| name | create-atomic-tool |
| description | Build a `BaseTool[InSchema, OutSchema]` subclass — input/output schemas, `BaseToolConfig`, `run()` (and optional `run_async()`), env-driven secrets, typed failure outputs. Use when the user asks to "add a tool", "create a tool", "wrap an API as a tool", "build a `BaseTool`", "make a calculator/search/weather tool", or runs `/atomic-agents:create-atomic-tool`. |
Create an Atomic Agents Tool
A tool is a deterministic capability an agent can invoke. In Atomic Agents, every tool is a BaseTool[InSchema, OutSchema] subclass with a typed run() (and optional run_async()). The input/output schemas double as the tool's signature for the LLM and as Pydantic validation at runtime.
For deep material (MCP interop, distributing as a standalone package, advanced error patterns), the authority is ../framework/references/tools.md. This skill is the action-oriented path: clarify → write → verify.
When this fires vs the umbrella framework skill
- This skill: the user is creating a specific tool — wrapping an API, building a calculator, scraping a page, querying a DB.
framework skill: questions about Atomic Agents in general, or the user is doing something other than authoring a tool.
Phase 1 — Clarify
Bundle into one message:
- What does the tool do? One sentence. This becomes the class docstring and feeds the LLM's tool description.
- Inputs and outputs. Names, types, units. If unclear, propose a schema pair and confirm.
- External dependencies. HTTP API? DB? Local computation only? If HTTP, what auth (API key env var, OAuth, none)?
- Sync, async, or both? If the rest of the project is async or the call is I/O bound, plan a
run_async() alongside run().
- Failure modes. Rate limits, not-found, network errors — how should the agent see them? Default: typed failure output, not raised exceptions.
Skip any question already answered in context.
Phase 2 — Plan
Confirm the location and shape in one short block, then proceed:
- File:
<project>/tools/<tool_name>_tool.py (in-project tool — see ../framework/references/project-structure.md).
- Schemas:
<ToolName>Input, <ToolName>Output, optionally a typed failure shape.
- Config:
<ToolName>Config(BaseToolConfig) if the tool needs API keys, base URLs, timeouts, retries.
- Sync vs async: pick one or both.
- Error pattern: typed failure output (preferred) vs raise (only for programmer error).
Phase 3 — Implement
Skeleton — local computation, no config
from pydantic import Field
from atomic_agents import BaseIOSchema, BaseTool
class CalculatorInput(BaseIOSchema):
"""Arithmetic expression to evaluate."""
expression: str = Field(..., description="Python-style arithmetic, e.g. '2 + 2 * 3'.")
class CalculatorOutput(BaseIOSchema):
"""Result of evaluating the expression."""
result: float = Field(..., description="Numeric result.")
class CalculatorTool(BaseTool[CalculatorInput, CalculatorOutput]):
"""Evaluate simple arithmetic expressions safely."""
def run(self, params: CalculatorInput) -> CalculatorOutput:
import ast, operator as op
ops = {ast.Add: op.add, ast.Sub: op.sub, ast.Mult: op.mul, ast.Div: op.truediv}
def ev(n):
if isinstance(n, ast.Constant): return n.value
if isinstance(n, ast.BinOp): return ops[type(n.op)](ev(n.left), ev(n.right))
raise ValueError("unsupported")
return CalculatorOutput(result=ev(ast.parse(params.expression, mode="eval").body))
Skeleton — HTTP-backed, with config and typed failure
import os
import httpx
from typing import Literal, Optional
from pydantic import Field
from atomic_agents import BaseIOSchema, BaseTool, BaseToolConfig
class WeatherConfig(BaseToolConfig):
api_key: str = Field(
default_factory=lambda: os.environ.get("WEATHER_API_KEY", ""),
description="API key for the weather service.",
)
base_url: str = Field(
default="https://api.weather.example/v1",
description="Base URL for the weather API.",
)
timeout: float = Field(default=15.0, ge=1.0, le=120.0, description="Request timeout (s).")
class WeatherInput(BaseIOSchema):
"""A request for current weather conditions."""
city: str = Field(..., description="City name, e.g. 'Brussels'.")
class WeatherOutput(BaseIOSchema):
"""Current weather conditions, or a typed failure."""
status: Literal["ok", "error"] = Field(..., description="Outcome code.")
temperature_c: Optional[float] = Field(default=None, description="Temperature in Celsius.")
summary: Optional[str] = Field(default=None, description="Human-readable summary.")
error: Optional[str] = Field(default=None, description="Failure message when status='error'.")
class WeatherTool(BaseTool[WeatherInput, WeatherOutput]):
"""Fetch current conditions for a city from the weather API."""
def __init__(self, config: WeatherConfig | None = None):
super().__init__(config or WeatherConfig())
def run(self, params: WeatherInput) -> WeatherOutput:
cfg: WeatherConfig = self.config
if not cfg.api_key:
return WeatherOutput(status="error", error="WEATHER_API_KEY not set")
try:
r = httpx.get(
f"{cfg.base_url}/current",
params={"city": params.city},
headers={"Authorization": f"Bearer {cfg.api_key}"},
timeout=cfg.timeout,
)
r.raise_for_status()
except httpx.HTTPError as e:
return WeatherOutput(status="error", error=str(e))
data = r.json()
return WeatherOutput(status="ok", temperature_c=data["temp_c"], summary=data["summary"])
async def run_async(self, params: WeatherInput) -> WeatherOutput:
cfg: WeatherConfig = self.config
if not cfg.api_key:
return WeatherOutput(status="error", error="WEATHER_API_KEY not set")
async with httpx.AsyncClient(timeout=cfg.timeout) as client:
try:
r = await client.get(
f"{cfg.base_url}/current",
params={"city": params.city},
headers={"Authorization": f"Bearer {cfg.api_key}"},
)
r.raise_for_status()
except httpx.HTTPError as e:
return WeatherOutput(status="error", error=str(e))
data = r.json()
return WeatherOutput(status="ok", temperature_c=data["temp_c"], summary=data["summary"])
Hard rules
- Generic parameters carry the runtime type info. Never also assign
input_schema / output_schema as class attributes — it shadows the framework-managed property.
run() returns the output schema instance, not a dict.
- Secrets via env /
BaseToolConfig, never hardcoded.
- HTTP tools always set a timeout. Tools run in the agent's request path.
- The async hook is
async def run_async, not arun — the framework calls run_async.
- Convert routine failures (rate limits, 404s, validation rejects from the upstream API) into a typed failure output. Reserve
raise for programmer error.
Phase 4 — Wire it into an agent
Two integration shapes (see ../framework/references/tools.md for more):
Single-tool agent — agent's output schema is the tool's input schema:
agent = AtomicAgent[UserQuery, WeatherInput](config=config)
tool = WeatherTool()
call = agent.run(UserQuery(question="weather in Brussels?"))
result = tool.run(call)
Router agent — agent picks among tools via a discriminated union of tool-call schemas. Use this when the agent has 2–10 tools to choose from. For dozens, see the search+execute pattern in ../framework/references/orchestration.md.
Phase 5 — Verify
uv run python -c "from <project>.tools.<tool_name>_tool import <ToolName>Tool, <ToolName>Input; t = <ToolName>Tool(); print(t.run(<ToolName>Input(...)))"
If imports fail with the docstring error, add the docstring on the schema. If self.input_schema is None, the generic parameters are missing — write class FooTool(BaseTool[FooInput, FooOutput]):, not class FooTool(BaseTool):.
Phase 6 — Hand off
Tell the user:
- Where the tool lives and what to import.
- How the agent should use it (single-tool or router shape).
- Optional next steps:
- The agent that calls it →
create-atomic-agent skill.
- Multi-agent wiring around the tool →
../framework/references/orchestration.md.
- MCP interop or packaging the tool for distribution →
../framework/references/tools.md.
Anti-patterns
class FooTool(BaseTool): with input_schema = ... class attributes — use generics: BaseTool[FooInput, FooOutput].
- Returning a dict or primitive from
run() instead of OutputSchema(...).
- Raising on routine upstream failures — model them as typed output.
- No timeout on HTTP / DB calls.
MCPTransportType.STREAMABLE_HTTP — the correct value is HTTP_STREAM.
- Implementing
async def arun(...) — the framework calls run_async.
For deeper material — MCP interop, packaging a tool for atomic download, advanced router patterns — load ../framework/references/tools.md.