Core concepts

Tools

Everything an agent can do becomes a schema-carrying Tool and registers into the gateway — so MCP servers, REST APIs, Skills, and RAG are all governed, audited, and argument-validated identically.

The Tool model

A Tool has a name, description, JSON-Schema parameters, and an invoke callable.

tools.py
from antraft import ToolRegistry, tool_from_function
 
def multiply(a: float, b: float) -> float:
    "Multiply two numbers."
    return a * b
 
tools = ToolRegistry([tool_from_function(multiply)])
# schema (incl. required + closed additionalProperties) is inferred from hints

MCP — local and remote

mcp.py
from antraft import tools_from_mcp_stdio, tools_from_mcp_http
 
# local MCP server over stdio
tools.extend(tools_from_mcp_stdio("npx", ["-y", "@modelcontextprotocol/server-github"]))
 
# remote MCP server over streamable HTTP (or SSE)
tools.extend(tools_from_mcp_http("https://my-mcp.example.com/mcp"))

REST / OpenAPI

rest.py
from antraft import rest_tool, tools_from_openapi
 
tools.add(rest_tool("get_user", "GET", "https://api.example.com/users/{id}",
                    description="Fetch a user"))
 
tools.extend(tools_from_openapi(openapi_spec_dict))   # one Tool per operation

Anthropic Skills & RAG

more.py
from antraft import tools_from_skills_dir, retriever_tool, InMemoryRetriever
 
tools.extend(tools_from_skills_dir("./skills"))
 
r = InMemoryRetriever().add(["doc one", "doc two"])
tools.add(retriever_tool(r, k=4))   # RAG, governed like any tool
Note
Whatever the source, every tool registers into the ToolGateway, so it inherits argument validation, RBAC, retries/timeouts, and the audit trail with zero extra code.

Schema validation

The gateway validates an action's arguments against the tool's JSON Schema before execution. Missing, mistyped, or unexpected arguments are rejected — a malformed or abusive call never runs.

Next →
Models