Introduction
Subako is the runtime your product's AI agents run on — long-running sessions, secure code execution, credentials, and skills, reachable from one API.
A model on its own answers a question and stops. Everything that turns that answer into work — keeping a run alive past a single HTTP request, holding the conversation so a dropped connection can pick it back up, executing code the model wrote, reaching a real system with a real credential — is runtime. It is also the part every team ends up building for itself, once per product.
Subako is that runtime, run as a service. Your product creates a session, streams its events, and answers the tool calls only it can answer. Subako holds everything else.
The pieces
Agents
A model, a prompt, and the grants that go with them — published as immutable versions a session binds to.
Sessions
One unit of work, as an append-only event log you can read, tail, and resume.
Skills
Bundles of knowledge and procedure, versioned in a workspace and granted to an agent version.
Vaults
Where a session’s tool credentials live. The broker uses them; the model never sees them.
How a session runs
You create a session
A session names an agent and, optionally, the vaults it may draw credentials from. It binds the agent’s current version and never re-binds — a version published later does not change a session already running.
You append input
A user message is appended to the session’s log. That starts a run: the engine is claimed, handed the committed log, and pointed at the model proxy and the tool broker for that run alone.
The run produces events
Assistant messages, tool calls, tool results, approvals — every one is appended to the log with a sequence number, then streamed. Nothing about a run lives only in a connection.
Your client answers what it can
A client registers the tools it offers and answers the calls routed to it. Tools that reach an external service go through the broker instead, which attaches the credential.
The run finishes — or you reconnect
Runs end in run_completed, run_failed, or run_cancelled. If your
connection dropped somewhere in the middle, you reconnect with the last
sequence number you saw and the log replays from there.
What ships on day one
The CLI and the API. There is no web console: everything an operator does is a subako command, and everything your product does at runtime is an HTTP call against the Core API.
That is a deliberate ordering, not a gap to work around. Workspaces, agent versions, skills, vaults, and API keys are all configuration, and configuration that lives in a terminal is configuration you can put in a script, a Makefile, or CI.