Get started

Jeff with coding agents

Jeff cannot serve as the underlying model for Claude Code, Cursor, Codex or other coding agents. It does not chat, write code or stream text; each function returns one typed decision. Coding agents can use Jeff's functions as tools and refer to these docs when writing code that calls Jeff.

Give the agent Jeff's functions as tools (MCP)

Jeff exposes its functions as MCP (Model Context Protocol) tools, one tool per function, plus jeff_dispatch (let Jeff pick the function) and jeff_list_functions (the same listing as GET /v1/functions). Any MCP-aware agent can then call the prompt injection guard on a piece of text without knowing Jeff's API. Each tool returns the whole decision, abstained included, so the agent can identify low-confidence decisions. There are two ways to connect.

Over HTTP. The Jeff server itself serves the tools at https://api.jeffhub.ai/mcp (streamable HTTP), behind the same bearer token as /v1/decide. Add it to your agent as a remote MCP server with the header Authorization: Bearer <your token>.

Over stdio, with jeff mcp. For an agent that starts its MCP servers as local commands, the jeff command serves the same tools and calls the server for each one. It is included in the Python package jeff-sdk with its mcp extra:

pip install "jeff-sdk[mcp]"

Then add it to the agent's MCP settings. Most agents accept a block like this:

{
  "mcpServers": {
    "jeff": {
      "command": "jeff",
      "args": ["mcp", "--server", "https://api.jeffhub.ai", "--token-env", "JEFF_API_KEY", "--timeout-s", "30"],
      "env": {"JEFF_API_KEY": "<your token>"}
    }
  }
}

--token-env names the environment variable that holds the token, so the token never appears in the list of running processes. --timeout-s is required: it bounds how long each call may take. MCP describes the tools, their arguments and their errors in full.

Give the agent the docs

When the agent writes code that calls Jeff, point it at:

• this site's pages, starting with the Quick start and Functions and /v1/decide;
• the OpenAPI description of every endpoint, request, response and error, at https://api.jeffhub.ai/openapi.json (also shown as the API reference);
• the catalogue page of each function it will call: its context fields, its input and its answers.

When Jeff is worth reaching for

An agent or application can use Jeff to:

• check text before a model acts on it (a prompt injection, a request it must not follow);
• pick where something goes (which tool, which team, which queue) and know how confident the choice is;
• put something on a scale (how urgent, how risky) and compare it with a threshold;
• avoid asking a chat model to "answer in JSON" and checking whether it did: a function can only return its own answers.