# doublegate

> doublegate is gated memory for AI agents. What an agent writes goes into a named work area, where it is held and scanned. No other agent's recall returns it until an identity that did not write it has reviewed and signed it. Every promotion is that gate's own and is marked provisional. Recall with no scope searches the caller's work area, then each parent, and stops at the first level that answers. Ask with scope `space` to search the whole space. By default a grading model reads each held claim twice, for and against, and those two readings can promote it with no person. An operator can require a person for any kind of content. Next: the Solo release of today's gate, installable, backed up and restorable, with rotating signing keys and erasable content. Then Team, one gate a group shares over the network, under one review policy.

## The API, as data

Rendered from the code that serves it and checked against it in the code repositories' continuous integration; published here as a snapshot stamped with the commit.

- [api/mcp-tools.json](https://about.doublegate.io/api/mcp-tools.json): the tool schemas of the gate's built-in plugin, which is being retired, as its `tools/list` answers them; the Claude Code plugin's tools are not in this file
- [api/daemon-rpc.json](https://about.doublegate.io/api/daemon-rpc.json): every verb on a gate's local socket — parameters, which role serves it, the proof each needs; a running gate answers the same to `dg.describe`
- [api/openapi.json](https://about.doublegate.io/api/openapi.json): OpenAPI for the gate's HTTP surface on this machine; it answers on loopback only, with a sign-in or an API key
- [api/note.schema.json](https://about.doublegate.io/api/note.schema.json): JSON Schema of a published note's `data`: the closed field set a subscriber may rely on — which claim, what happened, what kind; never the text
- [api/index.json](https://about.doublegate.io/api/index.json): which commit of which repository each document was rendered from

## If you are an agent wiring yourself in

- Claude Code: the `doublegate-agent` plugin gives an agent thirteen tools: remember, record, recall, context, workarea, lineage, workarea_relate, complete, status, ask, cancel_question, questions_from_crews and answer_question. It offers no act above an agent's. Over the gate's local socket, a process running as the gate's operating-system user is the gate's administrator, so the gate itself holds an agent to these acts only when the agent runs under another operating-system user.
- The gate's built-in plugin, which is being retired, namespaces its tools `doublegate.*` — `doublegate.remember`, not `remember`. Launch it with `doublegate-client mcp` (this machine's gate) over stdio and call `tools/list`.
- Every write names a work area. `remember` returns a handle, never a promise of promotion; nothing you remember is readable to anyone else through `recall` until a reviewer other than its author has promoted it. Your own held claims come back to you marked `held`.
- Never send `writer_identity` or `deployment_id`; identity is derived from the connection and a request carrying it is refused.
- There is no promote, sign, demote, delete or trust-class tool on any tier. The thing being gated does not operate the gate; a person, or a reviewer who is not the author, does that on the gate's own socket or in its console.
- Every hit carries `provisional`: true while no second authority has countersigned, which on one gate is every hit. The flag is shown and never changes the order of hits.

## Pages

- [Home](https://about.doublegate.io/): What an AI agent writes is held and scanned, and other agents recall it only after a reviewer other than its author signs it. The gate, its console and a Claude Code plugin are available today.
- [For business](https://about.doublegate.io/for-business.html): Keep the corrections, conventions and skills a company's agents learn as reviewed claims that any agent on the gate can find. Who runs it, what it costs, and what a group gets next.
- [For engineers](https://about.doublegate.io/for-engineers.html): Run the gate, connect Claude Code through its plugin, review in the console, and check files offline with the SDK. Interfaces, running costs and the release sequence.
- [How it works](https://about.doublegate.io/how-it-works.html): Follow one correction from the moment an agent writes it to the moment another agent recalls it. See each stage, from the work area to recall, and what happens when a step refuses it.
- [Evidence](https://about.doublegate.io/evidence.html): The research behind the problem, and the rules the gate enforces and how to check each.
- [Business governance](https://about.doublegate.io/governance.html): Who wrote each claim, who reviewed it and why, in a signed journal an auditor can verify. Quoted references to the ECB RDARR Guide, EU AI Act Article 12, GDPR and BCBS 239.

## Optional

- [Site source](https://github.com/doublegate-io/doublegate-io.github.io): how these pages are built and the checks they pass
- [Questions, objections and corrections](https://github.com/doublegate-io/doublegate-io.github.io/issues)