Use Colossus¶
Colossus supports quick one-shot requests and long-running, restart-safe work. Both use the same model routing, tools, authorization, hash-chained journal, and recovery semantics.
Connect a model provider¶
Model-backed workflows need one explicit provider route. Start with Connect a model provider, then choose the focused guide for your access method:
| Access method | Setup guide |
|---|---|
| ChatGPT/Codex subscription | Codex or ChatGPT subscription |
| OpenAI public API key | OpenAI API |
| OpenRouter API key | OpenRouter |
| Loopback model server | Local models |
| Hosted gateway or another compatible API | Other OpenAI-compatible endpoints |
Choose a working style¶
| Need | Start here |
|---|---|
| Run one prompt or produce JSON for automation | Agent runs |
| Work interactively with approvals and live output | Terminal UI |
| Resume a conversation or manage model context | Sessions and context |
| Capture commitments and approve execution | Tasks, decisions, and plans |
| Iterate on an objective or delegate bounded work | Goals and subagents |
| Preserve reusable non-secret context | Memories |
| Produce a durable report from repository, web, or MCP evidence | Deep research |
| Query an operator-configured provider-neutral search route | Web search |
What stays durable¶
Session messages, tasks, decisions, plans, goals, child jobs, memories, research runs, workflow runs, and audit evidence are canonical journal records. The keyless default stores payloads as plaintext; configured platform or environment keys add authenticated encryption. Terminal layout, search indexes, and projections can be rebuilt; they are not the source of truth.
When an external effect starts but its terminal outcome is not recorded, Colossus marks it unknown instead of assuming success or retrying silently.