Open standard for discoverability of paid HTTP APIs in agentic commerce (x402 / MPP). OASIS is a vendor-neutral discovery layer for paid HTTP APIs: a task ontology, a payment-aware endpoint index, JSON schemas, and a reference MCP server that map a natural-language task to the right paid endpoint — price and payment rails inline. Not a hosted product, no fees. See GOVERNANCE.md.
As paid endpoints multiply, agents hit registry noise, keyword collisions, and no reliable way to pick the right micropayment API. We could not find a high-quality, vendor-neutral discovery mechanism — so we built OASIS and open-sourced it.
| Artifact | Purpose |
|---|---|
ontology/intents/ |
Curated task capabilities — the controlled vocabulary agents route to (111 intents across 20 domains) |
dist/index.json (+ endpoints / capabilities) |
The unified, payment-aware paid-endpoint index (~19k gated endpoints) |
spec/ |
JSON schemas + entity vocab + agent traversal protocol |
mcp/ |
Reference MCP server (oasis_discover + contribution tools) + drop-in agent skill |
New here? Start with docs/concepts.md — the mental model and glossary (taxonomy, ontology, capability, domain, facet, entity, link, endpoint, binding).
Drop the reference MCP server into any agent. oasis_discover returns the right paid
endpoint for a task — plus a next_steps map to chain follow-ups — in a single call: a flat,
ranked endpoint list with price + payment rails inline. Works with any LLM provider.
A free hosted instance is live — no clone, no key:
claude mcp add --transport http oasis https://mcp.oasisindex.org/mcp(or any MCP client: { "mcpServers": { "oasis": { "url": "https://mcp.oasisindex.org/mcp" } } }).
Open + per-IP rate-limited, operated by the stewards as a convenience — not part of the
standard. To self-host the same image or run it locally over stdio, see mcp/ and
mcp/deploy/.
The server also exposes oasis_search (classify a query → task intents) and the
oasis_taxonomy / oasis_validate contribution tools. Teach your agent to call
oasis_discover first by dropping in mcp/skills/oasis.md (Claude
Code: ~/.claude/skills/oasis/SKILL.md).
No build needed to search or resolve — download dist/index.json (and siblings) from
Releases. To build the index yourself:
git clone https://github.com/Syntalic/OASIS.git && cd OASIS && pnpm install
pnpm run build # federated ingest → quality gate → semantic bind (needs network; GOOGLE_API_KEY for gemini binding)
pnpm run embed # vector index (111 curated intents, gemini-embedding-001)
pnpm testCLI (capindex):
pnpm exec capindex search "cheapest airpods pro" # vector search over curated intents
pnpm exec capindex resolve --intent commerce.compare_price # endpoints that satisfy a task
pnpm exec capindex validate # validate dist/index.json
pnpm exec capindex taxonomy --json # controlled vocab (to contribute)
pnpm exec capindex validate-source <intent.yaml> # SAME check CI runs on a PROn a battery of real, colloquial tasks, oasis_discover finds and picks a paid endpoint
in one call at 71% precision and the lowest token cost of every method tested (vs
keyword and the semantic-spec approach third-party registries use). On held-out queries phrased
away from the labels it routes at 95% discover@1 / 99% discover@3 (vs 41% for keyword). Full
numbers, the comparison to the largest live x402 layers, and reproduction:
docs/eval_results.md.
OASIS scales by LLM-assisted, contributor-funded curation: the service owner binds their
own endpoints into the taxonomy (their LLM, their cost) and opens a PR; CI runs an objective
validation gate (validate-source). See
docs/contributing-capabilities.md (bind endpoints into
the ontology) and docs/authoring-openapi-specs.md (author
a discoverable spec that passes the quality gate and ranks well).
- docs/concepts.md — start here: the data model + glossary (taxonomy, ontology, capability, domain, facet, entity, link, endpoint, binding)
- docs/taxonomy-methodology.md — how an intent is assigned to a domain (the finance↔blockchain rule, when a domain is justified, the tie-breaker ladder)
- ARCHITECTURE.md — index-build pipeline, search/retrieval, ontology→endpoint wiring
- spec/traversal.md — agent protocol: search → resolve → schema → execute
- docs/scaling.md — the endpoint-atomic direction and why it scales
- docs/eval_results.md — full benchmarks: accuracy, token cost, generalization, the agent probe
- docs/contributing-capabilities.md — add a service (bind endpoints into the task ontology)
- docs/authoring-openapi-specs.md — author a discoverable OpenAPI spec (what the quality gate + ranker expect)
- docs/index-snapshots.md — reproduce a pinned index anywhere (snapshot lock + restore)
- GOVERNANCE.md · CONTRIBUTING.md — what OASIS is (and isn't), and how to contribute
- docs/OASIS-explainer.pdf — visual explainer deck
MIT — see LICENSE.