forge-skills MCP server

---

Give your agent a better way to work

A convincing plan can rest on a missing dependency. A passing test can miss the real entry point. A new session can lose the reason behind yesterday's decision.

Forge Skills gives agents repeatable ways to inspect the evidence, challenge assumptions, keep code quality from slipping, and describe exactly what has been verified. Install the skills you need. Invoke them when the work calls for them.

Seven skills · Claude Code and Codex · Subscription-first agent work · Local memory included

See what the skills do

Two one-minute illustrated guides, explained through everyday examples. Click a cover to watch; each film is available in wide and vertical formats, with captions and downloads.

Check the plan before building

A booking-page example shows how Plan Hygiene helps an assistant inspect the real project, find gaps and define what success should look like. Watch Plan Hygiene →

Carry useful decisions into the next conversation

A fictional shop shows how explicitly saved decisions can inform later work, with a check that they still apply. Watch Agent Memory →

The films use original illustrations, on-screen explanations and music, without voiceover. They explain the workflows; they are not recordings of live assistant conversations. Browse the video files and captions.

Choose your skill

When you need to…	Use	What you get
Remember decisions across sessions	Agent Memory	A runnable local server with recall, capture, corrections, and optional maintenance
Find the next useful action	Project Forge	A grounded project assessment and next step
Check a plan before committing effort	Plan Hygiene	Dependencies, risks, scope, and proof-of-success gates
Stop quality debt from growing	Anti-Slop Gates	Executable checks with downward-only debt limits
Verify a completion claim	Completion Hygiene	A promise-by-promise evidence audit
Improve a reusable prompt or skill	Prompt Forge	Preserved constraints and a behaviour evaluation plan
Resolve a consequential decision	Crucible Forge	Evidence, dissent, a decision, and reversal conditions

Start with one

Requires Git and Python 3.11 or newer for the installer and tools. Instruction-only skills can also be copied manually. Run these in a terminal; change the project path to your own.

git clone https://github.com/qengine13/forge-skills.git
cd forge-skills
python3 scripts/install.py --agent claude --project /path/to/your-project --skill plan-hygiene

Open a new Claude Code session inside your project, then try:

/plan-hygiene Check my proposal to add bulk invoice export. Inspect existing
export paths and dependencies, define what would prove success, and do not implement it.

For Codex, use --agent codex during installation and invoke $plan-hygiene. Use --all instead of --skill plan-hygiene to install the collection. Agent Memory also needs a one-time server connection. The installer refuses to overwrite an existing skill. Installation, updating, and troubleshooting →

A workflow you can understand

flowchart LR
    O[Project Forge<br/>Understand the current state] --> P[Plan Hygiene<br/>Test the plan]
    P --> B[Build with<br/>Anti-Slop Gates]
    B --> V[Completion Hygiene<br/>Verify the promise]
    P -. A difficult choice .-> C[Crucible Forge]
    V -. Improve the instructions .-> F[Prompt Forge]
    M[Agent Memory<br/>Recall decisions; verify current facts] -. Context .-> O

You do not have to run every skill on every task. A small edit should stay a small edit. Walk through a concrete example →

Your subscription. Your available models.

The skills prefer the newest stable, suitable model available through your current subscription when the host can verify and select it. Your explicit model choice takes precedence. When discovery or entitlement is unavailable, they keep the configured session model and say that the latest eligible model is unverified. They never assume a chat subscription includes API credits.

Model choices stay fixed during a comparison so a newer model cannot masquerade as a better prompt. Read the full model policy →

Know what you are installing

• Instructions: Markdown workflows that the host agent interprets. They cannot guarantee compliance.
• Tools: A local installer, quality checks, offline decision validation, and a local memory server. Inspectable Python.
• Memory: Basic recall needs no model or API key. Optional model work is off by default; Claude adjudication uses a verified subscription CLI login.
• Permissions: The installer copies files only. It does not configure credentials, install hooks, change your model, or contact an AI provider. Invoked agents retain their own tools and permissions.
• Evidence: Automated package/tool checks are distinct from live model and end-to-end workflow tests. Verification and limitations →

Make it yours

Use existing project rules and conventions. Choose individual skills. Keep the scope small. Contributions should include a reproducible failure or a realistic scenario showing why a change helps.

Privacy review · Contributing · Provenance · License · Report an issue

Related work: Gauntlet Loop, an external project credited to Matt Shumer (technique) and Jay E / RoboNuggets (packaging), is not bundled here.