About
Woodentable gives AI agents a place to work: the files you choose, the tools you allow, and nothing else. It's made by Bandoti Ltd., in Nova Scotia, Canada.
Why it exists
An agent is most useful when it can act: read your files, run code, call other services. That is also exactly when handing it your whole machine is a risk. The usual answer is a container for every agent, which is heavy to set up and easy to get wrong.
Woodentable takes another route. It runs an agent's work inside a sandbox built into the language it is written in, so isolation is where you start, not something you have to arrange.
What makes it different
- Isolation without containers. Every script an agent runs goes into a Tcl safe interpreter that sees only what you've mounted. Real paths never reach it, and nothing crosses the network unless a tool you added allows it.
- One tool instead of a hundred. Your agent reaches every tool and every server you've added from short scripts in a single tool,
tcl_workspace. It spends its attention on the work, not on a long list of tool descriptions. - Tools your agent writes. With
-dev, your agent writes a tool, loads it and uses it on its next turn. You read it and decide whether to keep it. - Whatever drives it. The agent you already use, over MCP; a local model running inside Woodentable itself; or an agent in its own container, with Woodentable beside it. The same files, the same tools, the same isolation.
How it's built
Woodentable is written mostly in Tcl. Tcl has had safe interpreters, sandboxes for code you don't trust, since the 1990s, and Woodentable builds on them rather than reinventing them. Local models run through llama.cpp. Each system gets one program, no complex installer necessary.
It stands on the work of many open-source projects. woodentable licenses lists each one, with its licence.
What you can count on
- Your agent sees what you mount, and nothing else, unless you add a tool or server that reaches further.
- The tools and scripts you make with Woodentable are yours to use and share.
- The program sends us nothing about your work: no files, prompts or model output. The Privacy Policy says what it does send.
- You don't need an account to use it.
Get in touch
For help, see Support.