Certified protocol know-how for AI agents. Every skill in this repository is tested by real agents on a fork of mainnet before it is listed, and always runs inside an Altana session: an allowlist and spend cap the user sets, revocable in one transaction.
Agents discover these skills through the Altana MCP server (search_skills, get_skill).
Humans browse them at skills.altana.network.
skills/<id>/SKILL.mdis the skill: a Reference (protocol facts) and a Playbook of Plays (ready made, parameterized action plans).index.jsonis the registry agents read: id, description, required session scope, version, content hash. Only tested skills appear here.
Skills published by Altana are fork-tested before listing: their mechanics are verified
with scripted runs against a fork of BNB mainnet (real contracts, real pool math),
including a check that the session scope rejects out-of-bounds calls. The harness in
tools/skill-test can additionally run full agent-trial certification (N fresh agents
performing the task, judged from onchain state); that heavier gate is reserved for
community submissions when they open.
Open a pull request following CONTRIBUTING.md. In the current phase,
certification runs are performed by maintainers before merge; index.json entries are
added only by the certification process, never by the submitting PR.
tools/skill-test is the same harness we run for official certification. Run it on your
own machine with your own model API keys:
cd tools/skill-test && bun install
bun run skill-test --skill ../../skills/pancakeswap-trading --trials 3Skill content is GPL-3.0, matching the Altana SDK.