| name | forge-scout |
| description | Query external sources (MCP Registry, anthropics/skills, installed marketplaces) for candidate implementations of a specified capability. Applies the 6-rule trust heuristic to each candidate and returns ranked recommendations. Use when forge-gap has identified a gap and forge-lead needs to decide whether to install an existing implementation or author a new one. Only use for single-query bounded lookups — for competitive cross-roster analysis, forge-lead must delegate to research-github-miner instead. |
| disable-model-invocation | true |
| allowed-tools | Bash(gh api *) Bash(gh search *) Read Glob WebFetch |
Forge scout
Primary data sources
-
MCP Registry (STRONG-PRIMARY): https://registry.modelcontextprotocol.io/v0/servers
- Endpoint:
GET /v0/servers?search=<keyword>&limit=100
- Cursor pagination:
metadata.nextCursor opaque string
- Recent-updates:
?updated_since=<RFC3339>
- Public read, no auth.
-
anthropics/skills (STRONG-PRIMARY): gh api repos/anthropics/skills/contents/skills
- Returns the official Anthropic skill directories.
- For each match:
gh api repos/anthropics/skills/contents/skills/<name>/SKILL.md to pull the frontmatter.
-
Installed marketplaces (STRONG-PRIMARY, on disk): ~/.claude/plugins/marketplaces/*/
huggingface-skills/skills/<name>/SKILL.mdai-research-skills/NN-<topic>/<name>/SKILL.mdclaude-code-skills/engineering-team/<name>/SKILL.mdclaude-plugins-official/<plugin>/ via installed_plugins.json
-
Installed plugins (~/.claude/plugins/installed_plugins.json).
The 6-rule trust heuristic for MCP Registry entries
Before recommending any MCP server, apply these checks (ALL must pass for HIGH trust):
- Package registry preference:
packages[].registryType must be npm | pypi | oci | nuget. Reject smithery-only remote unless the Forge has a smithery API key.
- Github repository:
server.repository.url must be populated with a github.com URL. Reject opaque entries.
- Version maturity:
server.version >= 1.0.0 OR multiple iterations in the version history. Reject single one-shot 0.1.x unless no alternative exists.
- Description stability: if multiple versions exist, descriptions must be consistent across versions. Flag description churn.
- Transport availability:
packages[].transport.type must include stdio for local runnability. Accept streamable-http only if stdio is also available.
- Auth profile: prefer
authorization: none for read-only tools. Accept auth-required only if the server is load-bearing.
Servers passing >=5 rules: HIGH trust. >=3: MEDIUM. <3: REJECTED.
Query protocol
Input: capability description (e.g., "Semantic Scholar paper search") from forge-gap or forge-lead.
Steps:
- MCP Registry:
gh api "https://registry.modelcontextprotocol.io/v0/servers?search=<keyword>&limit=50". If nothing found, broaden to synonyms.
- anthropics/skills: check if one of the official skills already covers the capability (use the name + description match).
- Installed marketplaces: Glob for
SKILL.md matching keywords in their description.
- Apply trust heuristic to each candidate.
- Cross-validate high-value candidates by fetching the candidate's README or SKILL.md (via WebFetch or on-disk Read).
Output format
Write to ~/.claude/forge/scout-reports/<timestamp>-<capability>.md:
# Scout report — <capability>
## HIGH-trust candidates
1. <name> (<registry>/<identifier> v<version>, <stars>*)
- URL: <repo_url>
- Trust rules passed: 1,2,3,4,5,6
- Install command: <exact command>
- Why adopt: <1-sentence>
## MEDIUM-trust candidates
...
## REJECTED
...
## Recommendation for forge-lead
<INSTALL|ADOPT|AUTHOR|DEFER> — <rationale>
Hard rules
- Never recommend INSTALL without verifying the trust heuristic passes >=5 rules.
- Never rank by stars alone — the 6-rule heuristic is binding.
- If no HIGH-trust candidate exists, recommend AUTHOR or DEFER, not a MEDIUM adoption.
- If the capability needs cross-roster competitive analysis, recommend forge-lead delegate to research-github-miner via a research-request drop-file.
