APIblaze for Claude and ChatGPT

APIblaze’s own control plane is an MCP server. Add it to Claude (Settings → Connectors → Add custom connector, URL https://mcp.apiblaze.com) or to Claude Code:

claude mcp add --transport http apiblaze https://mcp.apiblaze.com

Sign in with your APIblaze account. Your assistant can then list and inspect your proxies, create a hosted MCP server for an API, mint and revoke API keys (shown once through a one-time link, never in the conversation), read usage, draft and apply access rules, explain a refusal, add a tenant admin and set rate limits. Destructive actions ask for confirmation.

Passing these checks means your MCP meets the requirements we can verify. Listing decisions are made by Anthropic and OpenAI.

npx apiblaze create --compliant --target https://yourdomain.com/openapi.yaml --domain mcp.yourdomain.com

One command, from the spec to the packet, to the letter of the directory's policies. create --compliant creates a directory-ready proxy from your OpenAPI document and checks it against the Claude directory (--compliant chatgpt or --compliant all adds OpenAI's; connector <project> runs the same flow on an existing one), runs the directory checks, applies the safe fixes (enables the MCP server, trusts Claude’s and ChatGPT’s OAuth clients, fills the docs URL from your portal), registers the MCP hostname and waits while you add its one CNAME record, hides utility endpoints such as health checks and docs pages, gives thin tool descriptions a first draft, asks for what only you know with every answer pre-filled from the spec and your site (name, tagline, privacy page, support contact, example prompts, the five attestations), creates the reviewer test account, and — when the report is green — writes the submission packet for both directories with the steps to submit it.

Stop any time and pick up where you left off. Every answer is saved the moment you give it, and the same scorecard — in the terminal and on dashboard.apiblaze.com/directory/<project> — says which of the twelve steps are done and what is left: finish the forms on the page, run the Claude Code pass from your terminal, or the other way round.

What directory-ready mode changes on your MCP server, and only there:

• tool names capped at 64 characters and de-duplicated when the catalogue is regenerated;
• PUT and DELETE marked destructive, PATCH and DELETE ask for confirmation, openWorldHint off unless you set it;
• errors become “status: what the API said”; request ids, traces and timings are stripped from results; results above the client budget are cut with a visible marker;
• a drained wallet answers “temporarily unavailable” (no checkout in a conversation);
• OAuth sign-in is accepted on the published URL even for key-only proxies; or, for public read-only data, no sign-in at all (--auth-mode open, asked once when every tool only reads; Anthropic allows it) while your API host keeps its own keys;
• browser origins are limited to the directories’ own and the ones you allow;
• 60 requests per minute per signed-in client;
• the OpenAI domain-challenge file is served on your custom MCP domain.

Nothing changes while the mode is off.

• A hostname on a domain you control for the MCP server, such as mcp.yourdomain.com, pointed at proxy.abz.run with one CNAME record. The command registers it and waits for it.
• A public privacy policy page and a square icon on your site. The command looks in the usual places (/privacy, /icon-512.png) and otherwise asks.
• A support contact, three example prompts and the five attestations — pre-filled from the spec, you confirm them.
• A reviewer test account — created for you (username + password sign-in on your tenant, stored encrypted for the packet, shown once). Not needed in open mode: the packet carries access instructions instead.
• Two verdicts, both binary: submit-ready (what the portal accepts) and review-ready (what a person checks on top: every tool answers, refuses a bad argument with a message, the reviewer account sees data, the docs read as prose, the three listing prompts work in Claude). The second comes from evidence the gateway records, never from a guess: npx apiblaze connector exercise <project> calls every tool as the reviewer (reads with a valid and an invalid argument; a write only after you answer y with the exact arguments shown) and, when Claude Code is installed, runs the listing prompts through it. The flow asks "Run every tool now? (Y/n)" at the end.
• The documentation page — generated on your Claude Code from the connector's facts (npx apiblaze connector docs <project>), served at <your portal>/connector, editable on the dashboard; or your own page, verified live. The portal's free-text answers (npx apiblaze connector answers <project>) and the packet laid out as portal steps 1–10. Data handling, health data and sponsored content are never pre-filled.
• A hosted report page on the dashboard (the first line every run prints) with the tool table, run history and replay, and an email at the end of every run — to you only; --no-email turns it off.

npx apiblaze check --directory-ready <project>

Prints the same report as the dashboard: “Claude: ready / ChatGPT: 2 blockers”, each blocker with the flag or field that fixes it.

See the privacy policy: API key reveal window, sign-in data, directory test accounts, AI assistants. Support: apiblaze.com/support.