I built Jimothy so agents can ship work without skipping the human gate

“What if labeling a ticket was enough to start the work—and approving the result was still required before anything left the machine?”

The short answer

Jimothy is live in free beta at jimothy.dev: AI agent pipelines that run on your machine, start from the tools you already use, and pause at a human gate before anything ships.

I built it because most “agent” demos skip the part that matters in production. They chain prompts. They do not give you retries, isolated worktrees, cross-model review, and an explicit approve/reject step before a PR opens or a message goes out. Jimothy does.

What Jimothy is

Jimothy chains the agents and CLIs you already have—Claude Code, Codex, Gemini CLI, Cursor Agent, Copilot CLI, local models, your own CLI—into real pipelines. Steps form a DAG. Tests and lint can run in parallel. A reviewer that says changes requested can send work back to the implementer with the feedback attached. Shell steps run your tests. Approval steps wait for you.

A typical engineering run looks like this:

1. Plan on a strong model (for example Claude Code on Opus)
2. Implement on a fast model (Sonnet)
3. Run tests with npm test or whatever your repo uses
4. Review with a different harness (Codex)
5. You approve
6. Only then: push the branch and open the PR

Triggers start the run: a Linear, Jira, or GitHub issue (labels, JQL, repo filters), a cron schedule, or a webhook. Jimothy can comment on the issue and move it—for example to In Review—when a run starts, succeeds, or fails.

It is not only for code. The same shape works for support reply drafts, weekly product updates, and inbound lead research: agents prepare the artifact; you approve before anything is sent.

Why local-first

Runs happen on your machine—macOS, Linux, or Windows via WSL2—with your tools, in folders you choose. Each run gets its own git worktree and branch (or a fresh clone / empty folder), so parallel runs do not step on each other.

Secrets live in the OS keychain. Runs, logs, and settings are plain JSON and JSONL you own. The webhook server listens on 127.0.0.1 by default. Issue text interpolated into shell steps is quoted so a ticket description cannot inject commands. Your code does not leave the machine unless you push it.

That is the same bar I argue for in production AI work: owned agents, constrained actions, an audit trail you can read, and a human at the irreversible edge.

Pipelines, not a chain of prompts

If you have used CI, the control plane will feel familiar: retries with backoff, timeouts, runIf, continueOnError. The difference is the steps are agents (and shell, and approval).

You watch a live pipeline graph and streaming logs—tool calls, thinking, results. Retry from any step. Cancel the process tree. Open the workspace in your editor. Get notified via desktop, Slack, Discord, Teams, ntfy, or a webhook when something needs you.

Setup is four pieces: connect an account (Linear / Jira / GitHub), add a trigger, pick or build a pipeline, then approve and ship.

Free beta — go label something

Jimothy is free to use in beta. Download, connect a tracker, import a starter pipeline (Issues → pull requests is the obvious first one), label an issue, and watch the run. When the gate lights up, approve or reject from the app, the tray, or your phone.

If you have been waiting for agent automation that still respects ownership, evals, and human authority—this is the product I wanted on my own desk.

Try it: jimothy.dev

If you kick the tires and something is wrong, tell me. Beta means the raccoon is still learning the neighborhood.