MiniADK

简体中文

A small Python Agent Development Kit. Build tool-using agents with a compact, readable API. The core stays small; providers, tools, policies, skills, MCP, CLI adapters, and presets live around it.

MiniADK is a development kit, not a finished agent product. We ship the building blocks; you write the smarts.

Install

pip install miniadk

That's it for the core. The default terminal UI (Ink, written in TypeScript) is downloaded on first use of run_cli and cached in ~/.cache/miniadk/tui/. If you'd rather skip the network fetch:

pip install miniadk[tui-textual]   # pure-Python fallback TUI

Then set MINIADK_TUI_NO_FETCH=1 in your environment, or pass backend="textual" to run_cli.

For development:

git clone https://github.com/Saglear/miniadk.git
cd miniadk
uv sync --extra dev
uv run --extra dev pytest -q

Three quickstarts

1. Headless agent (≤ 15 lines)

from miniadk import Agent, run, tool

@tool
def add(a: int, b: int) -> int:
    "Return a + b."
    return a + b

agent = Agent("calc", "Use tools when they help.", tools=[add])
print(run(agent, "What is 17 + 25?"))

2. Default terminal UI (≤ 10 lines)

from miniadk import Agent, make_tools, run_cli

run_cli(Agent(
    "repo",
    "Answer questions about this repository.",
    tools=make_tools(write=False, shell=False),
))

3. Custom React/Ink UI (~30 lines, see examples/custom_tui/)

import { mount, BridgeProvider, useBridgeSend, useBridgeEvents,
         Markdown } from "@miniadk/tui";

function App() {
  const send = useBridgeSend();
  const [items, setItems] = useState([]);
  useBridgeEvents("message", (e) => setItems(p => [...p, e.data.text]));
  return /* …your layout… */;
}

mount((bridge) => <BridgeProvider bridge={bridge}><App/></BridgeProvider>);

Point Python at your binary via MINIADK_TUI_BIN and run_cli does the rest.

Models

model() reads provider settings from environment variables and returns a configured adapter:

ANTHROPIC_API_KEY=...    # or ANTHROPIC_AUTH_TOKEN
ANTHROPIC_BASE_URL=...   # optional, for proxies / Anthropic-compatible APIs
ANTHROPIC_MODEL=claude-opus-4-7

OPENAI_API_KEY=...
OPENAI_BASE_URL=...
OPENAI_MODEL=gpt-5-pro

When both are set, choose with MINIADK_MODEL_PROVIDER=anthropic (or openai).

Tools

Any typed Python function with a docstring becomes a tool:

from miniadk import tool

@tool
def now_utc() -> str:
    "Return the current UTC time, ISO-8601."
    from datetime import datetime, UTC
    return datetime.now(UTC).isoformat(timespec="seconds")

Sync and async are both fine. The decorator reads the function name, docstring, and type hints to build the JSON schema.

For batteries-included tools, make_tools returns a curated set:

from miniadk import make_tools

tools = make_tools(
    files=True,    # read_file, list_files, glob_files, search_text
    shell=False,   # subprocess.run wrapper
    write=False,   # mutating file ops
    web=True,      # fetch_url
)

Each tool keeps its own path checks, permission prompts, and limits — the runtime stays atomic.

Composition

Agent carries its own policy and middleware. Adapters never learn about preset shapes; they call resolve_composition(agent) once.

from miniadk import Agent, RunDecision

class StopAfterThreeTools:
    def __init__(self): self.rounds = 0
    async def after_model(self, state):
        r = state.result
        if r and r.message and not r.tool_calls:
            return RunDecision.stop(r.message)
        return RunDecision()
    async def after_tools(self, state):
        self.rounds += 1
        return RunDecision.stop("hit cap") if self.rounds >= 3 else RunDecision()

agent = Agent("bounded", "Answer briefly.", policy=StopAfterThreeTools())

Same pattern for middleware (before_tool_call to gate, after_tool_call to log). See examples/05_middleware.py and examples/08_custom_policy.py.

Skills and MCP

Skills are slash-launchable playbooks (prompt + allowed tools). MCP servers are external tool providers connected over stdio. Both resolve to plain Tools before the runtime loop runs.

from miniadk import Agent, MCPServer, run_cli, skill
from miniadk.mcp import MCPHub
from miniadk.skills import SkillRegistry

run_cli(Agent(
    "assistant",
    "Use available capabilities.",
    skills=SkillRegistry.from_skills(
        skill("review", "Read $path and summarise.", tools=["read_file"], args=["path"]),
    ),
    mcp=MCPHub([MCPServer(name="docs", command="uvx", args=["some-mcp-server"])]),
))

Examples

A teaching ladder, plus practical tools — see examples/README.md for the index. Highlights:

• 01–05 — Agent, tools, streaming, sessions, middleware (concepts).
• 06–07 — Default CLI; a read-only repo assistant.
• 08 — Write your own RunPolicy (this is how you'd build ReAct or Plan-and-Execute, instead of importing a preset).
• 09–10 — MCP client; slash-skill router.
• custom_tui/ — Replace the entire terminal UI with your own React components while reusing the bridge.

uv run python examples/01_hello_agent.py
uv run python examples/06_run_cli.py

Architecture

Two-layer separation:

adapters/      tui_ink   tui_textual   json   web   ws    (lazy, opt-in)
core/          Agent  Tool  Model  Runtime  Session  Event  RunPolicy

import miniadk loads only the core — no Textual, no Ink, no React. TUI deps are resolved on attribute access. See docs/architecture.md for layering and extension points, and docs/tui-protocol.md for the JSON wire format between Python and the Ink subprocess.

What MiniADK does not provide

By design — these belong to your application, not the framework:

• No agent-loop preset zoo (ReAct, Plan-and-Execute, Tree-of-Thought, reflection). Compose them out of RunPolicy + middleware.
• No prompt template library. Strings work fine.
• No retrieval / vector store integration. Wrap one in a Tool if you need it.

The agentic() preset is the one opinionated composition we ship. Treat it as an example you can read and copy, not as the canonical way.

License

MIT