enlace_connector: Skill for Claude Code

corpus-connector — an ir corpus → a deployed MCP connector

A clean-separation pipeline (each layer has one owner; nothing knows about the next): ir indexes + searches, py2mcp wraps a search function as MCP, enlace_connector deploys any MCP connector, enlace_auth gates it. This skill is only the composition; don't push corpus/connector logic into the wrong layer.

sources ──ir──▶ named corpus ──ir.make_search──▶ search fn ──py2mcp/FastMCP──▶ MCP app
   └── enlace_connector.generate_deploy_bundle ──▶ app.toml + systemd + provision + allowlist + runbook
   └── ship data + provision + deploy.py + allowlist + client accounts ──▶ live, access-controlled connector

Inputs to gather first

• Corpus: either an existing registered ir corpus name, or a name + sources to index from scratch (a folder of files, a mapping, etc.).
• Client name / connector name (→ app name ­{name}_mcp, route /api/{name}_mcp).
• Allowed users (emails) — who may use this connector (access control).
• Free port on the box (8010/8011/8020 taken; use 8031+, one per connector).

Step 1 — Resolve or build the corpus (owner: ir)

Existing corpus → just use its name. New corpus:

import ir
src = ir.CorpusSource.from_files("/path/to/sources", name="<corpus>", pattern=r".*\.md$")
# or: ir.CorpusSource.from_mapping({...}, name="<corpus>")
ir.register("<corpus>", "files", root="/path/to/sources")   # persist to the registry
corpus = ir.build(src)                                        # embed + persist (XDG store)
print(ir.tools.search("smoke test query", corpus="<corpus>", k=3))  # verify it retrieves

If the sources need custom parsing/chunking (like trufflepig's P2/FileMaker), write a small indexer package with a CorpusSource + a strategy (see truffle as the reference) — that indexing code is the only per-client code, and it belongs in its own package, not here.

Step 2 — Generate the connector (owners: ir + py2mcp + enlace_connector)