AgentConnect

For agents and builders

Discover through
the engine.

The signed AgentConnect pipeline app keeps reviewed catalog records alongside your saved directory. Discovery reads the published catalog dataset; it cannot read your private saved-service dataset.

Choose your access path

The public catalog JSON is a reviewed snapshot. Live queries use your invited AgentConnect installation and its ordinary account permissions. Anonymous live discovery and hosted MCP are not enabled.

CLI

Use your existing RAC authentication, configured native DATA_ENGINE_URL and workspace scope. Replace INSTALLATION_ID and WORKSPACE_ID with the canonical installation and workspace returned by your app directory.

rac app invoke INSTALLATION_ID --tenant-id WORKSPACE_ID -- discover --where capability=research --where availability=preview
rac app invoke INSTALLATION_ID --tenant-id WORKSPACE_ID -- discover-mcp --limit 20
rac app invoke INSTALLATION_ID --tenant-id WORKSPACE_ID -- discover --where id=redgold-search

Filters execute in the engine. Interface commands are discover-api, discover-mcp, discover-cli and discover-website. Results include provider, pricing, access, review dates and official source URLs. Pages are bounded; a full page may have more results.

MCP tools

The repository provides a thin stdio adapter with discover_services and get_service. It invokes only the installed read pipelines through RAC. It starts no database, stores no credentials and executes no advertised provider tools.

python3 tools/discovery.py --installation INSTALLATION_ID --tenant-id WORKSPACE_ID --mcp

Configure that command in your MCP client from a reviewed checkout of the AgentConnect repository. Discovery supports exact capability, interface and availability filters, a limit of 1–50 and an offset up to 500. RAC must already be authenticated for that installation.

Native HTTP API

The app uses the engine's authenticated HTTP interface and native CBOR encoding. Read the current installed release and its exact pipeline pins first:

GET /api/apps/installations/INSTALLATION_ID/describe
POST /api/apps/shared/execute
Content-Type: application/cbor
Authorization: Bearer YOUR_ACCOUNT_CREDENTIAL
x-redgold-tenant: WORKSPACE_ID

Encode a native DataflowExecuteRequest containing app_installation_id = the exact installation ID, pipeline_spec_name = agentconnect-catalog-list and pipeline_spec_hash = the exact 32-byte hash from the installation. The response is native DataflowExecuteResponse CBOR, with rows_json containing the catalog rows. Use the pinned interface-specific pipeline for interface filtering. Do not substitute a source, dataset or pipeline hash from an untrusted listing.

These relative paths are served by the configured application API origin. Use the integration's verified API origin and current installation contract; the static JSON file is not that API. Missing access is an account setup result, not permission to borrow another user's credentials.

Connect Search

  1. Open the provider connection guide and confirm development account access.
  2. Create a web_search-scoped key with a spending limit in development key settings. Keep it in your credential manager.
  3. Request a quote. Quick is 1¢, Standard 5¢ and Deep 10¢ retail; Deep is the provider default.
  4. Authorize execution separately with a cap and stable request identifier.
  5. Read findings, reference links, billing state, charged amount and receipt ID. Preserve the original request on an uncertain response.
python3 tools/search_connection.py "your research question" --tier quick
python3 tools/search_connection.py "your research question" --tier quick --execute --maximum-charge-cents 1 --request-id YOUR-STABLE-ID

Prefix either command with rac run --with-secret YOUR_KEY=AGENTCONNECT_SEARCH_KEY -- to inject your own scoped key. The first command only quotes. The second can spend prepaid credit. The helper checks the live quote and sends max_cost_usd for provider admission before executing and uses the original provider receipt. Neither command saves keys or personal queries in catalog records.

Metadata maintenance

Trusted workspace editors can retain documentation/link checks as separate engine observations. A changed hash, failed request, expired review date or initial baseline requires review. The checker fetches only reviewed HTTPS sources, follows no redirects, caps responses at one MiB and never changes published catalog records. Review and apply metadata changes explicitly. A successful documentation retrieval does not prove service use, billing or destination security.

Back to AgentConnect