SoloSwarm blueprintsWhat works
- Session board for every Claude Code session on the machine, including subagents. Worktree sessions roll up to their project. Status: running, needs approval, waiting for you, usage limit, idle.
- Approvals through Claude Code’s HTTP hooks. Every tool call is classified (low / medium / high / critical) before it runs. High-risk calls wait in the dashboard or on Telegram; ten minutes without an answer means denied. Critical ones are refused.
- Costs at API list prices from the usage Claude Code logs, deduplicated per API response. Grouped by project, client, model, day or session.
- Budgets per project, client or globally, per day or month. When one is spent, every tool call in the projects it covers is refused until you raise it.
- Timeline per session: prompts, tool calls, approvals and usage-limit events.
- Desktop app (Tauri) that runs the daemon for you. The Linux build is done; Windows and macOS installers are next.
How we tested it
- Unit and integration tests with synthetic logs (no real session data in the repo).
- A full import of our own logs: about 1.7 GB across ~130 session files took about seven seconds, and the cost breakdown matched what we expected per project.
- The approval loop end to end: a destructive command refused instantly, a
git pushheld, denied from the dashboard, and the denial (with its reason) returned to the agent.
What isn’t there yet
- Windows and macOS installers.
- Adapters for Codex CLI, n8n and OpenClaw/Hermes.
- A rule editor in the UI (rules live in
config.jsonfor now). - Telegram approvals work in tests but haven’t run long in daily use yet.
A note on what the rules are
The rules stop the common expensive mistakes: pushing to the wrong branch, deploying, changing DNS, paying, messaging. They are not a sandbox. Keep using least-privilege credentials.
If you run many agents and want to try it, join the preview list.