A control room for your coding agent, not just a chat window

What it does

Pi-Harness is an Electron desktop app (macOS, Windows, Linux) that runs the Pi Coding Agent natively and builds an operational layer around it: a visual workspace for conversations, projects, files, Git, models, and Skills, plus tooling to inspect runs, enforce budgets, recover work, and coordinate agent teams. Your existing Pi sessions, config, tools, and extensions stay compatible — you can keep using plain Pi outside the app.

The interesting bit

The “harness” half is the real story: run traces, replay, run comparison, checkpoints, and regression baselines turn agent sessions from ephemeral chat logs into inspectable evidence. It also unifies local history from Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and Pi in one read-only browser — token-free until you ask for an AI summary.

Key highlights

• Run Tree with forks and retries, plus token/cost/duration tracking per run
• Tool policies, per-run budgets, checkpoints, and resume/fork/retry controls
• Multi-agent teams with task DAGs, review gates, and worktree isolation
• Local project architecture maps (JS/TS/Vue import graphs; file-level index elsewhere) exported as JSON, HTML, or PNG
• Custom skins via a separate starter repo; UI in eight languages

Caveats

• Architecture maps cover static import relationships only — no full call graph, and non-JS/TS languages get a shallower file-level index (the app says so itself).
• AGPL-3.0-only, which matters if you’re thinking of embedding it in anything commercial.
• Continuing a task from imported history starts a new Pi conversation rather than resuming the original.

Verdict

Worth a look if you run Pi (or several agents) daily and want budgets, traces, and recovery instead of a bare terminal. If you’re happy with your current CLI setup and never needed to replay a run, this is a lot of app for a problem you may not have.