policy bypass, telemetry, OTEL, connectors, and session settingsOfficial repositoryPinned shell permission implementationHost-shell execution, ask-by-default posture, safe-command allowlist, denylists, sudo sensitivity, timeout, and output capOfficial repositoryPinned built-in agent definitionsDefault, plan, accept-edits, auto-approve, explore-subagent, and Lean profiles with their actual overridesOfficial repositoryPinned CLI entrypointCurrent programmatic approval behavior, explicit auto-approve, budgets, structured output, worktrees, trust, and resume flagsOfficial repositoryPinned checkpoint and rewind managerMessage-level rewind, recorded file states, restoration selection, in-place history truncation, and restore-error reportingOfficial repositoryPinned Git worktree implementationOptional per-task branch and worktree creation, reuse, cleanup, and the limits of workspace rather than process isolationOfficial repositoryPinned local-session architectureDurable local session format, append-friendly messages, atomic metadata, resume, rewind, and compatibility constraintsOfficial repositoryPinned CI workflowPre-commit, CLI startup checks, pytest retry policy, and separately enforced snapshot testsOfficial repositoryPinned GitHub ActionFirst-party composite action for installing and invoking programmatic Vibe in CIOfficial repositoryPinned changelogVersion history for approval-default changes, rewind, subagents, MCP OAuth, hooks, OTEL, sessions, and provider supportOfficial repositoryVibe product evolutionOfficial distinction and handoff among terminal, editor, and separately managed remote Vibe surfacesOfficial announcementVibe overviewCoding-product scope, supervised local workflows, remote sandbox option, and common user tasksOfficial docs

Ecosystem discovery

CLIArena Mistral Vibe adapters and Terminal-Bench runs Independent, inspectable discovery signal only. Reported Terminal-Bench 2.0 runs use a modified Mistral Vibe fork and do not yet meet HarnessMatch's full immutable configuration and replication requirements, so no score is imported. Observed 2026-07-27.