nooa-tools-and-skills

Tools, Skills, and Integrations

Methods are the tools

There is no tool-registration abstraction: any visible method or attribute on self is callable from CodeAct-generated code, with doc(self) telling the LLM what exists. Mix deterministic helpers (SW1) with generation methods (SW3):

class InventoryAgent(Agent, llm=llm):
    """Checks inventory using deterministic helpers."""

    def get_stock(self, item: str) -> int:          # SW1: a "tool"
        return self.inventory.get(item, {}).get("stock", 0)

    async def can_fulfill(self, items: list[str], budget: float) -> Result:
        """Check whether the order fits the budget. Use self.get_stock()."""
        ...                                          # SW3: LLM orchestrates the helpers

Design rubric: exact semantics (matching rules, parsing, arithmetic) → deterministic helper; fuzzy judgment → generation method. Give helpers precise names, type hints, and docstrings — those are their tool schema.

Built-in tools

from nooa.tools import ShellTools, TodoManager

class DevAgent(Agent, llm=llm):
    def __init__(self):
        super().__init__()
        self.shell = ShellTools(cwd=".")          # persistent shell + file ops
        self.todos = TodoManager()                # t = self.todos.add("step", deps=[...]); self.todos.done(t.id)

ShellTools is the one shell/file tool (it replaced the removed BashTool/FileTool). One persistent bash session — cd, export, and cwd survive across calls — and four methods:

r = await shell.run("grep -rn 'def foo' src/")   # ShellResult: str subclass; .stdout/.stderr/.returncode/.success
await shell.run("python -", stdin=script)        # pass payloads via stdin= instead of quoting heredocs
region = await shell.read("f.py", lines=(10, 25))  # view a file region -> Match
await shell.replace(r.matches[0], new_code)      # edit at a Match anchor (pure grep/rg results carry .matches)
await shell.write_file("out.py", content)        # create/overwrite

Tools attached as public attributes are automatically visible; the LLM discovers their APIs via doc(self.shell).

For an object the LLM might misuse, prefer wrapping the dangerous surface in a small deterministic method over exposing the raw object.

Agent skills (Skill / TextSkill / SkillRegistry)

Skills inject curated guidance + optional helper APIs into an agent. Three forms:

from nooa import Skill, TextSkill

class MyAgent(Agent, llm=llm):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        # 1. Subclass with a docstring-as-usage-guide (methods become the API)
        self.git = GitWorkflow()                        # class GitWorkflow(Skill): """guide..."""
        # 2. Wrap a third-party library for discovery
        self.pd = Skill(pd)                             # LLM uses self.pd.read_csv(...)
        # 3. A SKILL.md directory on disk
        self.frontend = TextSkill(path=Path("skills/frontend-design"))

• Agents see a # Skills block (one-liner per skill) in their execution context and call doc(self.<skill>) for the full guide.
• TextSkill exposes read_file(path) and await run_script(name, *args) for files/scripts bundled with the SKILL.md.
• SKILL.md frontmatter: required name, description; optional compatibility, metadata, user-invocable, allowed-tools — Claude-Code-compatible format.
• Bulk discovery/activation is SkillRegistry(agent) + discover_skills_dirs([...]) / activate([...]). SkillManager does not exist — README/examples/quickstart/10_skills.py referencing SkillManager.install(...) are stale.
• @slash_command on a Skill method marks it user-invocable via a host that reads the agent's slash_commands queue (see InteractiveAgent).
• Docstring conventions: see skills/nooa-agent-authoring/SKILL.md and AGENTS.md.

MCP servers

Core library keeps MCP optional: uv sync --extra mcp (or uv add 'nooa[mcp]').

from nooa.mcp import MCPManager

class ConfluenceAgent(Agent, llm=llm):
    confluence = MCPManager.create_from_server("my-mcp-server")   # from .mcp.json
    # or inline, no .mcp.json:
    # confluence = MCPManager.create_from_server("name", url="https://.../mcp",
    #                                            transport="streamable-http", headers={...})

    async def respond(self, prompt: str) -> str:
        """Answer using the Confluence tool."""
        ...

MCP tools appear alongside regular methods — the LLM calls them like any other attribute. Servers are configured in .mcp.json (VS Code/Claude Code format). Transports: stdio, SSE, streamable-http; OAuth supported (nooa.mcp.OAuthConfig).

${VAR} placeholders in .mcp.json and inline servers configuration remain literal. If a server needs a secret, resolve it in trusted caller code and pass the resulting URL, header, argument, or environment value directly to MCPManager.create_from_server().

Multimodal media

from nooa import Image, Audio, Video, File

class MediaAgent(Agent, llm=llm):
    async def describe(self, image: Image) -> str:
        """Describe what you see."""
        ...

img = Image.from_file("photo.png")     # also .from_bytes(media_type=...), .from_url(...)
await MediaAgent().describe(img)

Only works with multimodal-capable models — text-only models error when handed media. See examples/quickstart/12_multimodal.py.

Pitfalls

• A tool attribute assigned as a bare function/lambda isn't introspectable — use class methods or proper objects.
• Skills/tools on self are per-instance; attach them in __init__ (after super().__init__()).
• Hide internal-only tools with Annotated[T, hidden] so they don't pollute every prompt.
• External API fragility: wrap flaky calls in a small deterministic method with clear errors, so CodeAct sees a clean failure it can react to.

Related skills

• nooa-agent-authoring — visibility rules and method design that make tools discoverable.
• nooa-capturing-traces — every tool invocation becomes a tool_execution.* span you can inspect.