10,000+ SKILLS · FREE FOREVER · NO KEYS · OFFLINE-FIRST
Reusable skills
for every agent
SkillForge is the open registry for SKILL.md agent skills — harvested from GitHub, npm and official MCP servers, safety-scored, hash-sealed. Search, compose and install for OpenClaw, Claude Code, Cursor and any MCP agent.
0
indexed skills
0
live sources
0
avg SkillScore
0
avg safety
01 — TOP SCORED
This week in the forge
Highest SkillScore records across all 10,000+ indexed skills. Every card opens a full breakdown: factors, seal, SKILL.md and one-command install.
02 — BROWSE BY CRAFT
Six crafts, ten thousand skills
Every record is classified into a craft at harvest time. Pick a lane — each card shows its reigning champion by SkillScore.
03 — ENGINE
Advanced under the hood, calm on the surface
Live signals, sealed fallback
GitHub, npm, MCP servers and Hacker News refresh the catalog. If any feed fails, the 10,000-record sealed offline catalog keeps every page, demo and first paint alive.
One engine, three consumers
SkillScore is deterministic and keyless: docs, breadth, compatibility, license, community, safety. The same function powers the UI, the REST API and the MCP tools.
Your agent can shop here
Five MCP tools — search, get, score, compose, verify — over JSON-RPC. Paste the mcp.json block and any capable agent can install skills by itself.
Integrity you can replay
Every record carries a SHA-384 hash chained to its predecessor. POST the ids to /api/seals/verify and watch 10,000 hashes recompute in under a second.
04 — FRESH FROM THE FEED
Trending on GitHub right now
Ranked by real stars, refreshed hourly, sealed on arrival. Side-scroll through what agent builders are starring this week.
Ranked by community stars · refreshed hourly · scroll sideways for more
05 — AGENT INTERFACE
Do not take our word for it — ask your agent
The console below hits the production MCP endpoint live. One click runs tools/list, search, scoring and seal verification.
Request
Press a preset to call the live MCP endpoint.
Response
Response appears here.
06 — SAFETY
Skills are code that acts. Score the risk before you grant it.
From prompt tricks to recursive self-improvement, the road to superintelligence runs through agents executing skills. SkillForge maps every skill to a machine-readable risk surface, a 0–100 safety score, and guardrails your agent must obey.
Read the alignment briefing07 — CONTRIBUTE
Forge the 10,001st skill
The catalog grows the way good infrastructure should: one reviewed, risk-declared, sealed record at a time.
01
Fork and add your skill
One record in src/lib/fallback.ts: id, triggers, workflow, guardrails, honest risk flags.
02
Declare risk honestly
exec-code, writes-files, network-egress, external-api, reads-secrets — the engine scores what you declare.
03
Pass lint and build
npm run lint plus npm run build, both green. No keys, no paywalled deps, ever.
04
Open the PR
Maintainers check scoring honesty and guardrail quality. Merged skills ship sealed to 10,000+ neighbors.
08 — QUESTIONS
Asked by humans and agents alike
+Does it cost anything? Do I need API keys?
No and no. The catalog, scoring, MCP tools, and installs are free forever with zero keys. Live feeds use keyless public APIs; everything degrades to the sealed offline catalog when a feed fails.
+Where do 10,000+ skills come from?
A harvester indexes real public sources — GitHub agent/skill/MCP repos, npm agent packages, official MCP servers, and exact SKILL.md bodies — then seals every record into a SHA-384 chain. Re-harvests refresh the catalog; the harvest script ships in scripts/harvest.mjs.
+How does SkillScore work?
Six deterministic factors: documentation completeness, capability breadth, multi-agent compatibility, license openness, community signal, and safety posture — plus a separate 0–100 safety score from declared risk flags. Same function serves the UI, REST, and MCP.
+How does my agent use it?
Add public/mcp.json to your agent config and it gets five tools: search_skills, get_skill, score_skill, compose_bundle, verify_seal. Humans get the same power over REST plus one-command curl installs.
+Can I trust a high score?
Scores are heuristics, not guarantees. Always read the SKILL.md, check the safety number, verify the seal, install least-privilege first, and never paste secrets into a skill. The safety briefing explains the full model.
+How do I publish my skill?
Fork the repo, add one record to src/lib/fallback.ts with honest risk flags, keep lint and build green, open a PR. The authoring guide in Docs walks through it in five moves.