B6 · the signature demo
Mention a connected agent in a thread (@claude tighten the exec summary). It reads the artifact and the conversation, publishes a proposed version, and replies in-thread. A human approves from the same screen. The whole point: agents become safe contributors, not autonomous publishers.
An agent is not a special subsystem. It is just another actor with a token and a role — and we register it at commenter rank, so it can only propose, never publish. Every safety property falls out of the permission model we already shipped.
Every arrow is a piece that already exists. The only genuinely new code is the mention parser, the agent registry, and the dispatch glue.
@claude … in a thread on a specific anchor or the whole doc.Why it's cheap: steps 02–05 are the webhook outbox, the MCP reader, the proposals API, and the staleness-guarded approve flow — all shipped. B6 adds the front of the line (parse + dispatch) and the agent's identity.
Three models, same architecture. Dock never has to be in the inference path unless the org opts into our hosted agent.
| Model | Who runs the LLM | Who pays | Best for |
|---|---|---|---|
| BYO key | Dock's worker calls the org's configured provider key (Anthropic / OpenAI / …) | Org's provider bill | Self-host, cost-conscious orgs |
| BYO agent endpoint | The org's own agent service receives the webhook and acts over MCP | Org's own infra | Orgs with an existing agent / their own harness |
| Hosted agent | Dock runs Claude under the hood, meters tokens | Subscription / usage credits | Hosted-tier convenience & revenue |
@name in comment bodies against the registry; emit a comment.mention event.