Origo technical reference

Origo exposes the WebCTX retrieval engine through stateless MCP Streamable HTTP. It provides five tools: read_link, map_site, research, inspect, and execute. Federated web search is CLI-only.

Endpoint and access

https://api.origo.ashray.xyz/mcp?key=<your-private-key>

An Origo deployment must have its own ORIGO_API_KEY. Treat the full URL as a credential: query parameters can end up in logs, analytics, clipboard histories, and copied links. Do not embed this URL in public documentation, screenshots, or version control.

Origo runs as a stateless Go function, so it does not require a persistent MCP session store. A client needs an access key to use the hosted endpoint.

Tool behavior

MCP tool	CLI counterpart	Notes
read_link	read-link	Optional question finds relevant passages and related sources.
map_site	map-site	Bounded URL inventory, filters, and continuation.
research	research	Finds data source operations; no paid execution.
inspect	inspect	Returns input/output contract and price.
execute	execute	May spend Alexandria credits.

The tools advertise the MCP readOnlyHint, but execute can still trigger paid provider calls. Inspect the contract and price before using it.

URL retrieval strategy

read_link tries increasingly general ways to represent a source:

1. GitHub’s structured/raw data for recognized repository, file, issue, PR, commit, and CI URLs.
2. Direct text for JSON, YAML, OpenAPI, and similar source documents.
3. Published Markdown representations of normal documentation pages when available.
4. Firecrawl Scrape with main-content extraction and the provider’s own 30-minute cache.
5. A more capable Firecrawl proxy or a short-lived Browser Sandbox session if extraction fails.

Question mode can follow relevant same-host documentation links and consult published llms.txt or OpenAPI materials. Native GitHub permissions and rate limits are not silently replaced with scraped versions of protected data. See How URL reading works.

Data execution and billing

research and inspect provide free discovery and schema lookup; execute runs paid Alexandria operations. The hosted server enforces a default 200-credit upper bound per request. It validates input fields before submission, preserves provider-native results, and reports estimated and provider-reported usage separately.

Third-party provider terms can require human acceptance. The response supplies the terms URL; Origo does not accept it automatically.

For uncertain paid retries, reuse the original request_id and unchanged inputs. Identical replays on the same warm worker can use a 15-minute in-memory cache and avoid another upstream execution. Replays across separate workers are not guaranteed by the local cache; Firecrawl’s upstream idempotency remains authoritative.

Origo has no custom Redis, KV, or persistent cache service. Firecrawl’s 30-minute scrape cache is controlled by Firecrawl, not by the MCP server.

Deployment

Two Vercel projects use the same GitHub repository but deploy independently:

Surface	Project	Domain	Trigger
Documentation	webctx-docs	webctx.ashray.xyz	Relevant docs/site changes on main
Origo API	origo-api	api.origo.ashray.xyz	Retrieval/MCP changes on main

The Origo entrypoint is api/mcp.go; MCP implementation lives in pkg/origo and the shared retrieval adapter in pkg/retrieval. The underlying engine is in internal/app. CLI release tags independently trigger GitHub binary builds and npm publication.

For contributor workflows, see CONTRIBUTORS.md.