Every agent run, step by step.
Sulcus shows what your agent is doing while it runs, and keeps the record when it’s done.
Illustration with sample data, not a live run: a hosted run named support-triage with three agents, 38,420 of a 50,000 token limit used, 1 minute 12 seconds of runtime, an agent tree with a status and a timeline for each agent, recent events, and a Stop run action.
Follow the run.
Status
Queued, running, then a final outcome, with the runtime counting up.
Structure
Graph nodes, crews and tasks, agents and handoffs, each with its status and place on the timeline.
Model and tool calls
Each call your framework exposes appears as it starts, finishes or fails.
Tokens
Input and output tokens as your framework reports them, for each call and the whole run.
Step in while it runs.
Stop a hosted run
Sulcus started the container, so it can end it.
Set a token limit
Once reported usage reaches it, Sulcus blocks the next model call or ends the run. It isn’t a billing cap.
Answer approvals
When Claude Code or Codex asks for permission, approve or deny it in Sulcus. In early access.
Ask a local agent to stop
Stop Claude Code at its next action, or interrupt a Codex turn, while the computer is online.
When a run fails, see why.
Run Details names the cause in plain language and keeps the logs that led up to it.
How a hosted run works.
- Point Sulcus at a public Git repository and choose your framework.
- Sulcus runs your entrypoint in its own container and instruments the framework, with no Sulcus imports in your code.
- Watch it live and stop it whenever you need to. The run keeps the commit and settings it started with.
What Sulcus records.
Which agents, steps and calls ran, their timing and the tokens reported. The framework integrations don’t collect prompt text, model responses or tool inputs and outputs, but a hosted run keeps what your app prints.
Security and data handling