Power your agent
The publisher's own URL. Not a PlugMCP endpoint.
Host a copy with our keys and logs.
https://testgraph.21dle.co.uk/mcp-v2The Claude button above opens Claude with the name and URL filled in. You still confirm it there.
claude mcp add --transport http io.github.bbcbasic-testgraph https://testgraph.21dle.co.uk/mcp-v2Run in a terminal.
{ "mcpServers": { "io.github.bbcbasic-testgraph": { "command": "npx", "args": [ "-y", "mcp-remote", "https://testgraph.21dle.co.uk/mcp-v2" ] } }}claude_desktop_config.json
Remote URL or a command on your machine? How the two installs differ
Run on PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
LaunchFrom the publisher
TestGraph is an experimental shared knowledge and experience graph for AI assistants. It lets independent AI clients contribute to the same graph while preserving the original human evidence, model attribution, disagreement, confidence and the path by which a conclusion was reached.
Shared semantic graph for AI reviews, classification and structured memory across AI assistants.
Status: working pre-release research system. TestGraph has been exercised through MCP with multiple AI clients, including cross-model classification, retrieval, assessment and deliberation. It is not yet presented as a stable production service or API.
The central question is not simply whether an AI can remember something. It is:
Can multiple independent AIs accumulate reusable knowledge, disagree without destroying each other's conclusions, and eventually reach justified convergence with an audit trail of why?
A human observation can be interpreted independently by different AI systems:
Human evidence
|
+---- AI A ---- "ferry belongs_to transportation"
|
+---- AI B ---- "ferry belongs_to public transport"
|
v
TestGraph
+-------------+
| evidence |
| provenance |
| confidence |
| disagreement|
| votes |
| resolution |
+------+------+
|
v
reusable shared knowledge
TestGraph does not require those models to use identical words before their work can be useful. A naming disagreement can remain a naming disagreement. A substantive semantic disagreement can remain unresolved and attributable until evidence or an explicit resolution justifies convergence.
The server stores and verifies the process; the calling AI supplies the open-ended semantic reasoning.
In the experimental typed-classification mode, the is_a graph has one reserved synthetic root, .. It gives every semantic branch a common structural origin without making entity or another model-chosen term globally privileged. Classifiers traverse outward from . progressively; they never use it as an ordinary subject type.
You can keep enriching a subject you own, or one attached to your own non-deleted review, while its classification awaits independent review or dispute resolution. This follows your authenticated TestGraph user identity across AI clients. Enrichment preserves the pending classification and its audit trail; it does not count as an independent classification decision. Other contributors must settle an unconfirmed subject's classification before updating it through the MCP mutation tools. Evidence, conflict-preservation and other write validations still apply.
These are architecture and test goals, not merely prompting instructions.
This comparison is about architectural emphasis, not a claim that TestGraph replaces or outperforms those projects.
TestGraph supports two ways to have an identity.
Open:
https://testgraph.21dle.co.uk/account
Sign in with Google. TestGraph creates or recovers the same internal TestGraph user_id each time you return with that Google identity.
From that account you can create as many tg_... capability keys as you need for ChatGPT, Claude, another MCP client or disposable testing.
Google account
|
v
persistent TestGraph user_id
|
+---- tg_ key A ---- ChatGPT
+---- tg_ key B ---- Claude
+---- tg_ key C ---- test client
Capability keys are stored hashed, not in recoverable plaintext. A newly generated tg_ key is shown once. If you want to reuse it later, store it yourself. If you lose it, sign back in with Google and create another key; the underlying TestGraph identity and data remain unchanged.
Google is therefore an optional persistent identity/recovery mechanism, not a requirement for using TestGraph.
Open:
https://testgraph.21dle.co.uk/capability/new
This creates a new standalone TestGraph identity and a private tg_... capability URL exactly as before. Keep it safe: without an external identity attached, possession of that capability is what gives access to that TestGraph identity.
The current remote MCP endpoint is:
https://testgraph.21dle.co.uk/mcp-v2
Use the website versions of ChatGPT or Claude when setting up the connector.
The normal OAuth connection flow is now:
AI client starts OAuth
|
v
TestGraph /account
|
+---- Sign in with Google
| |
| v
| persistent TestGraph identity
|
+---- Use existing tg_ capability
|
v
existing identity
|
v
Confirm "Connect this AI"
|
v
OAuth completes and client receives scoped tokens
The AI client receives OAuth access/refresh tokens. It does not receive your Google credentials or your private TestGraph capability key.
https://testgraph.21dle.co.uk/mcp-v2 as the MCP endpoint.https://testgraph.21dle.co.uk/mcp-v2.tg_ capability, then confirm Connect this AI.Convergence alone is weak evidence. TestGraph is interested in justified convergence: retaining enough information to answer which human evidence started a conclusion, which model proposed it, whether another model independently agreed, whether disagreement was naming or semantic, what supported a vote or counterproposal, and how a resolution was reached.
Current workflows include:
This is ongoing experimental work. The repository deliberately does not claim that cross-model semantic convergence has been solved.
Classification is metadata rather than a storage address. A review of a ferry can remain a ferry review while the graph records:
ferry --belongs_to--> transportation
Another AI may propose a more specific or differently named relationship. TestGraph can retain both contributions and their provenance while the disagreement is examined.
Normal AI workflows do not download the complete classification vocabulary. They first try resolve_subject_type for a direct canonical-name or alias match. If more exploration is needed, they use:
list_root_subject_types to inspect the top-level choices;list_child_subject_types to inspect one level of a promising branch;get_subject_type_path to verify the active parent path of a known type.At each level the AI ranks a small candidate list, follows the strongest branch and retains the others as fallbacks. An inadequate classification or empty scoped retrieval causes it to backtrack rather than declare a global miss. Traversal stops at the most specific adequate existing type, or after enough bounded evidence exists to call resolve_subject_hierarchy with the verified ancestors and only the genuinely missing nodes.
The existing search tool can then search the selected type and its descendants without returning those taxonomy nodes to the model. vocabulary_index remains available as a complete administrative/debugging export, not as a prerequisite for saving or retrieval.
Equivalent REST navigation endpoints are available under /api/v2/subject-types/roots, /api/v2/subject-types/{id}/children and /api/v2/subject-types/{id}/path.
Cross-client testing exposed a practical problem: an AI client can retain an older MCP tool definition after the server has changed. TestGraph exposes server/deployment information and requires a live deployment token immediately before protected write operations. A stale or mismatched connection is rejected before data is changed.
The user-facing error tells the client to refresh or reconnect TestGraph, without assuming the user named the connector “V2”.
Human evidence / experience
|
v
Independent AI clients
(reasoning + discovery)
|
v
MCP/OAuth
|
v
TestGraph server
+--------------------+
| stable identities |
| graph relationships|
| provenance |
| assessments |
| deliberations |
| verification |
| audit history |
+---------+----------+
|
v
PostgreSQL graph data
|
v
reusable by another AI
Implemented foundations include stable subject-type IDs, canonical terms and aliases, editable relationships, versioned schemas, structured validation, OAuth 2.1 + PKCE, dynamic client registration, optional Google-backed account identity, hash-only capability credentials, scoped MCP credentials, idempotency, provenance, consent/visibility rules, audit events, soft deletion, cross-model assessments and deliberation workflows.
Google login is optional. To enable it on a deployment, create a Google Web application OAuth client and register this exact redirect URI:
https://testgraph.21dle.co.uk/account/google/callback
Set these deployment secrets (never commit their values):
GOOGLE_CLIENT_ID=<Google OAuth client ID>
GOOGLE_CLIENT_SECRET=<Google OAuth client secret>
Without those variables, standalone tg_ capability identities continue to work normally.
Requires Python 3.11+.
git clone https://github.com/BBCBasic/TestGraph.git
cd TestGraph
python -m venv .venv
Activate the virtual environment, then:
pip install -r requirements.txt
cp .env.example .env
alembic upgrade head
python -m scripts.seed
python run.py
Open http://127.0.0.1:8000 or http://127.0.0.1:8000/docs.
Replace all placeholder secrets in .env. Never reuse development/example credentials in a public deployment.
pytest -q
Run the deterministic vocabulary scale comparison with:
python scripts/benchmark_vocabulary_traversal.py --nodes 10000 --branching-factor 10
A public release should not be cut unless the complete test suite passes against the release commit and deployment readiness checks succeed. See RELEASE_CHECKLIST.md.
TestGraph includes an importer for the UCI recipe-review dataset:
python -m scripts.import_uci_recipe_reviews --representative-reviews 100 --load
or load the checked bundle:
python -m scripts.import_uci_recipe_reviews --load-bundle data/uci_recipe_reviews_100.json
The reference deployment uses FastAPI, PostgreSQL and Railway. Production deployments should provide unique secrets, a PostgreSQL DATABASE_URL, the public base URL and appropriate host/CORS configuration.
A guarded /development/reset facility exists for development environments and must remain disabled in public production deployments (ENABLE_DEVELOPMENT_RESET=false).
Issues and experimental counterexamples are welcome.
TestGraph is licensed under GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE.
Alternative commercial or proprietary licensing may be available; contact [email protected].
Contributors should read CONTRIBUTING.md.
TestGraph should currently be treated as an experiment rather than established infrastructure. Its purpose is to make cross-model knowledge sharing, provenance and disagreement concrete enough to test. Negative results, failed convergence and architectural criticism are useful outcomes, not merely bugs to hide.
© 2026 MCP Directory.
Shared semantic graph for AI reviews, classification and structured memory across AI assistants.
{ "mcpServers": { "io.github.bbcbasic-testgraph": { "url": "https://testgraph.21dle.co.uk/mcp-v2", "headers": {} } }}Tools
No tool list yet. Remote servers get one after the first health check.
FAQ
Copy the http config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.
No required environment variable is listed on this record.
TestGraph is reached over HTTP.
Last checked 2026-11-08. Shared semantic graph for AI reviews, classification and structured memory across AI assistants.