io.github.Dgotlieb/verdict-mcp
Sandboxed test runs for coding agents: compact typed verdicts with failure fingerprints and history
Installation
Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).
verdict-mcpclaude mcp add verdict-mcp -e 'VERDICT_PROJECT=<VERDICT_PROJECT>' -- uvx verdict-mcp==0.1.0a4
{
"mcpServers": {
"verdict-mcp": {
"command": "uvx",
"args": [
"verdict-mcp==0.1.0a4"
],
"env": {
"VERDICT_PROJECT": "<VERDICT_PROJECT>"
}
}
}
}
Variables d'environnement
VERDICT_PROJECTPath to the project to verify (defaults to cwd)
Documentation
Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.
verdict
https://github.com/user-attachments/assets/62923912-98aa-4246-8bc8-4adf55ae3ff8
Structured, sandboxed verification feedback for coding agents.
An MCP server that replaces your agent's pytest shell-outs with something built for the agent inner loop: impact-selected tests, run in an isolated environment, returning compact typed verdicts instead of 40,000 tokens of raw runner output — with failure fingerprints that tell the agent whether a failure is its regression or was broken all along.
raw pytest dump: ~40,000 tokens, unstructured, run un-sandboxed on your machine
verdict: ~400 tokens, typed JSON, run in a rootless container, with memory
Why
The highest-frequency tool call in agentic coding is verification — and it's the least structured. Agents re-run whole suites when one module changed, burn context parsing ANSI-coded tracebacks, run arbitrary code directly on your machine, and routinely misdiagnose pre-existing breakage as their own regression (then "fix" code that wasn't broken). verdict fixes all four.
Tools
| Tool | What it does |
|---|---|
verify(scope?, base?) | Selects tests affected by your working-tree diff (static import graph via grimp), runs them via podman/docker with the worktree mounted read-only, returns typed failures with fingerprints and a preexisting flag |
explain_failure(check_id) | Full traceback for one failure, on demand — bulk never rides in the summary |
history(fingerprint) | First seen / last seen / times seen — regression vs. long-standing breakage |
run_checks(["ruff","mypy"]) | Lint and type checks, normalized into the same verdict schema |
Every failure carries a fingerprint: a stable hash of the normalized failure signature (volatile tokens — addresses, tmp paths, ids, durations — collapsed). Same logical failure, same fingerprint, across runs and refactors. Fingerprints are what give verdict memory.
Quickstart
No install step needed — uvx fetches it on first use. (Or uv tool install verdict-mcp / pip install verdict-mcp for a permanent verdict-mcp command.)
Claude Code — .mcp.json in your project root:
{
"mcpServers": {
"verdict": {
"command": "uvx",
"args": ["verdict-mcp"],
"env": { "VERDICT_PROJECT": "." }
}
}
}
Cursor — same shape in .cursor/mcp.json.
Optional verdict.toml in your repo root:
[project]
packages = ["your_package"] # for impact selection (auto-guessed if omitted)
[runner]
image = "ghcr.io/you/yourproj-test" # prebuilt env with your deps
setup_cmd = "pip install -e .[test]" # or install on the fly (runs with network; tests don't)
# prefer = "local" # escape hatch if you have no container runtime
[limits]
max_failures = 10
Try it without an agent:
cd examples/demo_project
VERDICT_PROJECT=. verdict-mcp # then connect any MCP client, or use the MCP inspector
Sandbox posture (v0.1)
Checks run in an ephemeral container (podman preferred, docker fallback): worktree mounted read-only at /src, copied to a writable /work inside the container, --network=none for the check run. Your host environment is never mutated by a test run. If setup_cmd is configured, that step runs with network before the check; prefer a prebuilt image for a tighter posture. No container runtime → explicit prefer = "local" fallback runs checks against a temp copy of your worktree (still never in place). See SECURITY.md for the full threat model and known limitations.
Troubleshooting: if a verdict says container engine 'podman' could not start the check, run the suggested podman pull <image> by hand — the engine's own error is the answer. One known trap on macOS: a "credsStore": "gcloud" line in ~/.docker/config.json makes podman call the gcloud credential helper for every registry, including docker.io; an expired gcloud login then breaks all pulls. Fix with gcloud auth login or remove that line.
Honest limitations
- Impact selection uses the static import graph — approximate by design. Dynamic imports, fixture-by-name resolution, and data-driven tests can be missed;
verify(scope="all")is always available and verdict says inselection_notewhenever it falls back. - Python/pytest only today, plus ruff/mypy. The adapter interface is small and documented — vitest and
go test -jsonadapters are the most-wanted contributions (CONTRIBUTING.md). - Flake detection and coverage-map-based selection are v0.2 (roadmap).
Roadmap
v0.2: coverage-based impact maps (precise selection), flake detection via fingerprint alternation, devcontainer.json support, result cache keyed on (tree hash, check, image digest). Later: vitest/jest, go test, cargo test adapters; per-repo verdict daemon mode.
License
Apache-2.0
Questions fréquentes
Qu'est-ce que io.github.Dgotlieb/verdict-mcp ?
Sandboxed test runs for coding agents: compact typed verdicts with failure fingerprints and history
Comment installer io.github.Dgotlieb/verdict-mcp ?
Il s'exécute sur votre machine. Dans Claude Code : claude mcp add verdict-mcp -e 'VERDICT_PROJECT=<VERDICT_PROJECT>' -- uvx verdict-mcp==0.1.0a4. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.
Faut-il une clé API pour io.github.Dgotlieb/verdict-mcp ?
Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.
io.github.Dgotlieb/verdict-mcp est-il open source ?
Oui. Son code est sur GitHub sous licence Apache-2.0.
Qui maintient io.github.Dgotlieb/verdict-mcp ?
Il est publié par Dgotlieb. AI.tn le référence sans le maintenir ni le recommander.
Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.