Murmuration

One flock.
Every tool.

Define your AI agents once, then compile them for every coding tool your team uses.

See how it works

Write it once, in plain files

Your agents, skills and rules live in your own repository. A single command turns them into the native format of each tool.

murmur/agents/analyst.md

---
description: "Use when: deep codebase or domain
  analysis is needed before planning."
tools: [read, search]
dispatch:
  tasks: [analysis, investigate,
    architecture, dependencies]
---

# Analyst — Deep Researcher

Then run

murmr compile --target claude

One definition becomes

  • Claude Code.claude/agents/*.md, CLAUDE.md
  • Cursor.cursor/rules/*.mdc, .cursor/agents/*.md
  • GitHub Copilot.github/agents/*.agent.md
  • gooserecipes/*.yaml, .goosehints
  • Google Antigravity.agents/plugins/<name>/
  • ACP.acp/manifest.json, .acp/agents/*.json

No leader. Simple shared rules.

A murmuration has no central controller. Thousands of starlings follow a few local rules, and the flock moves as one. The same idea shapes this framework.

Keep your agents

Switch from Cursor to Claude Code tomorrow and run one command. Your setup stays in Git.

Run it locally

It works on your machine with the keys and tools you already have, with no sandbox to provision.

Reuse knowledge

Agents stay generic. Project facts live in skills and instructions that any agent can load.

Send the right agent

Agents declare when to run and when to skip, so each task reaches the specialist that fits.

A team on day one

Fifteen generic agents ship with the framework, and none of them need to know your codebase in advance.

Coordination

masterBreaks down tasks and delegates to specialists

Strategy

architect, planner, prompt-engineerPlans the work and refines the prompts

Execution

implementer, analyst, curatorWrites code, analyses systems, maintains knowledge

Verification

researcher, fact-checker, verifierGathers sources and audits the result

Critics

critic, business-critic, data-critic, social-critic, research-criticReviews from technical, business, data, social and research angles

Browse the whole catalogue

Every agent, skill and rule below is the real definition that ships in the repository.

    The commands you'll use

    • murmr initReads your codebase and generates skills, instructions and agents
    • murmr compile --target claudeWrites native files for one tool
    • murmr classify "<task>"Routes a task to the agents that fit
    • murmr run <pipeline>Walks a governed multi-phase workflow
    • murmr score <doc> --rubric <name>Scores a document against a rubric
    • murmr doctor --fixValidates definitions and repairs missing references