TestGraph
TestGraph is a shared structured knowledge and review graph for AI agents. Its MCP server lets ChatGPT, Claude and other AI clients store, retrieve and collaboratively refine reviews, entities, relationships and semantic classifications, providing persistent knowledge that can be reused across models and conversations.
README
TestGraph
Experimental AI-native experience graph for shared, verifiable memory across AI assistants.
Status: working pre-release research system. The architecture and MCP workflows have been exercised with multiple AI clients, but the project is not yet presented as a stable production service or API.
TestGraph stores human reviews and AI-derived structure as durable graph knowledge while keeping evidence, provenance and server-verifiable execution separate from model claims. It is designed so that different AI systems can contribute to and reuse the same knowledge without requiring a complete domain schema in advance.
Project goals
- Schema emergence: Give several AIs unfamiliar experiences and obtain a useful structure without designing the categories beforehand.
- Controlled disagreement: Conflicting classifications converge through evidence, confidence and server rules instead of flip-flopping.
- Truthful execution: Models cannot claim that discovery, enrichment or reconciliation happened unless the server can verify it.
- Calling-AI capability: TestGraph deliberately uses the calling AI as its semantic and discovery engine. The AI should apply its available reasoning, retrieval and tool capabilities to unfamiliar subjects, derive useful structure and relationships, and reconcile evidence without waiting for TestGraph to prescribe a domain-specific form. TestGraph provides stable graph primitives, persistence and server-side verification; the calling AI provides the open-ended intelligence.
These goals are acceptance criteria for TestGraph's architecture and tests, not merely guidance for individual AI clients.
Standard vocabulary model
Reviews are stored against stable subject_type_id values, not DNS-style concept paths. Flexible input is resolved through canonical subject types and globally unique aliases; case, punctuation, possessives and ordinary plurals are normalised mechanically. Unknown types may be created as provisional entries after dictionary lookup.
Classification is separate metadata. For example, ferry belongs_to transportation improves broad transportation searches but never changes where a ferry review is stored. review is the record type, not a vocabulary node. Reusable structured fields have their own stable IDs and aliases and may be attached to multiple subject types.
Migration 0009_flat_standard_vocabulary deliberately discards the old v2 concept/review data while preserving users, OAuth state, capability credentials and other authentication data.
Local setup
- Create a Python 3.11+ virtual environment.
- Install dependencies:
pip install -r requirements.txt
- Copy
.env.exampleto.envand replace every placeholder secret with a private value. - Run migrations:
alembic upgrade head
- Seed schemas and demo identities:
python -m scripts.seed
- Start the application:
python run.py
Open http://127.0.0.1:8000 and http://127.0.0.1:8000/docs.
Do not reuse example or development credentials outside a local development environment. Secrets, API keys, OAuth connection codes and owner identifiers must never be committed to the repository.
MCP and OAuth
TestGraph includes a tool-only MCP app. The current multi-model integration is exercised through /mcp-v2. Its production connection uses OAuth 2.1 Authorization Code + PKCE. A connected AI receives a short-lived scoped token; it does not receive the connection code, API key or TestGraph owner ID.
The MCP surface includes review search/fetch/save operations and the newer graph/reconciliation capabilities used by multi-model experiments. Treat the deployed MCP schema as authoritative because this experimental surface is still evolving.
Before a deployment, configure a long random OAUTH_CONNECTION_CODE, set the appropriate OAUTH_OWNER_USER_ID, and ensure production secrets exist only in the deployment environment.
Import the open UCI recipe reviews
After migrations and python -m scripts.seed, run:
python -m scripts.import_uci_recipe_reviews --representative-reviews 100 --load
This downloads the CC BY 4.0 UCI dataset and chooses one evidence-rich review from each of its 100 recipes. It writes the converted records to data/uci_recipe_reviews_100.json and loads them into the configured database. Re-running it is safe: stable source IDs prevent duplicates.
The original review text, 0-5 star score, timestamp, votes, source record ID, licence and attribution are preserved in provenance. The importer interprets explicit statements about flavour, clarity, timing, ingredient availability, difficulty, repeat-worthiness and modifications. Each interpretation retains its supporting source sentence; anything unsupported remains null.
To load the checked bundle without downloading or regenerating it:
python -m scripts.import_uci_recipe_reviews --load-bundle data/uci_recipe_reviews_100.json
Development data reset
A guarded reset page is available at /development/reset. It is hidden and returns 404 unless explicitly enabled:
ENABLE_DEVELOPMENT_RESET=true
The page permanently removes v1/v2 review and knowledge data while preserving users, schemas, OAuth connections and capability credentials. It must remain disabled in a public production deployment.
Tests
pytest -q
A public release should not be cut unless the full test suite passes against the release commit and the deployment readiness check succeeds. See RELEASE_CHECKLIST.md.
Railway deployment
- Push the repository to GitHub.
- Create a Railway project from the GitHub repository.
- Add PostgreSQL.
- Reference the Postgres service's
DATABASE_URL; do not paste an unresolved Railway reference as a plain string. - Configure production variables, using unique random secrets:
ENVIRONMENT=production
APP_SECRET=<random secret>
DEVELOPMENT_API_KEY=<long random key; development/admin use only>
CLIENT_API_KEYS={}
OAUTH_OWNER_USER_ID=<owner UUID>
OAUTH_CONNECTION_CODE=<long random connection code>
PUBLIC_BASE_URL=https://<your-domain>
ALLOWED_HOSTS=["<your-domain>","<railway-domain>"]
CORS_ORIGINS=["https://<your-domain>"]
ENABLE_DEVELOPMENT_RESET=false
railway.json configures pre-deploy migrations, Uvicorn startup using Railway's $PORT, and readiness checking at /health/ready.
The application converts Railway's postgresql:// URL to SQLAlchemy's postgresql+psycopg:// form because the project uses Psycopg 3. A production deployment that cannot resolve PostgreSQL should stop rather than silently fall back to SQLite.
Architecture implemented
- Calling-AI capability as the open-ended semantic and discovery engine, with the server as verification/persistence layer
- Stable flat subject-type IDs
- Canonical terms and aliases
- Editable type relationships used for search rather than storage addresses
- Versioned schema registry
- Domain-specific Pydantic validation stored as JSON/JSONB
- Draft-first publication with explicit approval/version
- Scoped client credentials
- Central read policy
- Canonical subject resolution and version-checked draft editing
- OAuth 2.1 Authorization Code + PKCE, dynamic client registration, refresh-token rotation and scoped access tokens
- MCP endpoint for authenticated AI access
- Idempotency keys
- Provenance, consent, ownership and visibility
- Pairwise reviewer-reader alignment and reader-specific relevance
- Audit log and soft deletion
- Request IDs, consistent JSON errors, request-size limits and pagination-ready endpoints
- Alembic migrations
- SQLite locally; PostgreSQL in production
- Cross-model reconciliation and server-recorded deliberation/assessment workflows
Safe read and editing rules
- Public lists return only experiences that are both
publishedandpublic. - Exact IDs may also retrieve a published
unlistedexperience. - Drafts and private experiences require an appropriate read credential.
aggregate_onlyexperiences are never returned as individual reviews.- Resolve subjects before editing or attaching structured knowledge.
- Version checks protect concurrent draft editing.
- Optional client credentials must use unique, revocable secrets and the minimum scopes required.
Licence
TestGraph is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE.
The AGPL permits use, modification and redistribution subject to its terms, including its network-source obligations for modified versions used to provide a network service.
If AGPL-3.0 does not meet your requirements, alternative commercial or proprietary licensing may be available. Contact testgraph@21dle.co.uk to discuss a separate licence agreement.
Contributors should read CONTRIBUTING.md. Contributions are accepted only on terms that preserve the project's ability to offer alternative licences.
Before making the repository public
Read RELEASE_CHECKLIST.md and SECURITY.md. In particular, complete the secret-history review, confirm the public-data boundary, run the complete test suite, and verify a clean deployment from the exact release commit.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.