NemoClawd

🌐 Website • 📖 Wiki • 🏷️ Releases • 💬 Discussions • 🐛 Issues

What if the original ClawdBot and NVIDIA NeMo got together and had a baby?

You would get NemoClawd: a next-generation AI identity built at the intersection of rugged autonomy, adaptive intelligence, and real-world execution.

NemoClawd is an integrated workspace that combines the clawdbot Node.js/TypeScript runtime with the NeMo-Agent-Toolkit-develop Python toolkit.

The project uses a bridge-first architecture: Node orchestration calls the Python layer through NemoBridge, and Python executes NeMo workflows through the nat CLI with structured JSON results returned to Node.

In this repository, NemoClawd also includes canvas host branding updates, outbound transcript mirroring, and system-of-record anchoring so runtime behavior and identity settings stay consistent across builds and launches.

The goal is practical: keep Clawdbot-compatible operations while adding NeMo-backed workflow execution in one buildable source tree so that it can operate in enterprise environments.

NemoClawd currently focuses as a "Digital Project Manager".

Recent upgrades

• Canvas branding now uses NemoClawd assets in apps/clawdbot-main/src/canvas-host/a2ui/:
  • nemoclawd-favicon.ico
  • nemoclawd-apple-touch-icon.png
  • nemoclawd-logo.png
• Canvas A2UI page title is now NemoClawd Canvas in apps/clawdbot-main/src/canvas-host/a2ui/index.html.
• Canvas host fallback/default page title is now NemoClawd Canvas in apps/clawdbot-main/src/canvas-host/server.ts.
• Outbound assistant transcript mirroring is wired through deliverOutboundPayloads in apps/clawdbot-main/src/infra/outbound/deliver.ts.
• Mirrored transcript text normalization (including media filename extraction) is implemented in apps/clawdbot-main/src/config/sessions/transcript.ts.

What is included

• A root package.json that orchestrates the TypeScript build, the upstream Clawdbot build, Python environment bootstrap, NeMo source package builds, and artifact collection.
• A root pyproject.toml for a lightweight Python bridge package named nemoclawd-bridge.
• A TypeScript bridge in src/nemo-bridge.ts that lets Node code invoke Python NeMo workflows through a stable subprocess contract.
• A Python bridge in python_src/nemoclawd_bridge/__main__.py that exposes health and run commands.
• A unified Docker image in Dockerfile and Compose setup in docker-compose.yml.
• Build outputs in root dist/, local apps/clawdbot-main/dist/, and root artifacts/.

Workspace layout

This project contains both codebases directly under apps/ as normal directories:

• apps/clawdbot-main
• apps/NeMo-Agent-Toolkit-develop

NemoClawd builds and runs from this single source tree without requiring sibling repositories.

Architecture

Build orchestration

1. Root TypeScript bridge compiles to dist/.
2. Root Node build script runs pnpm build inside apps/clawdbot-main.
3. Root Python bootstrap script prepares apps/NeMo-Agent-Toolkit-develop/.venv, installs packaging tools, and installs the published nvidia-nat runtime.
4. Root Python build script builds:
   • the local nemoclawd-bridge package into artifacts/python-bridge/
   • the local NeMo source checkout into artifacts/python-nemo-source/
5. Artifact collection indexes Python package outputs in artifacts/python-nemo-source/ and writes artifacts/manifest.json.

Runtime communication

• Node uses NemoBridge from dist/index.js.
• NemoBridge launches python -m nemoclawd_bridge ... inside the NeMo virtual environment.
• The Python bridge performs one of two actions:
  • health: reports whether the nat CLI is available.
  • run: invokes nat run --config_file ... --input ... and returns structured JSON to Node.

This subprocess model keeps Python runtime dependencies isolated from the Node runtime while preserving Node-side workflow orchestration.

Execution guard policy

Each chat.send request is evaluated before execution:

• Job type detection: classifies the request as text-based or purpose-driven.
• Circadian phase detection: maps local time to startup (06-09), focus (09-18), winddown (18-22), or sleep (22-06).
• Execution policy:
  • blocks non-purpose text work during sleep hours,
  • allows overnight purpose-driven work only when urgency indicators are present (for example: urgent, critical, incident, outage, hotfix).

When blocked, the gateway returns a structured chat.send error payload with reason, phase, and detected job type.

System of record

Identity and core runtime defaults are defined in system-of-record.json.

• Canonical assistant name: NemoClawd
• Canonical session key: nemoclawd-main
• Canonical gateway port: 19001

To keep NemoClawd embedded in the system of record, run:

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
pnpm run sor:verify

Verification output is written to artifacts/system-of-record.status.json.

Verified status

The following were verified during setup in this workspace:

• clawdbot-main: pnpm install and pnpm build
• NeMo-Agent-Toolkit-develop: local .venv bootstrap, nat --help, and python -m build --wheel --sdist

The NeMo source build required a fallback SCM version because this checkout does not currently expose git metadata to setuptools-scm. The build scripts handle that by setting:

• SETUPTOOLS_SCM_PRETEND_VERSION=0.0.0
• SETUPTOOLS_SCM_PRETEND_VERSION_FOR_NVIDIA_NAT=0.0.0

Setup

Prerequisites

• Node.js 22.12+
• pnpm 10+
• Python 3.11-3.13
• Optional: uv
• Optional: Docker Desktop

Install root dependencies

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
pnpm install

Build everything

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
pnpm run build:all

Smoke test the bridge

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
pnpm run smoke:test

This checks the Node -> Python bridge path by asking the Python helper to report whether the nat CLI is available.

Use the bridge from Node.js

Example usage after pnpm run build:

import { NemoBridge } from "./dist/index.js";

const bridge = new NemoBridge();
const health = await bridge.health();
console.log(health.payload);

To run a workflow from Node:

import { NemoBridge } from "./dist/index.js";

const bridge = new NemoBridge();
const result = await bridge.runWorkflow({
  configFile: "workflows/nemo-sample.workflow.yml",
  input: "List five subspecies of aardvarks",
});

console.log(result.payload.stdout);
console.error(result.payload.stderr);

The sample workflow in workflows/nemo-sample.workflow.yml follows the upstream NeMo hello-world example and requires NVIDIA_API_KEY.

Start Clawdbot with NeMo environment variables prepared

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
pnpm run start:clawdbot -- --help

This launches the built Clawdbot entrypoint and injects default NEMOCLAWD_* environment variables pointing at the NeMo virtual environment.

Docker

The integrated container includes both Node and Python runtimes.

Build the container

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
docker compose build

Start the integrated runtime

Set-Location "E:\AIEnv_Adam\AIEnv\Storage_of_Programs\NemoClawd\NemoClawd-integrated"
docker compose up nemoclawd

Artifacts

After pnpm run build:all, expected outputs are:

• dist/ - compiled TypeScript bridge
• apps/clawdbot-main/dist/ - compiled Clawdbot output
• artifacts/python-bridge/ - nemoclawd-bridge wheel and sdist builds
• artifacts/python-nemo-source/ - NeMo source wheel and sdist builds
• artifacts/manifest.json - artifact summary

Notes

• The integration layer keeps upstream source trees local in apps/ and adds bridge/orchestration code in the root project.
• The bridge uses subprocess JSON exchange instead of embedding Python directly into Node, which keeps failure modes easier to debug.
• If the NeMo checkout later includes full git metadata, the fallback SCM version environment variables can be removed.

Repository owner

• Owner: Darrell Mesa
• Contact: darrell.mesa@pm-ss.org