An orchestration skill for coding agents. Forge takes ownership of complex tasks end-to-end — it helps you discover intent, then plans the approach, delegates the work to specialized sub-agents, and verifies quality before declaring the job done.
The goal: prevent the most expensive mistake in software — building the wrong thing, or building the right thing badly. Discovery costs minutes; rework costs hours.
Forge orchestrates sub-agents so the main context is almost never filled — each noisy step (code reading, planning, building, verification, review) runs in a fresh sub-agent with its own context window. What comes back is a concise synthesis, not a wall of tool output. In practice that means:
- Longer tasks without running out of context — the expensive reads and multi-file edits live in sub-agents that you throw away after each step.
- Cleaner main conversation — you see decisions and summaries, not
grepoutput andtsclogs. - Better decisions — each sub-agent is prompted for one narrow job, so it stays focused instead of drifting.
- Parallel work when it's safe — independent sub-agents run concurrently, which is only possible because they don't share context.
Just tell Forge what you want, using the slash command:
/forge add a rate limiter to the public API
Forge will take it from there — starting with a discovery conversation, then planning, then execution, with explicit checkpoints along the way.
The discovery conversation is where the real value lives. Don't rush through it — treat it as a collaborator thinking out loud, not a form to fill in.
DISCOVERY → APPROVE → RESEARCH → PLAN → [RECONFIRM*] → EXECUTION → VERIFY
* Reconfirm only fires when the plan deviates from the approved approach. If it tracks, Forge skips straight to execution instead of asking for another rubber-stamp.
Forge coordinates a team of specialized agents:
- strategist — turns goals into sequenced roadmaps
- analyst — investigates how the code works, and root-causes failures
- builder — builds the solution
- validator — validates against acceptance criteria with evidence
- reviewer — checks code quality
- auditor — security audit for sensitive surfaces (runs in parallel with reviewer)
Note: Installation differs by platform. Claude Code and Cursor have built-in plugin marketplaces. Codex requires a small amount of manual setup.
In Claude Code, register the marketplace first:
/plugin marketplace add mariomka/forgeThen install the plugin from this marketplace:
/plugin install forge@forge-marketplaceNote: The Cursor marketplace listing is pending approval. In the meantime, tell Cursor:
Fetch and follow instructions from https://raw.githubusercontent.com/mariomka/forge/main/.cursor-plugin/INSTALL.md
Once the marketplace listing is approved, install from Cursor Agent chat:
/add-plugin forge
Or search for "forge" in the plugin marketplace.
Tell Codex:
Fetch and follow instructions from https://raw.githubusercontent.com/mariomka/forge/main/.codex/INSTALL.md
Forge works with any harness that can load local skills and agent definitions — point its loader at skills/ and agents/ in this repo. If you've wired up OpenCode, Gemini CLI, Copilot CLI, or another harness and want to contribute the integration, PRs welcome.