Skip to main content

​
What it is

  • A platform for AI workforce + human governance, not a chat product.
  • Every prompt becomes a Redis-backed job. Every step emits a typed event. Every escalation is logged. Every paid call is cost-tracked.
  • The design centre is throughput, audit, and multi-agent governance.

​
Components

  • Mother AI — stateless ingest service (Rust/Axum). Authenticates requests, persists projects/workflows, enqueues jobs, streams events.
  • Worker — orchestrator (Python/LangGraph). Pulls jobs, runs the context engine, dispatches through an L1–L4 agent hierarchy, emits typed events.
  • Auth — identity and authorization service (Hono + Better Auth). Owns sign-in, sessions, roles, budgets, and all auth database access; the frontends hold no credentials.
  • Frontend — dashboard (Next.js / React). Renders live job streams, projects, workflows, and a 3D knowledge atlas.
  • Ingest — knowledge-hub pipeline. Reads markdown, chunks, embeds, upserts into Qdrant.
  • Knowledge Hub MCP — read-only stdio MCP server. Exposes the knowledge hub to any MCP client (Claude Desktop, Claude Code, agents, partner tools).
Deeper component pages: Scenarios runtime (the declarative flow engine), Intent classifier, Connectors, Job telemetry, QA answer panel, Agent workbench, Composer model picker, L4 ask-user, External tools platform, Diaries, and Notion ingest.

​
Core flow

  1. A user signs in through the Auth service; every app host shares the session cookie, and the frontend’s BFF resolves identity plus budget/ownership checks against Auth’s preflight guard.
  2. Client submits a prompt to Mother AI.
  3. Mother AI authenticates, persists project metadata, enqueues a job in Redis.
  4. Worker pulls the job, runs the context engine, classifies, dispatches through L1–L4.
  5. Worker emits typed AgentEvent JSON to a Redis stream.
  6. Frontend (or any client) subscribes via Mother AI’s GET /v1/jobs/:id/stream and renders backlog + live events.
  7. Worker persists job state, audit trail, and escalation history.
The event envelope and every payload contract are documented in HTTP API & event stream; storage layout in Data model; runtime configuration in Configuration.

​
Hierarchy

  • L1 — Team Lead: orchestration only — classify, plan, delegate.
  • L2 — Managerial roles: Architect, Tech Lead, Release Manager, QA / Security / Adversarial leadership.
  • L3 — Acceptance: verifier; pass/fail with deltas, not rewrites.
  • L4 — Executor: file-write specialist; emits net-new files and surgical edits.
Which model serves each level is decided per level, not per job. The budget a user picks on /new — Free, Cost-optimized (default), or Frontier — resolves to a band per role: Frontier buys a frontier architect and executor while planning and review stay cheap; Cost-optimized runs every role on cheap models; Free rides the quota-paced free pool. Within a band, capability requirements filter first, then the strongest (or cheapest, for short well-structured work) model wins. When a routed model’s request fails, dispatch walks the band strongest-first before any provider default. Same-level conflicts retry up to 3 times before escalating upward. The full mechanism — tier profiles, bands, failover, the effort ladder, and the free-quota ledger — is in Model routing.

​
Models

Which models exist and what they can do is data, in config/model_catalog.yaml; whether a team may use one is config/team_controls.yaml. Neither is a code change, and GET /v1/models serves the merged answer.
  • Free: hosted free tiers, reached through the Groq and OpenRouter routers (openai/gpt-oss-120b, openai/gpt-oss-20b, minimax/minimax-m3:free, z-ai/glm-5.2:free). Free means a quota, not free hardware — these fail by refusing, so a shared Redis ledger paces them across workers and a model whose window is spent is skipped rather than tried.
  • Paid: Claude (claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5), DeepSeek (deepseek-v4-flash, deepseek-v4-pro), GLM (glm-5.2), Kimi (kimi-k3), MiniMax (minimax-m3). Ten of the eleven providers speak the same HTTP shape, so adding one is usually a catalog row and no code.
  • Cost tiers are free → cheap → mid → frontier, and selection is cheapest capable first — a cheaper model that cannot do the work is not a saving. A tier choice resolves to a band of models (frontier currently: claude-opus-4-8, claude-opus-4-7, kimi-k3, glm-5.2), strongest first.
  • Model fallback: when a routed model’s request fails, dispatch walks the band — next model on the same provider, then the band’s other providers — before any default or a level failure. A refused credential (401/403) skips the rest of that provider’s candidates in one step.
  • Provider fallback is health-ranked, and tells a spent quota, an unfunded account and a missing capability apart, because the right response to each differs.
Self-hosted inference (the Ollama stack and its GPU node) is being retired — it is no longer a model source anything routes to on purpose.

​
Integrate

  • Submit a job: POST /v1/chat returns a job_id. Subscribe via GET /v1/jobs/:id/stream.
  • Resume an interrupted job: POST /v1/jobs/:id/resume for jobs paused on an ask_user interrupt.
  • Drive retrieval externally: install the Knowledge Hub MCP server and add it to Claude Desktop or Claude Code.