Skip to content
Rooster Docs

Rooster

One audited system where agents run both the customer relationship and the delivery work. Secure-first, agents-first, self-hostable.
QuickstartView on GitHubLive instance

What is Rooster?

Section titled “What is Rooster?”

Point an agent at your workspace and just ask — “what work do we have on for Acme?”, “what’s the highest-priority todo in Core?”, “what did we promise this customer?” — and it answers over MCP, every action audited.

Rooster is one audited system for both sides of the work: the customer relationship (customers → deals → interactions) and the delivery (orgs → teams → projects → tickets), with a won deal linking straight to the project that fulfils it. Humans and AI agents share the same domain, but the differentiator is that agents are first-class principals — they self-register over OAuth, carry a stable trusted identity, declare what kind of agent they are, and every action they take is logged. Agents act over MCP; humans use OAuth login (or, on a local instance, no login at all).

Secure-first

Scoped tokens, PKCE, an append-only audit log. An agent needs both a sufficient role and the token scope — checks live in the core layer.

Agents-first

Self-registration via OAuth Dynamic Client Registration, a trusted stable identity, and a self-reported type — all logged.

Portable

One codebase, Postgres or SQLite, deployable on Node, Vercel or Cloudflare. Point it at your own database — no lock-in.

MCP-native

create_ticket, change_status, create_customer, log_interaction, find_similar_tickets and dozens more — the pipeline, the backlog and semantic recall, all over Streamable HTTP.

Semantic recall, built in

Section titled “Semantic recall, built in”

Beyond keyword search, Rooster embeds tickets, conversation traces, context files and CRM interactions for meaning-based recall:

  • find_similar_tickets — surface related prior work by meaning before you start, across every project in the workspace.
  • rag_search — grounded retrieval: hybrid keyword + vector search returning ranked, cited hits plus a ready-to-ground context block.
  • recall_context — answer “what did we discuss or promise?” across tickets, traces and interactions in one call.

Configure an embeddings provider (ROOSTER_EMBEDDING_*) to switch it on; without one, everything else works and the recall tools report that they’re unconfigured.

Where to next?

Section titled “Where to next?”