2.4 compatibility in Nexa 3.0

This page describes the retained 2.4 language/runtime path. Sources without a version declaration use this path. Start new Open Programs with Nexa 3.0; compare contracts in the migration guide.

Tools, Protocols, and Python extensions

Use @tool to declare arguments and output, and uses to select exposed tools. python! embeds trusted local code in a tool implementation.

Complete tool program

Save this as calculator.nx, configure a model, and run uv run nexa run calculator.nx. The Agent can call add(a=2, b=3); the tool returns integer 5, which the model uses in its reply.

@tool("Add two integers")
fn add(a: int, b: int): int {
    python! """return a + b"""
}
agent Calculator uses add { prompt: "Calculate with the supplied tool." }
flow main { result = Calculator.run("2 + 3"); }

Schemas validate tool arguments. Unexposed tools, malformed JSON, wrong types, and extra fields are rejected. For tools writing files or contacting services, declare risk with the approval guide and handle unknown outcomes.

Structured output

This complete program asks for an object with an integer value, such as {"value": 42}. Once validated against the Protocol, the result is available by field.

protocol Answer { value: int }
agent Reader implements Answer { prompt: "Return the answer as structured data." }
flow main { result = Reader.run("Return 42"); }

Protocol, streaming, and cached outputs all pass through runtime output validation. Task-specific correctness needs explicit acceptance conditions after schema validation.

Extension boundary

python! executes trusted code with process permissions. sandbox: true requires a real isolation backend. Rust WASM has its own module execution scope. MCP tools use official SDK sessions; see Ecosystem and standard library.

---

Source and validation · tools / types