Update to current Orbit API and bundle Orbit's MCP server

[avdev4j]

Why

The skill was pointing at fabric-gateway.postmanlabs.com/api/search, which no longer exists. Orbit's API is now at api.buildwithorbit.ai/v1/search, documented at buildwithorbit.ai.

Rather than just swap in the new URL and wait for the next move, this bundles Orbit's MCP server so the API contract lives server-side. Orbit can change its parameters without breaking installed copies of the plugin.

Mirrors Postman-Devrel/orbit-claudecode-plugin#1.

What changed

Bundles the MCP server. New .mcp.json points at https://mcp.buildwithorbit.ai/mcp, referenced from the manifest as "mcpServers": "./.mcp.json". Installing the plugin now registers the search and integrate tools — no codex mcp add, no config.toml edits, no API key.

Skill drives the tools instead of curl. SKILL.md calls search and integrate and keeps everything that is actually this plugin's value: capability decomposition, reading Not supported: lines as design gaps, iteration, saving the blueprint to orbit-output/.

Adds integrate. Turns chosen endpoints into a task brief covering auth requirements, base URLs, ordered request steps, and gotchas. Wired into the README's design process as step 5.

Corrects the parameters. The old skill had several wrong or missing:

	Old	Current
cursor	mentioned, never used	tool argument / query param
limit	undocumented	default 10, max 25
q	unconstrained	max 512 chars
nextCursor	treated as nullable	absent on last page, not null
result fields	no resourceType	adds resourceType; also provider/product
pagination	unbounded	caps at 40 results total
integrate resources	n/a	1–10 items

Adds references/orbit-api.md. This repo had no reference file; it now documents the REST endpoints as a fallback for sessions where the MCP tools are unavailable.

README example output is now real. It previously showed a fabricated Stripe result; replaced with an actual Brevo result from a live call.

Codex-specific details

Codex's plugin format differs enough that copying the Claude Code .mcp.json verbatim would not have worked. Verified against the Codex source rather than the docs:

• The manifest needs an explicit pointer. Unlike Claude Code, Codex does not auto-discover .mcp.json — it requires "mcpServers": "./.mcp.json" in .codex-plugin/plugin.json. Without it the file is inert.
• The key inside .mcp.json is camelCase. OpenAI's docs say mcp_servers; PluginMcpServersFile carries #[serde(rename_all = "camelCase")], so the loader wants mcpServers. See Docs: plugin .mcp.json example uses mcp_servers but Codex expects mcpServers openai/codex#22105 (still open).
• Transport value is streamable-http. The native normalizer also accepts "http", but Codex's test fixtures use streamable-http on the parallel agent-plugin parse path, so that value is accepted by both. deny_unknown_fields on the transport enum makes a wrong key a hard parse error, not a warning.

Bundled HTTP servers are supported at all per the source comment: "Native plugin HTTP servers share the regular MCP transport configuration."

Verification

Probed the Orbit MCP server directly over JSON-RPC (initialize → tools/list → tools/call) and called both REST endpoints live, so the tool names and schemas are confirmed.

Codex-side loading is source-verified only — codex is not installed on the machine this was written on, so nothing was exercised end-to-end. Worth one /orbit:discover run on a real install. If the bundled server does not register, the README documents the ~/.codex/config.toml fallback.

Generated with Claude Code