Recall, in one page.
Recall is a hosted website that publishes your memory as a WebMCP tool surface. Your ChatGPT agent reads and writes that memory through the browser — origin-scoped, capability-token authenticated, and fully audited.
Try the tools
The interactive WebMCP Tool Playground lets you call each of the six tools against an in-memory demo vault — the same response shape ChatGPT would receive, with the same audit-trail provenance. No sign-in required; no data leaves your browser.
The tool surface
Recall registers six tools via document.modelContext.registerTool() when a signed-in user opens the app in a WebMCP-capable browser (ChatGPT in-app browser, or Chrome 149+ with the origin-trial flag). Each tool is annotated per the WebMCP spec.
| Tool | Read-only | Untrusted content |
|---|---|---|
| query | yes | no |
| addFact | no | yes |
| updateFact | no | yes |
| forgetFact | no | no |
| summarize | yes | no |
| timeline | yes | no |
The security model
- Browser-mediated. Tool handlers execute in the page's existing sandbox; there is no out-of-band channel.
- Origin-scoped. Tools are exposed only to the agent origins the user grants (default:
https://chatgpt.com). - Capability-token authenticated. Every tool call presents a short-TTL, audience-restricted, scope-limited token signed with the user's site key.
- Audit-logged. Every call is appended to an immutable, signed log; the user can export and verify it independently of the database.
The runtime
Recall runs zero LLMs of its own. The agent is ChatGPT — already running in the user's browser. Recall's job is to publish the right tools with the right schemas; ChatGPT does the reasoning, the tool selection, and the synthesis. This keeps the demo reliable (no second model to fail) and the cost zero.
The HTTP API lives at /api and is self-describing. The health check lives at /health.
The full architecture and rejected alternatives are documented in the repository README.
Read the README