Most people accumulate a dozen overlapping Claude Code skill packs and then have to remember which to invoke. forge replaces that pile with one cohesive plugin — the single best technique for each phase of software work (credited in CREDITS.md), wired into one lifecycle, plus a memory layer so your agent stops repeating mistakes.
understand → brainstorm → architect → plan → tdd ⇄ debug → review → verify → ship
( + orchestrate · eval · design-ui · principles · learn · recall )
[!IMPORTANT]
forge ships hooks that write a local activity log and run background claude -p calls on your own subscription to learn from your sessions. It's all local, fully disclosed in What the hooks actually do, and disabled with one setting (FORGE_AUTOLEARN=off).
Who it's for
forge is for you if:
- You juggle several skill packs and forget which to invoke → forge is one namespaced set with a clear hand-off chain.
- You want discipline enforced, not suggested → test-first, root-cause-first, and evidence-before-"done" as falsifiable rules.
- You want your agent to stop repeating mistakes → a memory layer that learns per-project and cross-project.
- You care about privacy → everything local, every hook disclosed, one-switch off.
Not a fit if you need a zero-hook plugin that never touches your tokens — forge's learning loop makes background model calls (off by a single setting).
Install
Option A — marketplace (one-liner):
/plugin marketplace add vasu-devs/Forge
/plugin install forge@forge
Option B — clone into your skills directory (recommended if you want the manual forge-store / forge-mem CLIs, which assume this path):
git clone https://github.com/vasu-devs/Forge ~/.claude/skills/forge
Then /reload-plugins (or restart Claude Code) and verify:
claude plugin list # look for: forge@skills-dir (or forge@forge)
Skills now appear as forge:tdd, forge:debug, forge:design-ui, etc. The data store (~/.claude/forge-data/) is created lazily on first use; the optional recall add-on needs one extra npm install.
Why forge
A raw model has predictable failure modes: it assumes silently, over-builds, edits code it doesn't understand, patches symptoms instead of root causes, and claims success it never verified. forge is an opinionated cure for each, expressed as falsifiable rules rather than vibes.
| |
|---|
| 🔁 Full SDLC | 16 hand-off-linked skills, from understanding code to shipping it |
| 🗺️ Live code map | a self-updating dependency graph — god-nodes · modules · blast-radius · cycles · dead-files — never stale |
| 🧠 Learns as you work | a background distiller turns your sessions into durable "instincts" |
| 🌍 Two memory tiers | per-project instincts + anonymized cross-project lessons |
| 🔎 Cross-session recall | (opt-in) local semantic search over past chats + claude --resume links |
| 🧩 Namespaced | forge:* — no collisions with other skill packs |
| 📦 Zero core deps | plain Node; only the recall add-on needs an npm install |
| 🔒 Local & disclosed | no telemetry; every hook documented; one-switch off |
| 🖥️ Cross-platform | macOS, Linux, Windows (path-normalized project keys) |
