Woodentable
Menu

A workbench for AI agents.

Just add tools, point to local files, and your agent does the heavy lifting.

Download Woodentable

Version 0.7.3 · Linux · macOS · Windows

Set up your workbench.

  1. Download Woodentable

    One program holds the whole workbench, so there's nothing to install.

    Linux macOS Windows

  2. Gather your tools, files and folders

    Connect the tools you want and share only the folders you choose, read-only where it matters.

    woodentable mcp add comfy -oauth https://cloud.comfy.org/mcp
woodentable resources add gallery -mount ~/gallery -readonly ~/reference-images

  3. Invite your agent to the table

    Once Woodentable is added to your agent, everything you've gathered is within reach. Just ask.

    claude mcp add gallery -- woodentable -resources gallery serve -name gallery

Help along the way.

To keep the cogs whirring, Woodentable comes with documentation written for two readers: you at the command line, and your agent at the workbench.

User guides

Every part of the workbench has a guide: adding tools, sharing folders, connecting your agent, and the rest. Each guide arrives typeset to fit your terminal, and opens in a pager when it runs long.

# every topic, then one in full
woodentable help topics
woodentable help mounts

# the same, printed without the pager
woodentable -nopage help mounts
A terminal showing woodentable help mounts: a guide to the form of a mount, with a table of mount options and what the catalog receives.

Agent guides

Your agent has guides of its own: working at the bench, the language it writes in, and every tool you connect. It finds what it needs by searching for the task at hand.

# search the guides for a task
file_search -under /usr/share/help/agent "show an image"

# search the Tcl manual
file_search -under /usr/share/help/tcltk -limit 3 "sort a list"

# find a connected tool by what it does
tools_find "workflow templates"
Search results from the Tcl manual for sort a list: the lsort page first, then passages from the pid and eval pages.

Everything in its place.

Keep your workbench organized with a catalog for each project, so your agents see only what matters for the task at hand. Watch them work, and fine-tune the mechanism.

How the workbench contains your agent's work and records it Your disk mounts folders into the workbench's own file system, one of them read-only. Your agent, or a local model running inside Woodentable, sends scripts to a sandbox that sees only that file system. Commands are bound into the sandbox by name: built-in helpers, your own tools, which run in your interpreter with access to the whole disk, and the tools of connected servers, which run on those servers. Its work can be traced over OTLP to a backend such as Jaeger, Langfuse or Datadog, which stays off until you set an endpoint. YOUR DISK ~/gallery ~/reference-images ~/.woodentable/*.tcl your tools, read at start WOODENTABLE Catalog file system /gallery /reference-images (read-only) /usr/share/help /var/tmp (scratch) the only files the sandbox can see Local model llama.cpp, in-process an agent of its own Scripts in a sandbox safe interpreter · no exec, no sockets · a budget per call the agent's script image_mask /reference-images/fox.jpg \ -output /gallery/fox.png comfy::search_templates -q "style transfer" COMMANDS, BOUND IN BY NAME file_read file_search tag_set built-in image_mask image_resize your tools comfy::search_templates comfy::submit_workflow playwright::browser_navigate playwright::browser_click connected servers' tools Your interpreter your tools run here, with your permissions: exec, sockets, the real disk Your agent any MCP client Connected servers comfy playwright each tool runs on its server Your traces Jaeger, Langfuse, Datadog… off until you set an endpoint mount tcl_workspace one tool whole disk OTLP

Scripts in a sandbox

Agents write scripts that run in their own safe interpreter, sealed off from your system and the network. Several can share one catalog and work side by side.

Catalog file system

Work from the comfort of your cluttered computer. Mix and match files from anywhere, and agents see only those.

Agents in view

Each script your agents run can be traced, along with the server calls it makes and every turn of a local model. You can see what an agent did, how long each step took, and where a run stopped.

Bring your own agent.

Woodentable works with any agent that speaks MCP, or local models powered by llama.cpp. It runs on Linux, macOS and Windows, and it travels well: strap on a balloon and float into the cloud.

Download Woodentable

Version 0.7.3 · Linux · macOS · Windows