extending neon

60s

Add a tool, a sensor, a persona or an MCP client: no refactors needed.

a composed tool

tools/g1_mytool.py:

from strands import tool
from ._g1_common import ensure_dds, get_loco_client, read_fsm_id, decode_code, HANDSHAKE_FSMS

@tool
def g1_mytool(param: int = 0, network_interface: str = "eth0") -> dict:
    """One-line intent the LLM reads."""
    if err := ensure_dds(network_interface):
        return {"status": "error", "content": [{"text": err}]}
    if read_fsm_id() not in HANDSHAKE_FSMS:
        return {"status": "error", "content": [{"text": "wrong FSM"}]}
    rc = get_loco_client().MyMethod(param)
    return {"status": "success" if rc == 0 else "error",
            "content": [{"text": f"MyMethod rc={decode_code(rc)}"}]}

Then in tools/__init__.py: from .g1_mytool import g1_mytool, append it to the right bundle (the catalog counts follow at the next docs build) and run python3 -m pytest tests/ (ToolResult shape, import sanity). Every tool handed to an agent is wrapped by tools/tool_log.py, so your calls appear in the activity log without any code of yours.

a universal SDK call

You don't wrap it, it's already reachable:

use_unitree(service_name="loco", operation_name="MyNewRpc", parameters={"foo": 1})

Only make a composed tool if it needs FSM gating, mutex, or rich parsing.

a sensor (no DDS)

Same shape without ensure_dds; return {"text": ...} and {"json": data} content blocks and register it in G1_SENSING_TOOLS.

a persona

g1.py is the one place: a prompt builder next to _telegram_prompt, a tool list from build_tools or build_voice_tools, and a NEON_PERSONA name so its calls are attributed in agent_log. telegram_listener.py is the smallest example (one agent per incoming message).

an MCP client

uvx --from neon-the-g1 neon-mcp --safe          # stdio for Claude Code / Desktop, no walking
neon-mcp --http --port 8022                      # HTTP, several clients
neon-mcp --no-robot                              # cross-persona stack only, no DDS

neon/mcp.py serves the toolset over the Model Context Protocol. Without --safe a remote client can walk the robot; read safety first.

a doc page

1. docs/section/page.md (open with a <span class="read-badge">…</span>)
2. add to nav: in mkdocs.yml; numbers come from {{facts:...}} tokens (docs/hooks/facts.py), never typed
3. pip install -r requirements-docs.txt && mkdocs build --strict (CI runs the same on push to main and deploys to Pages)

Match the tone: minimalist, dense, specific. No fluff.

contribute

Fork → branch → python3 -m pytest tests/ green → Conventional Commits (feat:/fix:/docs:) → PR. Non-trivial? Open an issue first.

---

architecture composed tools

Edit page