Headless infrastructure for agent products
Gooselake is a machine-side runtime for agents that operate on real hosts. It keeps sessions, event history, processes, worktrees, and team coordination durable while any web app, desktop UI, or CLI stays thin.
The architectural bet
01 / BUILD THE DISTRIBUTED SYSTEM ON PURPOSE
Agent products begin as interfaces and quietly accumulate a runtime: a second provider, background execution, resumable sessions, approvals, delivery queues, and another client. Gooselake makes that boundary explicit from the first turn.
The result is one long-lived Rust control plane behind HTTP and SSE. Provider adapters translate. SQLite remembers. Clients render what the runtime can prove.
Read the story behind GooselakeWhat the runtime earns
02 / INFRASTRUCTURE AFTER THE DEMO
Gooselake is valuable because these concerns reinforce each other. The event ledger makes recovery inspectable. Session ownership makes process control and message delivery safe. Worktree claims give team operations a real machine boundary.
Codex, Claude, and ACP map their native lifecycles into one runtime model for sessions, turns, approvals, and terminal state. Real capability differences remain explicit instead of being papered over.
Compare providersImportant transitions land in SQLite with scoped sequence numbers. Clients can resume from a cursor, replay the gap, and follow the live tail without inventing missing state.
Follow the event modelStartup reconciles sessions, turns, approvals, deferred deliveries, processes, and worktree claims. Work either resumes safely or becomes an inspectable failure, never a convincing ghost.
Understand recoveryThe runtime starts commands, enforces ownership and concurrency, records lifecycle, and keeps bounded stdout and stderr logs. Live output is a sample; the log files are authoritative.
Inspect process controlManaged Git workspaces have durable records, explicit claims, cleanup policy, and startup repair. Team spawn treats session, worktree, membership, and onboarding as one journaled saga.
See worktree lifecycleTeam messages produce per-recipient delivery records with policies, deferral, retries, cancellation, and replayable state. Coordination is transport infrastructure, not prompt roleplay.
Open teams and commsThe flight recorder
Every important event is a scoped, sequenced receipt. A client remembers the last cursor it processed, asks for everything after it, renders the backlog, and then follows live state.
Session and process streams subscribe before replay handoff to close the live-event gap. Lifecycle transitions are critical; high-volume output can be droppable because full process logs stay authoritative.
Read events and recoveryHost work and agent teams
03 / OWNERSHIP ALL THE WAY DOWN
An agent that can edit a repository and run a build needs more than a chat transcript. Gooselake makes the host resources and the handoffs between agents explicit, durable, and queryable.
Creation, claims, lifecycle, logs, and cleanup stay separate so the runtime can explain exactly who owns what.
Follow the host workflowsA direct message or broadcast becomes one message plus a delivery record per recipient. Restart the runtime and deferred delivery is still runtime state.
See delivery semanticsThin by design
04 / SPEND COMPLEXITY ON THE PRODUCT
A good client can be killed, restarted, and reconnected without changing the runtime’s understanding of the work.
Documentation flight plan
05 / ENTER FROM WHERE YOU STAND
The manual is organized around how the runtime behaves, not around a list of crates. Pick the route that matches the work in front of you.
Browse the complete manualInstall the server, verify provider readiness, create a session, and watch replayed and live events.
Start with setup↗Get the vocabulary and boundaries for sessions, turns, providers, events, teams, processes, and worktrees.
Read core concepts↗Use records and cursors to build a web app, desktop UI, CLI, or operations console that can reconnect cleanly.
Open the client guide↗Inspect health, diagnostics, recovery summaries, provider state, logs, and host resources before guessing.
Use the runbook↗Control tower, not cockpit
Start locally, create one provider-backed session, and watch the runtime retain its own receipts.
Open the quick startInspect the source on GitHub