MCP
Every Flux API ships with a built-in Model Context Protocol server — seven tools over the same collections and permissions as the REST endpoints.
An agent connects with a key and finds out for itself what it can reach, what shape the data is, and how to search it. The tool definitions come from the server, so there is no SDK to install and no integration code to keep in step with your schema.
Where the server lives
Two addresses reach the same server, with the same tools and the same permissions.
https://mcp.foxnose.net/_mcp — the single endpoint. The key carries the environment, so this one URL works for every environment and every API. Use it for a catalogue listing, a hosted connector, or any client with a single URL field. It accepts Simple and Bearer.
https://{environment_key}.fxns.io/{api_prefix}/_mcp — the environment's own address. The environment and the API are named in the URL. Use it when requests are signed with Secure, whose signature covers the request path — and here the path names the API — or when the API is public and the client sends no credential at all.
Single Endpoint compares them in full.
What the server can do
| Tool | What it gives an agent |
|---|---|
discover_resources | Every reachable collection, and what can be done with each |
describe_resource | A self-contained JSON Schema, and which fields are searchable |
query_records | Records by filters, sorting and a cursor |
search_records | Text, semantic, hybrid or vector-boosted retrieval in one call |
get_record | One record, in full |
create_record | A new record, published immediately |
update_record | A new published revision, with optimistic concurrency |
The five read tools are always present. The two write tools appear when the key grants them.
Retrieval is hybrid in one call. search_records matches literal text and meaning together and merges them into a single ranking, over the fields marked vectorizable. The question is sent as text and embedded for you, there is no separate vector store to run alongside, and filters apply to a semantic question exactly as they do to a plain query.
An agent sees only what its key sees. Collections are connected to an API with the read methods you choose, and the key's role decides the rest. MCP adds no second access-control surface — a collection a key cannot read over REST is not reachable over MCP either.
Agents can write back. With the grants for it, an agent creates and updates records through the same connector it reads from, running the full publish pipeline: schema validation, a revision, and re-vectorization. There is no delete tool.
What the Tools Do covers how these compose into a session.
Where to go
- What the Tools Do — the seven tools and how an agent composes them into a session.
- Single Endpoint —
https://mcp.foxnose.net/_mcp, one address for every environment, resolved from the API key. Start here unless you needSecureor anonymous access. - Connect Claude — wire up Claude Code, another MCP client, or a hosted connector using a bearer token.
- Connect ChatGPT and OpenAI — the OpenAI Responses API path, the ChatGPT UI path, and the current public-vs-private limitation.
- Preparing a Collection — how to describe a collection so an agent picks the right one, which read methods to connect, and which fields to make searchable.
- Authentication — which of the three schemes to send, and how fast each way of revoking access takes effect.
- Retrieval & Cost — what the default page size and text truncation mean for what you get back, and why the bill is dominated by re-reads.
- Limits — the protocol version served, and what isn't implemented yet.
For the concept, start at Agent-Native Flux APIs; for inputs, outputs and error codes, go to the MCP Server reference.