pi-dispatch CLI
Background runs of pi coding agents on your machine by schedule and label, each job in a Docker container with a spending limit.
- 197 ★ on the repository
- Last commit
Broad access
Why this level
- Agents edit code and open pull requests in the background without you present.
- The AI provider key and a GitHub or GitLab token are passed into the container and the agent can read them.
- A job on a local folder changes it directly with no undo.
Install
Manual install
npx --yes --ignore-scripts @edgehero/pi-dispatch@5.1.0 up --consoleRun it in a new empty folder on Mac or Linux. For pi there is also `pi install npm:@edgehero/pi-dispatch-admin` and the `/dispatch` command.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Agents edit code and open pull requests in the background without you present.
- The AI provider key and a GitHub or GitLab token are passed into the container and the agent can read them.
- A job on a local folder changes it directly with no undo.
This is third-party code. Review the repository files before installing.
What it does
pi-dispatch runs the pi coding agent in the background on your own computer or server and shows every job on one web page, the console at `127.0.0.1:4790`. A job starts on a schedule or when a GitHub issue gets a label such as `pi:fix`. A second agent can review the result. Each job runs in a sealed container that is thrown away afterwards and stops at the spending limit you set per day, week or month. For projects on GitHub, GitLab, Forgejo and Azure DevOps the work ends at a proposed change: pi-dispatch merges nothing, and what ships stays a person's decision. You pick the level of involvement per kind of work: watch the agent, answer its questions or look only at the finished result.
Who it is for. Tech leads and developers whose same tasks come back every week and who are tired of sitting next to an agent.
Good fit when
- Every night at three the agent tidies the code and sends a pull request.
- An issue gets the `pi:fix` label and a proposed fix appears later.
- The agent answers review comments on its own changes.
- Several projects share one queue, one budget and one status page.
Not a fit when
- A one-off edit on your own machine where plain pi is enough.
- Docker cannot be installed on the machine.
- Agents are expected to merge changes into the main branch themselves.
Example request
Run the tidy job on my-project every night and send a pull request, with a spending limit of 5 dollars a day.Limitations
Docker and Node.js 22.19 or newer are required, plus an AI provider key or subscription. On Mac and Windows jobs run only while you are signed in. Browser setup works on Mac and Linux. A job on a local folder edits it directly with no undo, so commit your work first. Two secrets reach the container and the agent can read them: the provider key and the GitHub or GitLab token. Use a narrow token and set a limit at your provider.
How to disable. Stop the service with `./node_modules/.bin/pi-dispatch service stop --console`, delete the install folder and revoke the GitHub or GitLab tokens. Job containers delete themselves.
FAQ
How do I install it with one command?
In a new empty folder on Mac or Linux run `npx --yes --ignore-scripts @edgehero/pi-dispatch@5.1.0 up --console`. Before starting anything the program shows a plan and asks for consent.
Can I use a subscription instead of an API key?
Yes, if you set it up on purpose: ChatGPT sign-in or Claude through a community bridge. Check subscription terms with the provider.
What does the `pi:fix` label do?
The rule fires when the label appears on an issue, the agent works on a copy of the repository and returns a pull request. It never merges the change.
How do I limit spending?
Limits in dollars and jobs are set per day, week and month, for all projects or one. A job over a limit is refused before it starts.
What happens if the machine restarts?
The queue is kept on disk and waiting jobs stay. A running job is lost or restarted unless its rule opts into continuing from a checkpoint.
Discussion
No questions or reviews yet.
Related
CLI toolsA web app for the Pi coding agent: sessions survive restarts, teammates can join, and you can work from a phone with no cloud VM
A minimal terminal coding agent extended with skills, prompt templates and TypeScript packages
Personal Pi agent setup with four commands for plan, todos, execution and review, passing state through Markdown session files.
Repeated pi agent steps accumulate as JavaScript programs with a manifest and tests, and pi calls them instead of redoing work.
A web dashboard for the pi coding agent with sessions grouped by branch, a built-in terminal and a mobile mode
An extension for the Pi agent: one programmable fabric_exec tool calls tools, MCP, agents and workflows from code
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| pi-dispatch (this card) | no | 197 | ||
| Ralph Orchestrator | no | not required | 3,203 | a loop that runs an agent until done; pi-dispatch keeps a queue, spending limits and a container per job |
| Dispatch | no | 414 | launching agents from a session; pi-dispatch works in the background on schedules and labels and shows every job on a web page |
README in short
The project calls the approach dark engineering: agents work without you while you look at one page with counters. It tracks how often a person intervened per project and offers fallback models, model roles and budget splits.