Serveur MCP LearnLoop
Commit 17 oct. 2026Évalue et accompagne l'équipe sur ses prompts en temps réel, en créant une bibliothèque de prompts partagée et un wiki pour améliorer l'adoption de l'IA sur Claude Code, VS Code et Copilot.
Aussi connu sous Bogzx/LearnLoop
- Étoiles
- 4
- Outils
- —
- Téléchargements, 30 jours
- —
- Dernière version
- —
- Dernier commit
- 17 oct. 2026
- Licence
- MIT
Qu'est-ce que LearnLoop MCP ?
Évalue et accompagne l'équipe sur ses prompts en temps réel, en créant une bibliothèque de prompts partagée et un wiki pour améliorer l'adoption de l'IA sur Claude Code, VS Code et Copilot.
Résumé généré à partir du README du projet
À propos du MCP server LearnLoop
LearnLoop permet aux développeurs utilisant Claude Code, VS Code ou Copilot Chat d'améliorer l'adoption de l'IA au sein de leur équipe grâce à un accompagnement en temps réel sur l'ingénierie de prompts. Le système évalue chaque prompt selon cinq dimensions spécifiques : la clarté de l'objectif, la spécificité, le chargement du contexte, l'articulation des contraintes et la spécification du format de sortie. En utilisant l'outil de coaching, le serveur fournit un feedback immédiat et des boucles d'apprentissage pour aider les utilisateurs à affiner les prompts de faible qualité avant qu'ils ne soient envoyés au LLM.
Il s'agit d'une solution auto-hébergée que les équipes déploient via Docker et une base de données Postgres. Bien qu'il puisse fonctionner en mode hors ligne avec un système de notation basé sur des règles, l'intégralité des fonctionnalités nécessite une clé API Gemini. Le serveur gère une bibliothèque de prompts partagée et un wiki d'équipe, qui capturent automatiquement les conventions de prompting réussies et les prompts éprouvés. Cela crée une base de connaissances cumulative où le contexte des futurs prompts est enrichi par les standards établis et les workflows documentés de l'équipe.
Ce que vous pouvez faire avec
- 1Coaching de prompts en temps réelUn client IA peut utiliser l'outil de coaching pour lancer une boucle d'apprentissage et de révision afin d'améliorer un prompt de faible qualité.
- 2Capture automatisée des connaissancesLe système peut enregistrer automatiquement les conventions de l'équipe dans un wiki pour construire une base de connaissances partagée.
- 3Création d'une bibliothèque de promptsLes prompts de haute qualité répondant à des critères de notation stricts sont automatiquement ajoutés à une bibliothèque d'équipe partagée.
- 4Prompting contextuelLes développeurs peuvent récupérer les conventions spécifiques à l'équipe et les prompts éprouvés pour fournir un meilleur contexte lors de nouvelles tâches.
Extrait du README du projet. Chaque cas d'utilisation est vérifié par rapport à une citation du document.
Configuration
TRAILHEAD_API_URLTRAILHEAD_IT_DATABASE_URLTRAILHEAD_LLM
| Nom | Description | Type |
|---|---|---|
| TRAILHEAD_API_URL | — | |
| TRAILHEAD_IT_DATABASE_URL | — | |
| TRAILHEAD_LLM | — |
Vérifications automatisées
Exécuté le 14 nov. 2026Rapport complet →Pas encore vérifié. Un serveur nécessite un score de 80, aucune faille sérieuse et suffisamment de données pour que nous puissions effectuer les vérifications. Fonctionnement des vérifications · Ce qu'il faut corriger pour gagner des points
- Dépôt source lié et accessiblebogzx/learnloop is public and reachablegithub.com/bogzx/learnloop ↗Réussi
- Le nom du package correspond au dépôt sourceNon vérifié
- Provenance du build (attestation Sigstore)Non vérifié
- Licence open source (OSI)MITspdx.org/licenses/MIT.html ↗Réussi
- Commit ou version publiée au cours de l'année écouléeLast commit or release 2026-10-17github.com/bogzx/learnloop ↗Réussi
- OpenSSF ScorecardNon vérifié
- Nom du package non proche d'un package populaireNon vérifié
- Aucun avis critique dans le package ou ses dépendancesNon vérifié
- Aucune instruction cachée destinée au modèle0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersRéussi
- Les outils destructifs sont étiquetésNon vérifié
- Le point de terminaison distant répond à la poignée de main MCPNon vérifié
README
Voir sur GitHubLearnLoop
A prompting coach for teams. Every prompt sent through Claude.ai, Claude Code, VS Code or Copilot Chat is scored on five dimensions before it goes out. Weak prompts get a short teaching loop; strong ones join the team's prompt library, and the conventions people state are collected into a team wiki that is fed back into the next person's context. A team's way of prompting compounds without anyone writing docs.
- Try the scorer in your browser: https://learnloop-gules.vercel.app/#try
- Video walkthrough: https://www.youtube.com/watch?v=kD6nnJAmRK8
- Run the whole stack, no API key needed: Try it below
Built at PoliHack v19 (April 2026, BMW track "Applications that encourage AI adoption") by @Bogzx, @KunMihai1, @bbeatricecretu and @CosovanuGabi912. Open source (MIT) and self-hosted: there is no hosted service to sign up for.
Trailhead is the codename used inside the code (package names, settings, env vars); LearnLoop is the product name.
How it works
flowchart LR
subgraph clients["Where people prompt"]
BX["Chrome extension<br/>on claude.ai"]
VS["VS Code extension<br/>sidebar"]
MCP["MCP server<br/>Claude Code · Copilot Chat"]
end
DASH["Dashboard<br/>Next.js"]
subgraph api["apps/api · Hono"]
ROUTES["/score · /coach · /improve · /diff<br/>/wiki/* · /prompts/* · /onboard/*"]
LLM["llm.ts"]
end
PG[("Postgres<br/>wiki · library · score history")]
GEM["Gemini<br/>gemini-3-flash-preview"]
RULES["rule-based scorer<br/>TRAILHEAD_LLM=offline"]
LF["Langfuse<br/>optional"]
BX -- "X-Team-Token" --> ROUTES
VS -- "X-Team-Token" --> ROUTES
MCP -- "X-Team-Token" --> ROUTES
DASH -- "read-only proxy" --> ROUTES
ROUTES --> PG
ROUTES --> LLM
LLM --> GEM
LLM -.-> RULES
GEM -. traces .-> LF
The loop, end to end:
- Score. A prompt is sent (Chrome extension) or a code task starts
(Claude Code / Copilot Chat via the MCP
coachtool). The API scores it on the five dimensions below and records the scores for the team's skill arc. - Coach. At 7/10 or more it goes through untouched. Below 7, the
extension holds the send and offers Improve / Send as-is / Edit;
nothing is sent automatically. In Claude Code the
coachtool runs a short teach → revise loop (at most 5 rounds) on the weakest dimension, using a teammate's proven prompt as the example when there is one. - Promote. A prompt that clears a strict bar (unrounded mean ≥ 7, no dimension below 5, and an independent re-score that agrees) joins the team's prompt library, optionally after a teammate's review.
- Remember. When someone states a team convention,
wiki_saverecords it; repeated mentions are deduplicated and become durable after three. - Reuse. The next prompt about the same part of the codebase is scored and coached with that wiki node's conventions and library prompts in context, so the rubric reads against how this team works.
The rubric, and how the score is measured
Every prompt is scored 0–10 on:
goal_clarity— the desired outcome is stated unambiguouslyspecificity— the change itself is named (files, functions, errors)context_loading— the relevant code, docs or examples are referencedconstraint_articulation— what must not change; limits and invariantsoutput_specification— the shape the answer should take
overall is the rounded mean. The rubric text lives in
packages/scoring/src/score-prompt.mjs, and the score comes from one Gemini
call, so it varies from run to run. apps/api/eval/ measures how much on a
golden set of 30 prompts plus a held-out set of 16, reporting run-to-run
spread, how often the coaching and library gates flip, band hits, and whether
better prompts are ranked above worse ones.
What has been measured so far is the rule-based scorer
(packages/scoring/src/heuristic-score.mjs), a transparent baseline that a
model scorer should beat; the two have not been compared yet. CI checks it on
every change:
| Prompt set | Overall in expected band | Pairs ranked the right way round |
|---|---|---|
| golden (30, visible while its rules were written) | 29/30 | 152/152 |
| held-out (16; per its author, written before the rules were frozen and not tuned on, but added in the same commit, so this can't be checked from history) | 14/16 | 38/38 |
The bands were written in this repo, not by independent reviewers, and the
Gemini scorer has not been run on either set yet (it needs a key; one command
in apps/api/eval/README.md). Known misses and
the full method are documented there.
Try it
In your browser, 10 seconds: the Try it section of the landing page runs the rule-based scorer on whatever you type. Nothing leaves the page.
The whole stack, no key, about 2 minutes (Docker with Compose v2):
git clone https://github.com/Bogzx/LearnLoop && cd LearnLoop
TRAILHEAD_LLM=offline docker compose --profile demo up --build
# API on http://localhost:3000, Postgres schema applied, demo team seeded
curl -s -X POST localhost:3000/score -H 'content-type: application/json' \
-H 'X-Team-Token: trailhead_demo_acme_2026' \
-d '{"prompt":"fix the retry","user_id":"me"}'
Offline mode has no model: scoring uses the rule-based scorer, coaching uses
the static templates, and every response says so. --profile demo loads a
fictional team ("Acme Fintech") with a small wiki, four library prompts and
six days of synthetic activity, so the dashboard has something to show.
With Gemini: cp .env.example .env, put a key from
https://aistudio.google.com/apikey in it, and docker compose up.
Then point the clients at it, all described in SELFHOSTING.md:
# Dashboard on http://localhost:3001 (server-side env; the secret never reaches the browser)
TRAILHEAD_API_URL=http://localhost:3000 TRAILHEAD_TEAM_TOKEN=trailhead_demo_acme_2026 \
npm --workspace=apps/dashboard run dev
# Claude Code / Copilot Chat in one of your repos (the MCP package is not on npm,
# so the CLI runs from this clone; `init` registers the repo's team)
cd /path/to/your/repo
node /path/to/LearnLoop/apps/mcp-server/bin/cli.mjs init
Packages of the Chrome extension (.zip, load unpacked) and the VS Code
extension (.vsix) are built by every CI run (the extensions artifact) and
attached to GitHub releases
from v0.1.0; neither is on a store.
Before exposing the API beyond localhost, read
SELFHOSTING.md → Security model.
What's in the box
| Surface | What it does | Code |
|---|---|---|
| API | Hono + TypeScript on Node 22, raw pg. Multi-tenant: each team has a public id and a server-minted secret (stored hashed). Scoring, coaching, library, wiki, rich bootstrap, rate limits, Langfuse tracing. |
apps/api — routes in its README |
| Chrome extension | MV3, on claude.ai. Score card under the composer, score badges on each message, Compare to team diff, Improve chat, outcome chips, wiki toasts, context picker. Every API call goes through the service worker. | apps/browser-ext |
| VS Code extension | Sidebar with the same score card, team examples for the open file, wiki-update toasts. | apps/vscode-ext |
| MCP server | STDIO server for Claude Code and Copilot Chat: coach, wiki_lookup, wiki_save, wiki_bootstrap, wiki_proven_prompts (+ ping), and a CLI: init, bootstrap, reset. |
apps/mcp-server |
| Dashboard | Next.js 16: team skill arc, metrics, wiki tree, onboarding view; reads through a server-side proxy so the team secret stays on the server. | apps/dashboard |
| Landing page | Static page with the in-browser scorer. | apps/landing-page |
| Shared packages | shared (every request/response type), scoring (prompts, rubric, rule-based scorer, renderers), score-card (DOM score card), db (schema, migrations, demo seed). |
packages/ |
Postgres holds eight tables (packages/db/schema.sql): teams, wiki nodes,
learnings, library prompts, captures, skill observations, and the two
rich-bootstrap job tables.
Development
Node ≥ 22.6, npm workspaces.
npm ci
npm run typecheck # tsc --noEmit in every workspace
npm run lint # ESLint flat config
npm test # every workspace's tests, including the scorer baseline
npm run build # every workspace with a build script
# API integration tests against a throwaway Postgres (they wipe it):
docker run -d --rm --name trailhead-it -p 55432:5432 -e POSTGRES_USER=trailhead \
-e POSTGRES_PASSWORD=trailhead -e POSTGRES_DB=trailhead_it postgres:16-alpine
TRAILHEAD_IT_DATABASE_URL=postgresql://trailhead:trailhead@127.0.0.1:55432/trailhead_it \
npm --workspace=apps/api run test:integration
CI runs all of that, plus npm audit, a Docker build with a
docker compose up smoke test in offline mode, and packaging of both
extensions. A v* tag attaches the extension packages to a draft release
(.github/workflows/release.yml).
Without Docker, the API runs against any Postgres:
psql "$DATABASE_URL" -f packages/db/schema.sql, optionally
node packages/db/seed.mjs, then npm run dev (reads .env).
Deployment
- API: self-hosted.
docker composeas above, or any Node 22 host (railway.jsonis kept for Railway). Every environment variable is in SELFHOSTING.md → All variables. - Dashboard: Vercel or any Next.js host. Set
TRAILHEAD_API_URLandTRAILHEAD_TEAM_TOKENserver-side; anyone who can open it can read that team's data, so restrict access. - Landing page: static files on Vercel.
- Extensions: from the release packages, loaded unpacked / installed from
the
.vsix.
Specs
The project was specced before it was built; the specs are the record of why:
docs/superpowers/specs/2026-04-25-trailhead-design.md— master specdocs/superpowers/specs/2026-04-25-mcp-plugin-ux-design.md— MCP install story and (then) four-tool surfacedocs/superpowers/specs/2026-04-25-trailhead-browser-ext-design.md— Claude.ai content-script architecturedocs/superpowers/specs/2026-04-25-demo-completion-design.md— dashboard and seeding plandocs/superpowers/specs/2026-04-26-trailhead-educational-loop-design.md— the teach → reveal coaching loopdocs/superpowers/specs/2026-04-26-wiki-bootstrap-rich-design.md— async rich bootstrapdocs/superpowers/specs/2026-04-26-improve-widget-design.md— multi-turn improve widgetdocs/roadmaps/— per-surface 24-hour build plans
Each app and package has its own README with surface-specific details.
License
MIT — see LICENSE.
Questions sur LearnLoop MCP
Comment fonctionne la boucle d'accompagnement dans Claude Code ?
Dans Claude Code, l'outil de coaching exécute une boucle d'apprentissage et de révision allant jusqu'à cinq cycles sur la dimension la plus faible d'un prompt.
LearnLoop est-il un service hébergé ?
Non, il s'agit d'un produit open-source à auto-héberger, sans service d'abonnement hébergé.