Skip to content

Platform

The brain service, oa the CLI, and the inbound gateway they share.

  • Quick start: A brain against a local Postgres, in about five commands.
  • Configuration: Every environment variable, its default, and whether it is used yet.
  • oa, the CLI: Every command and subcommand, and the flags that apply to all of them.
  • HTTP API: Every endpoint the brain serves, and how a caller authenticates.

What runs today

The brain runs agents. A run starts from oa runs start or the HTTP API, executes inside sdk/agent against the team's model, and streams every step back as an event; the dashboard shows each run's trace.

  • Agents: a model, a reasoning pattern (ReAct, plan-then-act, ReWOO or reflection), the MCP servers and skills they are granted, who they may delegate to, and whether they may use the web.
  • The hub: every team has a front door. It is the built-in orchestrator unless the team names one of its own agents, and it can plan, find an agent in the graph and hand it the work, search the web and read a page.
  • The graph: each team's agents, MCP servers, tools and skills, projected from Postgres into FalkorDB. A run's graph tool searches it, scoped to what the caller can reach before anything is scored.
  • Skills and MCP servers: uploaded, imported from a repository and kept in sync; servers discovered for their tools, with their credentials in the secret store.
  • The dashboard, served by the brain at /ui: teams and members, agents, skills, MCP servers, channels, sessions with each run's trace, the graph and the team's settings.
  • Underneath: OIDC authentication, role-based authorisation where roles and their permissions are rows, an inbound gateway that verifies each channel's webhook against its own signing secret, versioned migrations, and health and readiness probes.

What does not exist yet

  • Answering a channel. A message that arrives on a channel is stored in its session, and nothing starts a run from it, so nothing replies. Runs start from the API and the CLI.
  • Slack, Discord and Telegram adapters. The seam they plug into is written, and custom is a working generic webhook in the meantime.
  • Learnings in the graph. It holds agents, servers, tools and skills; what a run learned is not projected into it.
  • A sandbox. code is a pattern name the runner refuses, because running model-written code needs process confinement that is not written.