More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    swarmer

    1

    A swarm of 100+ specialist agents your coding agent convenes for one request: nine strategies, blind lanes, voting juries, evidence for every claim.

    Free

    Security scanned

    About this skill

    The problem

    One agent answering a hard question tends to agree with itself: it picks the first plausible fix, reviews its own work, and calls it done. Real decisions need independent specialists, blind to each other, and a process that separates claims from evidence.

    What it does

    swarmer turns your coding agent into the coordinator of a swarm of 100+ specialist professionals and sub-professionals (formal methods, statistics, security, architecture, debugging, LLMs, perception and estimation, writing, economics, and more), convened as subagents for one request with one of nine strategies:

    • single: one specialist for one clear question;
    • fan-out: independent slices in parallel, or a race with a selection rule declared first;
    • classic: blind lanes, triage, a falsifier tournament, branches and verdicts, for decisions, designs, fixes and proofs;
    • jury and light jury: a prepared proposal voted blind (unanimity approves, majority revises, minority rejects);
    • full and plenary: every point of view for strategic choices;
    • expert: a provisional card for a competence the swarm lacks;
    • auto: the coordinator picks the cheapest strategy that suffices, and says why.

    Every brief is self-contained, every claim carries its evidence and a state (proven, verified, supported by data, refuted, unresolved), and agent reports are evidence, not proof: the coordinator reruns the checks.

    How it works

    The skill is small: SKILL.md and a standard-library script. The script reads on demand, one piece at a time, the public catalog generated from the open-source repository (one JSON file per role, strategy and document on GitHub Pages): only the protocol of the chosen strategy and the cards of the roles convened enter the context. It sends no user data; the role search runs locally. New professionals accepted in the repository reach every user automatically.

    Works in Claude Code, Codex, Antigravity, Gemini CLI and any agent that reads SKILL.md. Without subagents, the coordinator performs the roles in sequence and says so. The full installation (sealed and recounted juries, local overrides, Italian) is free on GitHub.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse All Claude Skills.

    The jury and convening scripts of the full installation (sealed, recounted votes; convening folders) are not included: the coordinator applies the same rules itself. English role cards. Local overrides and the roster that grows from requests need the full installation from GitHub.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed today

    • Free with an Agensi account

    Software
    Python 3.8+Network access to GitHub Pages

    Permissions

    Allowed Hosts

    grithat.github.io

    scripts/swarm.py only sends GET requests for fixed JSON files of the public swarmer catalog (role cards and protocol texts, generated from github.com/GritHat/swarmer). It sends no user data; the keyword search runs locally. Nothing is written to the project unless the user's request asks for it.

    Listedtoday

    What's inside

    Frequently Asked Questions

    It turns your coding agent into the coordinator of a team of specialists. For one request it picks a strategy (single, fan-out, classic, jury, light jury, full, plenary, expert, or auto), chooses the roles, writes a self-contained brief for each, launches them as subagents, and checks their reports before answering. Every claim is tied to its evidence and to a state: proven, verified, supported by data, refuted or unresolved.

    Claude Code, Codex, Antigravity, Gemini CLI and any agent that reads SKILL.md. With subagents, the roles run in parallel and blind to each other; without them, the coordinator performs the roles one at a time and says that independence is reduced.

    No. The bundled script only downloads fixed JSON files of the public swarmer catalog (role cards and strategy guides) from grithat.github.io, with plain GET requests. The search for the right roles runs locally on your machine, and nothing of your request, your code or your files is sent.

    To keep the skill small and current. The catalog has more than a hundred full role cards and nine strategy guides; loading only what a convening needs keeps the context small, and new professionals accepted in the open-source repository reach you without reinstalling the skill.

    The skill itself is about 1.5k tokens. A convening adds the guide of its strategy (about 2k tokens for single, 4.5k for classic) and the card of each role convened (about 1.5k each), plus the work of the agents themselves, which depends on your request. The coordinator always picks the cheapest strategy that meets the success criterion.

    68 professionals and 41 sub-professionals in 8 cores: theory and formal methods, AI and models, experiments and measurement, software and systems, knowledge and communication, design and decision, critical and divergent roles (adversary, debugger, goal guardian), and perception, estimation and motion. Run `swarm.py roles` to see the full roster.

    A prepared proposal is voted blind by 5 to 9 jurors, always including a goal guardian and an adversary. Unanimity approves (or recommends, by default), a majority asks for a small revision (at most three rounds, with no regressions), a minority rejects. Paid, irreversible or outward-facing actions always stay with you.

    The full installation (free, MIT) writes the agents into your project for each tool, and adds sealed juries recounted by a script, convening folders with gates, local overrides of the roles, Italian, and a roster that grows from expert requests. This skill installs nothing: the coordinator applies the same rules itself.

    The coordinator reports missing competences at the end of its answer, and the `expert` strategy creates a provisional card for the current request. To add a professional for everyone, propose it on GitHub with the "new professional" template: once accepted, it appears in the catalog for every user.

    Yes. The skill and the full installation are MIT-licensed and free; the only cost is the usage of your own coding agent.

    Free