An atomised index of TouchDesigner, a live bridge into a running instance, and an offline .toe reader — exposed to AI agents over MCP.
- ✓Open-source license (MIT)
- ✓Actively maintained (<30d)
- ✓Clear description
- ✓Topics declared
- ✓Documented (README)
- !Install pipes a remote script into a shell (curl | sh)
claude mcp add td-atlas -- uvx td-atlas{
"mcpServers": {
"td-atlas": {
"command": "uvx",
"args": ["td-atlas"]
}
}
}MCP Servers overview
English | [Русский](README.ru.md) | [简体中文](README.zh-CN.md)
<div align="center">
<p align="center"><img src=".github/cover.png" width="720"
alt="The td-atlas mark over a field of waveform lines."></p>
<h1>td-atlas</h1>
<p>
An atomised index of TouchDesigner, a live bridge into a running instance,
and an offline reader for saved projects — exposed to AI agents over MCP.
</p>
<p>
<a href="#install">Install</a> ·
<a href="#as-an-mcp-server">Use it from an agent</a> ·
<a href="plugin/skills/touchdesigner/references/tools.md">Tools</a> ·
<a href="docs/cli.md">Command line</a> ·
<a href="docs/troubleshooting.md">Troubleshooting</a> ·
<a href="CHANGELOG.md">Changelog</a>
</p>
<p>
<a href="https://github.com/grigabyte/td-atlas/actions/workflows/ci.yml"><img src="https://github.com/grigabyte/td-atlas/actions/workflows/ci.yml/badge.svg" alt="CI: pytest and ruff on macOS and Windows"></a>
<img src="https://img.shields.io/badge/python-3.11%2B-blue" alt="Python 3.11 or newer, on the same computer as TouchDesigner">
<img src="https://img.shields.io/badge/platform-macOS-lightgrey" alt="Platform: macOS; the Windows code paths run in CI, TouchDesigner on Windows is unverified">
<a href="LICENSE"><img src="https://img.shields.io/badge/licence-MIT-blue" alt="Licence: MIT"></a>
</p>
</div>
td-atlas gives an agent three things. The exact operator and parameter names
from *this* machine. Hands inside a running instance, with one step of undo.
And a way to read a saved project without opening it.
## What it does
- **[The atom index](docs/atom-index.md)** is the name list. Two passes read
every operator and parameter out of your own copy of the application into one
SQLite file. The values in it are the values that copy reports.
- **[The bridge](docs/bridge.md)** lets an agent change your project while
TouchDesigner is running. A script builds it there out of four ordinary
nodes: a Web Server DAT, a callbacks DAT, a status panel, and an Execute DAT
that fires when you save. You can read it, see its changes in git, and
upgrade it in place.
- **[What TouchDesigner does not report](docs/health.md)**. `errors` covers
what TouchDesigner itself calls an error, and `td_health` covers the
breakage it stays quiet about.
- **[The call journal](docs/journal.md)** writes one line per bridge call to a
file on disk, and it outlives the session.
- **[Reading projects offline](docs/offline-projects.md)** inspects, searches
and compares a project with TouchDesigner closed, and the original file is
left alone.
- **[The network text, written on save](docs/network-text.md)** is what the
bridge can put beside the `.toe` on every save, so a project gets a diffable
history in git.
## Install
Three things have to be there first.
- **TouchDesigner**, already installed. The [index](docs/atom-index.md) is
built from *your* copy of the application and holds the values that copy
reports. There is nothing to download.
- **Python 3.11 or newer**, on the same computer as TouchDesigner.
- **An AI agent that speaks [MCP](https://modelcontextprotocol.io)**. The
agent is what calls these tools. td-atlas was built and measured against
[Claude Code](https://docs.claude.com/en/docs/claude-code/overview), whose
command is the `claude mcp add` line further down. Any MCP client reaches
the same tools.
[Compatibility](docs/compatibility.md) lists the platforms and what the
connector can change in your project.
On macOS, one line from a terminal. The installer is a POSIX shell script,
so Windows takes the step-by-step sequence below.
```bash
curl -fsSL https://grigabyte.github.io/td-atlas/i | sh
```
If that address does not answer, the same script comes out of the repository:
```bash
curl -fsSL https://raw.githubusercontent.com/grigabyte/td-atlas/main/install.sh | sh
```
[`install.sh`](install.sh) finds a Python, clones the repository, makes a
virtualenv beside it, installs the package, builds the index and stages the
[bridge](docs/bridge.md). Every command is printed before it runs.
It asks two questions: where to clone, and whether to build the index now.
The default answer to the first is `~/td-atlas`. With no terminal to ask, it
takes the default for both.
Two directories end up on disk: the checkout, `~/td-atlas` unless you named
another, and `~/.td-atlas`. Besides those, `uv` or `pip` fills its own package
cache, as it does for any package. No sudo. System directories and your shell
startup files are left alone. Run it again on an existing checkout and it
updates that checkout.
<details>
<summary>By hand, or on Windows, the same sequence step by step</summary>
The block below uses [`uv`](https://github.com/astral-sh/uv), a fast installer
for Python packages. Each line carries a second form in the comment beside it,
and that one runs on the `python3` you already have.
```bash
git clone https://github.com/grigabyte/td-atlas
cd td-atlas
uv venv # or: python3 -m venv .venv
uv pip install -e . # or: .venv/bin/pip install -e .
```
There is no package on PyPI. `pip install td-atlas` and `uvx td-atlas` will
find nothing. Then, from the checkout:
```bash
.venv/bin/td-atlas build # offline index, 13–16 s, no TouchDesigner process
.venv/bin/td-atlas install # stage the bridge, print the bootstrap and MCP lines
```
</details>
Commands from here on are written `td-atlas` for short. Unless the virtualenv
is activated, call it by path. In the checkout that path is
`.venv/bin/td-atlas`. From anywhere else it is the whole path,
`~/td-atlas/.venv/bin/td-atlas` for the default directory. On Windows it is
`.venv\Scripts\td-atlas`. A system Python does not see the package.
`td-atlas install` prints two things to paste. The `curl` line ran it for you,
so both are already at the end of what it printed; running it again prints them
again. First, into TouchDesigner's textport (Dialogs → Textport and DATs),
once per project:
```python
exec(open('/Users/you/.td-atlas/bootstrap.py').read())
```
The textport answers with `[td-atlas]` lines. The last one names your
TouchDesigner build and the project the bridge attached to.
```
[td-atlas] registered as /Users/you/.td-atlas/instances/9977.json
[td-atlas] bridge ready at /tdatlas on port 9977 (auth: token)
[td-atlas] TouchDesigner 2025.32460, project 'NewProject.1.toe'
```
Second, a `claude mcp add` line for your MCP client. See
[As an MCP server](#as-an-mcp-server). To also write that same entry into
`DIR/.mcp.json`, or merge it into one that is already there, pass
`--write-mcp-json DIR`.
Once the textport has answered, finish the index.
```bash
td-atlas probe
```
This pass adds the facts only a running instance knows.
Last, check the install. If you came in halfway, start here.
```bash
td-atlas doctor
```
A finished install answers with every link `ok`, and one closing line.
```
environment : ok — running from /Users/you/td-atlas/.venv …
touchdesigner : ok — build 2025.32460 at /Applications/TouchDesigner.app …
index : ok — 667 ops, 24251 params, 2060 articles …
index build : ok — index and installation agree: 2025.32460
probe : ok — runtime pass complete: 647 ops probed …
bridge : ok — connected on port 9977 to 'NewProject.1.toe' …
mcp server : ok — 'claude mcp add …' launches …
every link checked out.
```
The counts come off your own copy of TouchDesigner, so yours will differ.
Every link that is not `ok` comes with its repair, and a broken one makes the
command exit non-zero. Before the index is built you see
`index : FAIL … fix: td-atlas build`, and before the bridge is staged,
`bridge : warn … fix: td-atlas install`.
## As an MCP server
Run `td-atlas install` and paste the `claude mcp add …` line it prints. That
line names the interpreter by absolute path, so it keeps working from any
working directory, with or without an activated virtualenv. Wiring it in by
hand looks like this:
```bash
claude mcp add td-atlas -- /path/to/python -m td_atlas.cli mcp
```
From here you say what you want in plain language. Here is what the agent calls
on it:
| What you say to the agent | What it calls |
| --- | --- |
| *"Which operator displaces an image with noise? Give me the exact parameter names before you build anything."* | `td_search_operators`, then `td_operator_schema` |
| *"Build a noise into a blur into an out TOP in the project I have open, and check nothing is silently dead."* | `td_build` — one undo block — then `td_health` |
| *"It looks like nothing is happening."* | `td_health`, then `td_flags` on whatever it names |
| *"Show me what that looks like right now, and the motion over a second."* | `td_render`, and a contact sheet for the motion |
| *"What is inside `/project1` of `myproject.toe`? TouchDesigner is closed."* | `td_project_read` — the file is copied to a cache and read there |
| *"What did you change since we started?"* | `td_snapshot` before and after, then `td_project_diff` on the two — components in `~/.td-atlas`, never your own file |
| *"Undo that."* | `td_undo` — a whole `td_build` batch is one step |
41 tools in three groups. **9 index** tools work offline, **23 live** tools act
on a running instance, and **9 project-file** tools read and write
`.toe`/`.tox` from disk. Each one, with its arguments and what it is for, is in
[`plugin/skills/touchdesigner/references/tools.md`](plugin/skills/touchdesigner/references/tools.md),
and a test holds that list to the code.
`td_build` and `td_set_params` check parameter names against the index before
sending, so the usual mistakes come back as corrections:
```
- t: is a parameter group, not a settable parameter (try: tx, ty, tz)
- typ: no such parameter (try: type, ty)
- type: 'simplex5d' is not a valid menu entry
(try: simplex4d, simplex3d, simplex2d, sparse, perlin4d)
- period: -3 is below the clamped minimum 0.0
```
## Documentation
| Document | For |
| --- | --- |
| [What people ask about td-atlas
What is grigabyte/td-atlas?
+
grigabyte/td-atlas is mcp servers for the Claude AI ecosystem. An atomised index of TouchDesigner, a live bridge into a running instance, and an offline .toe reader — exposed to AI agents over MCP. It has 0 GitHub stars and its last recorded update is dated 2026-09-15.
How do I install td-atlas?
+
You can install td-atlas by cloning the repository (https://github.com/grigabyte/td-atlas) or following the README instructions on GitHub. ClaudeWave also provides quick install blocks on this page.
Is grigabyte/td-atlas safe to use?
+
Our security agent has analyzed grigabyte/td-atlas and assigned a Trust Score of 87/100 (tier: Trusted). See the full breakdown of passed checks and flags on this page.
Who maintains grigabyte/td-atlas?
+
grigabyte/td-atlas is maintained by grigabyte. The last recorded GitHub activity is dated 2026-09-15, with 0 open issues.
Are there alternatives to td-atlas?
+
Yes. On ClaudeWave you can browse similar mcp servers at /categories/mcp, sorted by popularity or recent activity.
Deploy td-atlas to your cloud
Ship this repo to production in minutes. Each platform spins up its own environment with editable env vars.
Maintain this repo? Add a badge to your README
Drop the badge into your GitHub README to show it's tracked on ClaudeWave. Each badge links back to this page and reflects the live Trust Score.
[](https://claudewave.com/repo/grigabyte-td-atlas)<a href="https://claudewave.com/repo/grigabyte-td-atlas"><img src="https://claudewave.com/api/badge/grigabyte-td-atlas" alt="Featured on ClaudeWave: grigabyte/td-atlas" width="320" height="64" /></a>More MCP Servers
Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.
User-friendly AI Interface (Supports Ollama, OpenAI API, ...)
An open-source AI agent that brings the power of Gemini directly into your terminal.
Real-time global intelligence dashboard. AI-powered news aggregation, geopolitical monitoring, and infrastructure tracking in a unified situational awareness interface
🕷️ An adaptive Web Scraping framework that handles everything from a single request to a full-scale crawl! Don't be shy, join here: https://discord.gg/EMgGbDceNQ
The fastest path to AI-powered full stack observability, even for lean teams.