Product · Agents
Agents with a job description
Tokenalytics runs Claude Code and OpenAI Codex sessions with your own accounts. Each stage gives the agent a narrow job, the tools it needs, and a clear point where it stops and waits for you.
Four roles, separated on purpose
Instructions live with the stage
Each stage has a short instruction file and a tool list. The builder can edit files and run your checks. The reviewer can read and comment, but not push.
Because the rules are files in your repository, changes to them go through review like any other code.
stage: buildingagent: codexmax_parallel: 3requires: plan.approvedchecks: - bun run test - bunx tsc --noEmiton_success: open_pull_requestLimits an agent cannot cross
- 01
No merges
Agents open pull requests. Merging is always a person's action on GitHub.
- 02
No unapproved builds
Building starts only from a plan someone approved.
- 03
No shared workspace
Every card gets its own worktree, so parallel agents cannot overwrite each other.
- 04
Stop means stop
Stopping a card ends the session and frees its slot immediately.
Give your agents a process
Start with the default six stages, then change what you need.