An end-to-end workflow for complex coding tasks. Forge turns vague requests into verified changes through discovery, planning, delegated execution, and review.
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.
Use the forge skill by asking it for the thing you want to build:
/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 → [REFACTOR → RE-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.
Refactor runs only after a green Verify and skips docs, config, generated output, and trivial mechanical diffs. If it changes code, Forge runs the full verification team once more; if it skips, the task is already done.
For small, unambiguous tasks Forge declares a light tier at Approve: the approved approach doubles as the plan, so Research, Plan, and Refactor are skipped while Verify runs in full. Anything that breaks those assumptions mid-flight upgrades the task back to the full lifecycle.
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
- refactorer — simplifies verified changes without altering behavior
- 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 Codex use plugin marketplaces. Cursor marketplace approval is still pending.
In Claude Code, register the marketplace first:
/plugin marketplace add mariomka/forgeThen install the plugin from this marketplace:
/plugin install forge@forge-marketplaceRegister the marketplace, then install Forge:
codex plugin marketplace add mariomka/forge
codex plugin add forge@forge-marketplaceRestart Codex and open a new conversation so it loads the plugin.
Note: 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.
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.