Skip to main content

nodus-mcp-server

Status: published on PyPI.

An MCP server that connects AI assistants to the Nodus language runtime — giving them persistent memory, sandboxed code execution, and checkpoint/resume orchestration workflows, all powered by .nd scripts running on the Nodus VM.

Supports Claude Desktop (stdio), and claude.ai and ChatGPT (streamable HTTP at /mcp, through a tunnel, with sign-in).

Tools

Tool What it does
nodus_remember Store a fact in persistent memory with optional tags
nodus_recall Search memory by free-text query and/or tags
nodus_forget Delete a memory entry by ID
nodus_run_goal Run a built-in Nodus goal (structured multi-step result)
nodus_run_workflow Run a built-in Nodus workflow (returns a graph_id for resuming)
nodus_resume_workflow Resume a workflow from a checkpoint using its graph_id
nodus_exec Execute arbitrary Nodus code in a sandbox (no file I/O, no network, no subprocess, no environment variables, 10 s timeout)

Requirements

  • Python ≥ 3.10
  • pipx (recommended — keeps the server in its own isolated environment)
  • Claude Desktop or ChatGPT Desktop (the downloadable apps, not browser versions)

Install

pipx install nodus-mcp-server

Claude Desktop setup

1. Find your config file

Setup Config path
Standard install %APPDATA%\Claude\claude_desktop_config.json
Windows Store app %LOCALAPPDATA%\Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude\claude_desktop_config.json
macOS ~/Library/Application Support/Claude/claude_desktop_config.json

2. Add the server

{
  "mcpServers": {
    "nodus": {
      "command": "nodus-mcp-server",
      "args": ["--stdio"]
    }
  }
}

If nodus-mcp-server isn't on your PATH, use the full path to the executable. On Windows with pipx that's typically C:\Users\<you>\.local\bin\nodus-mcp-server.exe.

3. Restart Claude Desktop

The seven nodus_* tools will appear when you click the tools icon (the + button or tool picker) in a new conversation.

How to use

Memory

Store anything you want Claude to remember across conversations:

Use nodus_remember to store: "Project deadline is 2026-07-01" with tags ["project", "deadlines"]

Retrieve it later:

Use nodus_recall to find memories tagged "deadlines"

Or search by content:

Use nodus_recall to find memories about "deadline"

Memory is stored in a local SQLite database at ~/.nodus-mcp-server/data/memory.db and persists across upgrades.

Sandboxed code execution

Run Nodus (.nd) code in a fully sandboxed runtime:

Use nodus_exec to run: print("Hello from Nodus!")

The sandbox enforces: no file I/O, no network, no subprocess, no environment variables. Use print() to surface results — top-level return values are not captured.

Goals (structured multi-step tasks)

Goals run a fixed sequence of named steps and return each step's result:

Use nodus_run_goal with name "summarize" and params {"text": "your text here"}

Built-in goals:

Goal Params What it does
summarize {text} Counts characters, classifies size (short/medium/long)
pipeline {items, label} Validates a list and produces a labelled report

Workflows (checkpoint/resume orchestration)

Workflows are like goals but support checkpoints — they can be paused and resumed from a saved state:

Use nodus_run_workflow with name "research" and params {"topic": "LLM context windows"}

The response includes a graph_id. Use it to resume the workflow later:

Use nodus_resume_workflow with graph_id "g_abc123" (and optionally a checkpoint label)

Built-in workflows:

Workflow Params What it does
research {topic} Two-step plan + execute workflow with checkpoints at each step

Adding your own goals and workflows

Goals and workflows are .nd files (Nodus source) placed in the goals/ or workflows/ directory of the installed package. The file should only define the goal or workflow — the server calls it for you.

// goals/my_goal.nd
goal my_goal {
    step process {
        if (input_text == nil) { throw "missing required param: input_text" }
        let result = len(input_text)
        return {"length": result, "has_content": result > 0i}
    }
}

Then call it:

{"name": "my_goal", "params": {"input_text": "hello"}}

For a goal, params are injected as top-level variables in the .nd execution context. Check nil before using them — missing params surface as nil, not an error, unless you throw explicitly.

A workflow declares its params instead, so they are stored with the run and restored when it resumes. A missing or unknown param is an error:

// workflows/my_workflow.nd
workflow my_workflow(input_text) {
    step process {
        checkpoint "processed"
        return {"length": len(input_text)}
    }
}

See the Nodus language guide for the full .nd syntax reference.

About Nodus

The goals, workflows, and nodus_exec sandbox all run on the Nodus VM — a lightweight, embeddable language runtime designed for AI-native orchestration. Nodus scripts (.nd files) define the step logic; the MCP server wires them to Claude over the Model Context Protocol.

This package is part of the Nodus ecosystem — a family of companion libraries built around the nodus-lang runtime (MCP and A2A servers, an SDK, memory and workflow engines, and more; see the ecosystem overview). The runtime's own suite runs this server's tests on every release, so an upstream change that would break it is caught before it ships.

Architecture

server.py          — MCP tool definitions, NodusRuntime setup, request dispatch
runner.py          — goal/workflow execution via ModuleLoader + VM
memory_store.py    — SQLite-backed thread-safe memory store
oauth.py           — sign-in for --http --public-url (OAuth provider, consent page, grants)
goals/             — .nd goal definitions (bundled + custom)
workflows/         — .nd workflow definitions (bundled + custom)
~/.nodus-mcp-server/data/memory.db  — SQLite DB (persists across upgrades)
~/.nodus-mcp-server/data/oauth.db   — OAuth clients and grants (with --public-url)
~/.nodus-mcp-server/passphrase.json — the sign-in passphrase, scrypt-hashed

claude.ai and ChatGPT setup (remote, with sign-in)

claude.ai and ChatGPT call the server from their own cloud, not from your machine, so they need a public HTTPS URL. They cannot use localhost or 127.0.0.1. Use a tunnel such as ngrok, and turn on sign-in so that the URL alone gets nobody in.

With --public-url, the server is its own OAuth authorization server. Each app registers itself, sends you to a consent page on your server, and gets a token only after you enter your passphrase and choose what it may do. Without --public-url, --http has no sign-in at all, exactly as before 0.2.0, and must stay on loopback.

1. Set a passphrase

nodus-mcp-server set-passphrase

It is prompted for, never passed as an argument, and stored as a scrypt hash in ~/.nodus-mcp-server/passphrase.json. Setting a new one signs every app out.

2. Start the server with your public URL

nodus-mcp-server --http --port 8765 --public-url https://<your-static-domain>.ngrok.io

--public-url is the https:// origin the tunnel serves, with no path. You can set NODUS_MCP_PUBLIC_URL instead. The server refuses to start with it if no passphrase is set.

3. Start the tunnel

ngrok http --url=<your-static-domain>.ngrok.io 8765

4. Connect

The server URL in both apps is https://<your-static-domain>.ngrok.io/mcp.

  • claude.ai: Settings → Connectors → Add custom connector, enter the URL, and choose dynamic client registration (DCR). This server does not support a client metadata URL (CIMD).
  • ChatGPT: Settings → Apps → Advanced settings → enable Developer Mode, then Create app with the URL and OAuth authentication.

Your browser opens the consent page. It shows the address the app will send you back to. Check that address, since apps choose their own names. Tick what to allow, enter your passphrase, and press Allow.

Created a ChatGPT app before 0.1.11? Its URL probably ends in /sse. 0.1.11 replaced the SSE transport with streamable HTTP at /mcp, and /sse now returns 404. ChatGPT's app settings have no field for changing the URL: delete the app and create it again with the /mcp URL.

What an app can be allowed to do

Scope Tools Ticked by default
memory.read nodus_recall yes
memory.write nodus_remember, nodus_forget no
workflow.run nodus_run_goal, nodus_run_workflow, nodus_resume_workflow no
exec nodus_exec no, and only offered with --allow-remote-exec

An app sees only the tools its token allows. nodus_exec runs code on your computer. It is sandboxed, but a sandbox bug there is a bug on your machine, so it is not offered over the public URL unless you start the server with --allow-remote-exec.

Managing apps

nodus-mcp-server grants list            # who is connected, with what, and when it was last used
nodus-mcp-server grants revoke <id>     # sign one out now

Access tokens last an hour. Refresh tokens last 30 days and change on every use. If an old refresh token is used again, that app is signed out.

Note: Memory is shared with the Claude Desktop instance (same SQLite database at ~/.nodus-mcp-server/data/memory.db).

Local HTTP clients

nodus-mcp-server --http with no --public-url serves http://127.0.0.1:8765/mcp with no authentication. Anything that can reach it can use every tool. Use it only for local clients, and never put a tunnel in front of it. With --public-url on, local HTTP clients sign in too, because the server cannot tell a tunnelled request from a local one. Claude Desktop uses stdio and is unaffected either way.


Upgrading

Stop-Process -Name "nodus-mcp-server" -Force   # Windows — close before reinstalling
pipx install nodus-mcp-server --force

Then restart Claude Desktop / ChatGPT Desktop.

Development

git clone https://github.com/Masterplanner25/nodus-mcp-server
cd nodus-mcp-server
python -m pip install . mcp pytest httpx
python -m pytest -q

CI runs the suite across a matrix of OS (Ubuntu, Windows), Python (3.10 and 3.13 — the ends of requires-python), and the mcp SDK (1.x and 2.x, since the server branches on the SDK major). A separate job builds the wheel and completes a real initialize handshake against the installed server. main is branch-protected: changes land through a branch, a pull request, and green CI.

License

MIT

Metadata

Release files for nodus-mcp-server 0.2.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for nodus-mcp-server 0.2.1
File Size Uploaded Provenance Actions
nodus_mcp_server-0.2.1.tar.gz 41.8 kB Without provenance Details for nodus_mcp_server-0.2.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for nodus-mcp-server 0.2.1
File Interpreter ABI Platform Actions
nodus_mcp_server-0.2.1-py3-none-any.whl Python 3 none any Details for nodus_mcp_server-0.2.1-py3-none-any.whl

Total release size: 71.9 kB

Release files / nodus_mcp_server-0.2.1.tar.gz

Download URL nodus_mcp_server-0.2.1.tar.gz
Size 41.8 kB
Tags Source
SHA-256 checksum
How to use checksums
45938b0b1b317cc680174bccd44202b39c1a8f1fc91d292f85435bb3318215c7
BLAKE2b-256 checksum
How to use checksums
8e79b8d360d57ed630f09157d7020f7b5cbd2c3d0b188eaa21e4473bfd75f18e
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.11.9

Release files / nodus_mcp_server-0.2.1-py3-none-any.whl

Download URL nodus_mcp_server-0.2.1-py3-none-any.whl
Size 30.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
c3441c5f3ba6919f159bf7fca0b1afa8367e340b34a5bc64bfad1ce84c148e46
BLAKE2b-256 checksum
How to use checksums
b3fd10da72848f53d8ceb237eae8ff8f0467d487a02b77d76756adf45563ca22
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.11.9

Security for nodus-mcp-server 0.2.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page