agent-dojo MCP server
Commit Nov 15, 2026Provides plugins, skills, and habits to help coding agents perform professional-grade work with guardrails and coordination capabilities.
Also known as dimokol/agent-dojo
- Stars
- 2
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 15, 2026
- License
- MIT
What is agent-dojo MCP?
agent-dojo provides plugins, skills, and habits to help coding agents perform professional-grade work with guardrails and coordination capabilities.
Summary generated from the project's README
About agent-dojo MCP server
Agent-dojo provides a suite of plugins, skills, and habits designed to bring professional standards and coordination to coding agents. It is built for developers using Claude Code or Codex who want to implement guardrails and automated workflows. Users can install the full starter pack via the Claude plugin marketplace or add individual components like git-guardrails and machine-pressure to manage specific risks.
The system functions through hooks, skills, and an MCP server. For Claude Code users, hooks can prevent unauthorized git actions or manage system resources. The agent-chat MCP server allows multiple agents in different terminals to communicate and cooperate without a middleman. While hooks are exclusive to Claude Code, skills are compatible with Codex and other agents. Setup requires Node 24.15+ for certain features and specific configuration within a CLAUDE.md file.
What you can do with it
- 1Prevent unauthorized git operationsStop an agent from merging a pull request or pushing to protected branches without explicit permission.
- 2Manage system resource usageAvoid crashing your computer by refusing heavy commands when CPU or memory usage is too high.
- 3Coordinate multiple agentsEnable agents running in separate terminal sessions to message each other and work together.
- 4Maintain workspace cleanlinessAudit your environment to find and propose the removal of orphaned dev servers or idle dependency folders.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachabledimokol/agent-dojo is public and reachablegithub.com/dimokol/agent-dojo ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/dimokol/agent-dojo ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubagent-dojo
Plugins, skills and habits that help you do professional-grade work with coding agents.
Taken straight out of the personal workflows I apply daily in my professional work as a web and game developer. In Claude Code, hooks stop an agent from merging a PR you didn't approve or starting a second test run on a machine that's already struggling. agent-chat gives agents in different terminals, Claude Code or Codex, a room to coordinate and cooperate between themselves without requiring a middleman. The skills, which also work in Codex, carry the review and verify habits that keep agent work up to your standard.

Install
claude plugin marketplace add dimokol/agent-dojo
claude plugin install starter@dimokol
Restart Claude Code. starter brings git-guardrails, machine-pressure, worktree-hygiene and
agent-chat, which needs Node 24.15+, the native Claude Code install, one setting and three
lines in your CLAUDE.md (see its README). Claude Code may say some options are "not yet set"; the
defaults apply. To see one work, ask "run a health check".
Skills only, for Codex and other agents: npx skills add dimokol/agent-dojo.
What it prevents
- An agent merges a PR, pushes to main or runs
reset --hardbefore you said so. - Four agents start four test suites, and the laptop swaps until nothing responds.
- Two agents in two terminals guess the same field name differently.
- Merged worktrees, idle
node_modulesfolders and orphaned dev servers fill the disk. - An agent rebuilds something that already merged on the trunk.
- A PR goes up with review findings nobody applied or answered.
New to running more than one agent? Running several agents at once is a five-minute read.
Parts
| Part | What you get | Type |
|---|---|---|
| git-guardrails | Agents can't merge PRs, push to protected branches, delete remote branches or run reset --hard unless you said so. Strict mode also gates every commit and push. |
hooks |
| pr-task-link-guard | gh pr create is refused when the body has no link to a task, once you set a task-link pattern. |
hook |
| machine-pressure | Refuses a second e2e or Docker run, and heavy commands while CPU, memory or disk is in the red, plus a statusline script you add by hand. | hook, script |
| context-nudge | One line when a session passes 250k, 400k and 600k tokens, or after an hour idle at 150k or more, so you compact or start fresh. | hook |
| worktree-hygiene | An audit to run at session start: merged worktrees, idle dependency folders, orphaned dev servers, free disk. Proposes, never deletes. | skill |
| storage-reclaim | Disk cleanup that opens with the biggest safe wins and a one-word "rec", and won't touch what a process or another session is using. macOS. Bundles disk-map, a treemap for any disk. | skills |
| agent-chat | Agents in different terminals message each other, and a message wakes the idle Claude Code or Codex session it's for. Mail waits for closed sessions. | MCP server, hooks |
| verify-before-building | Fetches and checks the trunk before you branch, so nobody rebuilds what already merged. | skill |
| pr-review-loop | Asks the reviewer once, applies or answers every finding, reruns your checks, stops when the PR is ready. Never merges on its own. | skill |
| integration-branch-qa | Test several approved PRs at once on one local branch that never gets pushed, with a clickable checklist and a merge gate. | skill, hook |
| setup-audit | A weekly read of the Claude Code changelog plus your own drift checks, written to a log. | skill |
blocks/ holds text to paste into your own CLAUDE.md or AGENTS.md: a working
agreement between you and the agent, web project standards, a writing tone, a product copy voice,
and a docs layout.
Settings and other installs
Pick single parts with claude plugin install <part>@dimokol. Every setting has a default. Change
one with /plugin configure <part>@dimokol; each part's README lists them. Your own hooks stay as
they are. Turn a part off with claude plugin disable <part>@dimokol. If you installed starter,
disable starter@dimokol first, because it holds its four parts in place.
Without the plugin system, install.sh copies a part into place and prints any hook or MCP snippet
for you to paste:
curl -fsSL https://raw.githubusercontent.com/dimokol/agent-dojo/main/install.sh | bash -s -- --list
curl -fsSL https://raw.githubusercontent.com/dimokol/agent-dojo/main/install.sh | bash -s -- git-guardrails
# Codex: copy skills into Codex's folder
curl -fsSL https://raw.githubusercontent.com/dimokol/agent-dojo/main/install.sh | SKILLS_DIR=~/.codex/skills bash -s -- worktree-hygiene
In Codex the skills and agent-chat work; hooks are Claude Code only.
Also worth installing
Tools that fit with these parts. Linked, not copied.
| What | Why | Install |
|---|---|---|
| superpowers, pr-review-toolkit, feature-dev, claude-md-management, skill-creator | Planning, subagent-driven builds, review agents, CLAUDE.md upkeep | claude plugin install <name>@claude-plugins-official |
| mattpocock/skills: grilling, diagnosing-bugs | Sharp questions before building, a debug loop that starts from a failing repro | claude plugin install mattpocock-skills, or one skill: npx skills add mattpocock/skills --skill grilling |
| pstack: unslop, blast-radius, show-me-your-work | Cuts AI tells from writing, finds what a change could break, keeps a decision log | npx skills add https://github.com/cursor/plugins/tree/main/pstack --skill unslop -a claude-code (Cursor: /add-plugin pstack) |
| ccstatusline | A configurable statusline; machine-pressure ships widgets for it | npm install -g ccstatusline |
| claude-notifications (from me as well!) | A sound and an OS banner when a Claude Code session in a VS Code terminal finishes or needs you. On macOS and Windows a click jumps to that window and terminal tab. | VS Code marketplace |
Credits
- ccstatusline by sirmalloc draws the statusline that machine-pressure's widgets plug into.
- superpowers by Jesse Vincent shaped how these parts were planned, built and reviewed.
- mattpocock/skills: its grilling skill settled the design questions, and it has a git-guardrails skill with the same aim as this one.
- pstack: its unslop skill is the stricter checklist
blocks/writing-tone.mdpoints to, and its worktree-cleanup is the fuller playbook worktree-hygiene points to. - The demo is recorded with VHS by Charm.
License
MIT. See LICENSE.
agent-dojo MCP questions
What does git-guardrails do?
It prevents agents from merging PRs, pushing to protected branches, deleting remote branches, or running hard resets without explicit permission.
Can agents communicate with each other using this tool?
Yes, the agent-chat feature allows agents in different terminals, such as Claude Code or Codex, to coordinate and cooperate directly.
Does it help prevent system resource exhaustion?
Yes, the machine-pressure tool refuses heavy commands or additional test runs when CPU, memory, or disk usage is too high.