An API does not become useful to agents just because it has an OpenAPI file. The spec gives you paths, parameters, request bodies, responses, and auth. Agent hosts also need a small amount of product judgment: which tools are safe, which tools need approval, which host gets which config, and what a maintainer reviews before anything goes live.
That is the job of an Agent Interface Pack.
The pack is a generated artifact that turns one OpenAPI spec into a reviewable bundle for ChatGPT, Grok, Claude, Hermes, OpenClaw, Cursor, VS Code, Cline, Continue, and generic MCP clients.
What goes in the pack
Sourced's Agent Interface Pack includes more than source code:
- A normalized operation inventory.
- A TypeScript MCP server with stdio and HTTP entrypoints.
- A safety report that separates read, sensitive read, write, and destructive operations.
- Host-specific connector files for ChatGPT, Grok, Claude, Hermes, and OpenClaw.
- Local install snippets for desktop MCP clients.
- A default allowlist that starts narrow.
- Harness tests that prove the generated metadata matches the tool inventory.
- A draft PR plan that explains files, risks, test output, and approval gates.
The important part is that this is one generated review package. The user does not have to collect URLs from five dashboards and paste them into five different tools before they can understand whether their API is safe to expose.
Why OpenAPI is the right source
OpenAPI already contains most of the hard inputs:
operationIdbecomes a stable tool name.summaryanddescriptionbecome the first draft of tool copy.- path/query/body schemas become tool arguments.
securitySchemesexplain auth.serversexplains where the generated server calls.
Sourced does not treat all operations equally. Listing invoices and refunding invoices are both OpenAPI operations. They are not equal agent tools.
The first review screen
Before a connector is installed, Sourced shows:
- total tool count
- safe default tools
- sensitive read tools excluded from defaults
- write and destructive tools
- vague descriptions
- missing or synthesized operation IDs
- OAuth or OpenID Connect blockers
- base URL and HTTPS problems
The page answers one question: can this API safely start with a read-only connector review?
Host-specific output
Each host wants a slightly different shape.
ChatGPT API uses remote MCP tool config for the Responses API. Claude web uses remote MCP connectors, while Claude Desktop can still use a local stdio config. Grok needs a public HTTPS MCP endpoint and a careful allowed_tools list. Hermes can use an mcp_servers config. OpenClaw gets an MCP/plugin config bundle unless the API is actually replacing a native agent harness.
The generator keeps those differences out of the user's head as much as possible. Upload once, review once, then copy or install the right target.
What Sourced generates today
Sourced's OpenAPI-to-MCP generator emits an Agent Interface Pack with:
adapter.ir.jsonadapter.manifest.jsonagent-readiness.mdconnector.chatgpt.mdand.jsonconnector.grok.md,.json, xAI SDK, OpenAI-compatible SDK, and cURL examplesconnector.claude.mdand.jsonconnector.hermes.config.ymlconnector.openclaw.bundle/- harness tests
PR_PLAN.md
It keeps generated repo writes behind explicit approval. It also keeps MCP access tokens out of generated files; token-bearing snippets are copied only when the user explicitly chooses that action.
The right launch path
The safest launch order is:
- Generate the pack from OpenAPI.
- Review the safety report.
- Run local tests.
- Deploy the MCP HTTP endpoint behind HTTPS.
- Verify endpoint auth.
- Copy the host connector config.
- Enable only the default read-only allowlist.
- Add write tools later, one by one.
That is not just safer. It is less annoying. Sourced keeps the first action focused on generating a pack, then uses the selected host to show the next missing setup step.
Generate an Agent Interface Pack or start with the Grok connector guide.