Skip to main content
Glama
Sourcey

Sourcey

Official
README.md
# Sourcey

<img src="assets/sourcey-mark.png" alt="Sourcey mark" width="96" />

Find evidence-backed startup credits and deals, inspect Agent Readiness report cards, and
prepare canonical service declarations from an agent client.

Sourcey is already a hosted Streamable HTTP MCP server:

```text
https://mcp.sourcey.com/mcp
```

Gemini CLI can install the hosted connection and bundled `sourcey-research` skill directly from this repository:

```bash
gemini extensions install https://github.com/sourcey/mcp-server --ref main
```

`--ref main` installs the current extension from the repository. Without it, Gemini CLI selects the older v1.0.0 release archive, which predates the extension manifest. The extension uses the remote Sourcey data MCP; it does not start the npm stdio bridge or invoke Sourcey's separate documentation generator.

Use that URL directly when your client supports remote MCP. This package is the thin stdio
bridge for clients that require a local command:

```bash
npx -y @sourcey/mcp-server
```

```json
{
  "mcpServers": {
    "sourcey": {
      "command": "npx",
      "args": ["-y", "@sourcey/mcp-server"]
    }
  }
}
```

The bridge downloads no offer dataset and contains no tool implementations. At startup it negotiates
with the live Sourcey server, reads the current tool definitions once, and relays tool calls. A
data or tool-description update therefore needs no package release.

## Agent plugins

This repository also packages the **hosted Sourcey connection** for Claude, Cursor, Grok Build and portable ChatGPT/Codex distribution. Host manifests are in `.claude-plugin/`, `.cursor-plugin/` and `.grok-plugin/`; root `plugin.json` carries Agent Plugins metadata for ChatGPT/Codex. Portable `mcp.json` and host-specific `.mcp.json` point directly to `https://mcp.sourcey.com/mcp`. The bundled [Sourcey research skill](skills/sourcey-research/SKILL.md) shows how to check evidence and eligibility before recommending an offer. Marketplace review and live cards are separate from the source package; no ChatGPT/Codex approval is claimed here.

The Claude directory uses the scoped [plugin folder](plugins/sourcey/README.md) at `plugins/sourcey`. It contains only the hosted connection, research skill, mark, license and README. `npm run generate:metadata` copies its shared files from the existing root sources; `npm run check:metadata` verifies the exact file set and bytes. The optional stdio bridge, its dependencies and repository ownership proofs stay outside that installable folder. Other clients can continue using the root manifests.

The plugin connects to Sourcey's startup-offer and Agent Readiness data. It does **not** run Sourcey's separate documentation generator or pay for Sourcey's HTTP services. The current square mark is in `assets/sourcey-mark.png`; [privacy](https://sourcey.com/privacy), [terms](https://sourcey.com/terms) and [support](https://sourcey.com/support) are available on the product site.

Try these prompts after installing the plugin:

- “Find current cloud-hosting startup offers in Sourcey. Show the original source, last-checked date, and any eligibility unknowns for three results.”
- “Compare two of those offers using their recorded terms and evidence. Separate confirmed conditions from anything Sourcey has not verified.”
- “Find a Sourcey Agent Readiness report card and explain the tested workflows, blockers, and claims that remain unverified.”

The Claude plugin's `.mcp.json` declares only the public HTTPS Sourcey MCP server. It does not run this repository's optional npm stdio bridge or read a local credential. Search terms and record references deliberately sent to the hosted MCP are handled under [Sourcey's privacy policy](https://sourcey.com/privacy); request bodies are not persisted as product records.

## What agents can do

- search and compare current startup credits, discounts, free tiers, programs, perks, and deals;
- inspect exact evidence, freshness, eligibility unknowns, and the signed change feed;
- search and inspect Agent Readiness report cards, including visible grades, five-stage findings,
  blockers, remediations, methods, coverage, and freshness;
- prepare canonical Agent Readiness declaration YAML without assigning a grade; and
- discover Sourcey's paid HTTP services and x402 endpoint without paying inside MCP.

Call `tools/list` for the exact live schemas and descriptions. Sourcey's full agent guide is at
[sourcey.com/SKILL.md](https://sourcey.com/SKILL.md), its OpenAPI contract is at
[api.sourcey.com/openapi.yml](https://api.sourcey.com/openapi.yml), and x402 discovery is at
[api.sourcey.com/.well-known/x402](https://api.sourcey.com/.well-known/x402).

## Boundary

This package performs only two actions: MCP over local stdio and MCP over the configured hosted
HTTP endpoint. It has no wallet, payment signer, credentials, telemetry, browser, background
daemon, mutable plugin, Catalog data, Agent Readiness executor, or submission authority. Paid
work remains on Sourcey's canonical HTTP API.

`SOURCEY_MCP_URL` may select another HTTPS Sourcey endpoint. Plain HTTP is accepted only for a
loopback test server. URLs containing credentials or fragments are refused.

## Development

```bash
npm ci
npm run verify
docker build -t sourcey-mcp .
```

The verification path builds from clean output, checks all generated metadata, tests the real
HTTP-to-stdio boundary, creates an SBOM, installs the exact packed tarball into a blank project,
and repeats protocol introspection from those installed bytes.

MIT. See `SECURITY.md` and `SUPPORT.md` for reporting paths.

TDQS

A3.9/5.0

Scored across 8 tools

Disambiguation5/5

Each tool targets a distinct action and resource: search/compare offers, inspect evidence, track changes, search/get readiness, prepare a declaration, and discover services. The offer and readiness tools are clearly separated, and search vs. get/compare boundaries are explicit.

Naming Consistency5/5

All eight tools follow a consistent snake_case verb_noun pattern: compare_offers, search_offers, inspect_evidence, get_changes, search_readiness, get_readiness, prepare_declaration, get_services. There are no mixed styles or vague standalone verbs.

Tool Count5/5

Eight tools is well scoped for a server covering startup offers/credits plus Agent Readiness reports. Each tool has a clear role, and the set avoids both thinness and unnecessary sprawl.

Completeness4/5

The offer side covers discovery, comparison, provenance, and change tracking; readiness covers search, detail, and declaration preparation. Minor gaps remain, such as no direct single-offer retrieval tool and no MCP tool for submitting or publishing a prepared declaration, though the latter is explicitly delegated to the product HTTP API.

Maintenance

ActivityMaintained
ResponsivenessLimited data