The brain of a modern coding agent, pointed at the web. Magpi inspects, acts, verifies, recovers, and keeps going.
$ git clone https://github.com/Magpi-works/magpie.git
$ cd magpie
$ npm run cli$ npm i magpi
$ magpiStart now
Install from npm for the fastest start, or clone the source when you want the full project locally.
Next: authenticate with /login, then give the open browser a job with /browser-start.
Full setup guideTeach your existing agent when to hand browser-heavy, repetitive work to Magpi. Paste once. Then give it goals.
Agent skills · Copy and paste
Copy this command to your Grok Bot Chief of Staff.
Give OpenClaw the Magpi parent skill once, then send it goals.
Teach Hermes to delegate browser-heavy work to Magpi.
Coding agents succeed by grounding every step in the environment. Magpi applies the same discipline to the browser.
Look before acting. Verify every action. Recover when the page disagrees.
observe → act → inspect
Criteria are evaluated in code against the live page—not accepted from the model’s report.
criterion: text_visible
State what must be true. The process exits non-zero unless those criteria pass.
exit_code !== 0 on failure
Submit, send, pay, and delete require approval by default. Automation cannot commit by surprise.
policy: ask
Runs preserve readable events, metrics, payloads, screenshots, and checkpoints under one goal ID.
events.jsonl · artifacts/
Measure tokens, cache behavior, and context per turn. Attribute cost instead of guessing about it.
metrics <goalId>
Find, qualify, prepare, wait for the few decisions only you can make, then keep going.
Product direction · Coming next
Campaign
Pentesting vendor qualification
Qualify and prepare
Approve as a batch
Continue and verify
Measure the work, attribute the context, and keep deterministic checks deterministic.
Capable model
Reason about the goal and constraints.
↓Model + tools
Act through bounded browser controls.
↓Deterministic checks
Accept success only when criteria pass.
Maximum reliable work per dollar.
npm install
npx playwright install chromium
npm run agent -- run "apply for the staff engineer role" \
--url https://example.test/jobs \
--criterion "text_visible:Application submitted" \
--policy askRun non-interactively with explicit criteria, steer it through the Pi CLI, or use the local web UI. Chromium stays on your machine.
Pi supplies the agent runtime. Magpi adds browser-native skills, bounded tools, evidence, and verification you can inspect or extend.
Coming next
Long-running campaigns keep working while collecting a small batch of decisions that genuinely require your attention.
Research, qualify, prepare
Approve 9 requests
Follow up, compare, verify
Inspect the architecture. Run it locally. Change the agent.