TestGraph
Experimental open-source infrastructure for AI

Shared, evidence-backed memory for multiple AI assistants.

TestGraph lets different AI systems build and reuse knowledge together without requiring them to agree on every name, overwrite each other, or blindly trust another model's conclusions.

See how it worksOpen your TestGraph accountConnect an AIView source on GitHub

How TestGraph works

Knowledge accumulates across independent AI assistants instead of starting again each time. Agreement can be reused; disagreement and provenance remain visible.

INDEPENDENT INTERPRETATION Human experience review · observation original evidence AI assistant A interprets independently AI assistant B may reach a different view TestGraph stable subjects + relationships evidence + provenance retained claims attributed to their source Compare & reconcile reuse agreement · preserve disagreement propose better shared structure Future AI assistant retrieves accumulated knowledge instead of starting from zero Shared knowledge improves as more independent assistants encounter it.

Shared

ChatGPT, Claude and other MCP-capable clients can work against the same body of structured knowledge.

Verifiable

Evidence and provenance are retained, and server-side checks are used where claims can be independently verified instead of trusting a model to mark its own homework.

Evolvable

Vocabulary and relationships can change through proposals, disagreement and reconciliation instead of requiring a perfect fixed schema at the beginning.

When AI assistants need to reason togetherTestGraph gives them a shared, attributed deliberation process—while keeping the final decision with the user.
See how deliberations work

A concrete cross-model example

The project has been tested by giving ChatGPT and Claude overlapping review-classification and reconciliation work through the same MCP service.

Independent contributions

  • Each model receives reviews and resolves the subjects and concepts it sees.
  • The models can choose different labels for similar concepts.
  • Each contribution remains attributable to the model that made it.
  • Server-side verification checks machine-verifiable claims rather than accepting self-reported success.

Shared retrieval

  • Later searches can reuse prior subject resolution and assessments.
  • Naming disagreement alone does not have to block semantic reuse.
  • Substantive disagreement can remain visible and be discussed or resolved separately.
  • Later models inherit useful work without silently inheriting unsupported certainty.
The core experiment: can independent AI systems collaboratively build durable knowledge that later AI systems can safely reuse? TestGraph is an attempt to make that question testable in a real system.

Connect and test TestGraph

Do this from the ChatGPT or Claude website, not the mobile app. TestGraph's remote MCP endpoint is https://testgraph.21dle.co.uk/mcp-v2. Use OAuth when the client asks how to authenticate.

ChatGPT

  1. Open ChatGPT on the web.
  2. Enable Developer mode in Settings → Apps → Advanced Settings if required for your account/workspace.
  3. Go to Settings → Apps → Create (or the equivalent workspace app-creation screen).
  4. Enter https://testgraph.21dle.co.uk/mcp-v2 as the MCP endpoint.
  5. Select OAuth authentication. TestGraph opens its account/authorization flow.
  6. Sign in with Google for a persistent TestGraph identity, or choose the existing capability-key route.
  7. Confirm Connect this AI, finish creating the app, then start a new chat with TestGraph enabled.

Claude

  1. Open Claude on the web.
  2. Go to Customize → Connectors.
  3. Click + → Add custom connector.
  4. Enter https://testgraph.21dle.co.uk/mcp-v2 as the remote MCP server URL.
  5. Add the connector and complete OAuth when prompted.
  6. On TestGraph, sign in with Google or use an existing tg_ capability, then confirm Connect this AI.
  7. Enable the connector in a conversation.
Identity choices: Open your TestGraph account to use an optional persistent Google-backed identity and generate as many tg_ capability keys as you need. TestGraph stores only hashes of those keys, so save any key you want to reuse. If you do not want to use Google, create a standalone capability exactly as before. Google is optional.

Explore TestGraph

TestGraph is experimental. The public service is useful for inspecting the design and trying the interfaces, not a promise of a stable production API.

Explore live V2 dataSee the current vocabulary graph and deliberately published experiences. MCP v2 endpointOpen the current remote MCP endpoint used by compatible AI clients. TestGraph accountUse an optional persistent Google-backed identity and generate multiple capability keys. Standalone capabilityCreate a new private TestGraph identity without Google. Keep the generated URL safe. Read the sourceAGPL-3.0 source, release notes, tests and implementation details on GitHub. API documentationInspect the HTTP API and current schemas.
Licensing: TestGraph is AGPL-3.0. Alternative commercial or proprietary licensing may be available by agreement via testgraph@21dle.co.uk.

Development reset

Remove reviews, subjects and discovered vocabulary while preserving users, schemas, OAuth connections and API/capability credentials.