The idea
Treat every token like it costs money.
Most AI coding burns the biggest model on every step, including the cheap ones. Metis spends on purpose. It breaks work into focused tasks, sends a cheap model to explore, a mid-tier model to build, and the expensive one only where judgment actually matters. It plans before it writes, verifies before it commits, and runs up to four agents at once under a budget you set. Less waste, fewer wrong turns, and work that scales past a single chat window.
Set a budget
Give the swarm a token ceiling. It stops when it gets there.
Right model, right job
Haiku explores, Sonnet builds, Opus judges.
Four at once
Up to four agents in parallel, dependency-aware.
It learns your repo
Every finished task leaves notes the next run reads.
Orchestrate
Swarm
Point the swarm at your backlog and it runs the work for you. Opus breaks each task down, cheaper agents do the building, and nothing is committed until it has been verified. It tracks dependencies, keeps up to four agents busy at once, and halts the moment it reaches the budget you set. Each agent is handed only the instructions its task needs, which cuts the prompt it carries by 60 to 80 percent.
Verify
iOS QA
The ios-qa plugin drives the iOS Simulator and checks your app the way a picky user would: it screenshots each screen, holds it against the task's acceptance criteria, fixes the small stuff itself, and escalates the real bugs. It is the same harness that swept Finn in [a blind, peer-judged model bug-hunt](/blog/blind-ai-bug-bounty-benchmark).
Features
A Claude Code marketplace
Two plugins, sixteen commands, and a set of reusable capability packs. Add the marketplace and install only what you need.
Cost-aware by design
Model tiering, per-task capability subsetting, and a hard budget flag keep token spend deliberate instead of accidental.
Open source, MIT
Built in the open. Read how every command works, then bring your own skills and capabilities.
Metis is a marketplace of Claude Code plugins built around one idea: an AI coding session should be cheap, disciplined, and able to run more than one thing at a time.
Add it
/plugin marketplace add chensagi/metis
/plugin install metis@metis
/installThe optional iOS QA companion installs the same way:
/plugin install ios-qa@metisWhat’s inside
- Swarm orchestration. Parallel task execution with dependency tracking, per-task verification, and a budget ceiling.
- iOS QA. Spec-driven visual QA that drives the iOS Simulator and bundles its simulator driver, so there is no separate setup step.
- Capability packs. Reusable instruction bundles for TypeScript, React Native, Expo, Python, Go, and more, handed to an agent per task instead of all at once.
- Cost tracking. An optional metrics server that reports token spend while the swarm runs.
How it works
Every task runs the same loop: ask the clarifying questions first, explore the code, research the approach, plan it, then execute and verify. Cheap models do the cheap steps. The expensive model is saved for judgment. The payoff is fewer wrong turns and a token bill you can predict.
Metis is open source under the MIT license. The skills, capabilities, and plugins are all readable, and the project is built to be extended.