Tools

A tool is a program your AI can work in: Blender, your Obsidian or Logseq notes, a browser, a 3D printer, your smart home. Under the hood each is an MCP server - the open protocol for giving an AI actions - and the app knows how to start it, what it needs, and what to ask you for.

The Tools page

Add-ons › Tools is laid out like the Offline Models page. Your Tools comes first, as compact rows that each say in one line whether the tool is ready or what it is waiting on. Available Tools follows: the tools listed in the directory, reviewed, and a card there has one button, Add. Adding downloads nothing - it puts the tool in Your Tools and opens its Set up list. The page says so when Your Own AI Build is not installed.

Add your own takes any MCP server: a program with arguments, or an address on this computer or your own network. It also asks for a few things you would say to use the tool (optional, one per line), so a chat message that sounds like one reaches it and small talk does not. Tools you add yourself get the same Set up list as the listed ones.

One Set up list per tool

Each tool card opens a single list of everything that tool needs, in order: Your Own AI Build (the free add-on that lets your AIs use tools), the programs the tool runs on, the tool itself, its settings, which AI uses it, and any check the tool declares (for Blender, its add-on). Each line shows where it stands and has its own button - nothing downloads or installs until you press that line's button - and the list ends on Ready, or on exactly what is still missing. It also says when a tool is switched off in chat for the AI that has it.

Programs are installed from the list: Install first shows exactly what would run and runs it on the second click - uv through its official installer, Git, Node.js, Python and Docker through your system's own package manager, or your terminal when a password is needed. Some tools bring their source with them: Fetch and add says what it fetches and from where (a git repository into your home folder) before it does. Check for updates on such a tool is a single check you press; nothing is checked on its own.

Blender

Blender's own Lab server needs Blender 5.1 or newer and its MCP add-on. The Set up list offers to install the add-on through Blender's own extension installer - it also turns on Blender's Allow Online Access setting, which the add-on requires before it will open its connection. That connection is local, on this computer only; the setting is Blender's, and it also lets Blender check its own extensions site for updates. The list says this before it runs. Restart Blender afterwards if it was open.

Obsidian

Your AI searches your Obsidian notes, reads them, lists folders and tags, and - when you allow it - writes and edits notes. It works on the notes folder directly, so there is no Obsidian plugin to install and Obsidian does not need to be open. It starts read only; one switch in its settings allows writing. A second switch, on by default, also reads those notes into the documents of each AI that uses the tool and keeps them in sync, so the AI remembers your notes in every chat, not only while the tool is running. Needs Node.js 20 or newer; the first session fetches the tool itself (a few MB).

Logseq

Your AI searches your Logseq graph, reads pages and blocks, follows backlinks, runs queries, and - when you allow it - creates and edits pages and blocks. It works with both kinds of graph, files and the newer database, through Logseq's own HTTP API on this computer, so Logseq has to be open with that server started; the Set up list checks for it and says how. An API token is required. It starts read only; one switch allows writing. For a graph kept as files, a second switch also reads the folder into the documents of each AI that uses the tool and keeps it in sync. Needs uv; the first session fetches the tool itself (about 45 MB).

Settings a tool asks for

Some tools need an address or a key - Home Assistant needs where it is and a token you create; a printer server needs the printer's address. A listing describes those settings; it never carries values. The Set up list asks for them in a small form; settings can be text, switches or folders (with a folder picker). Save is a real button, an amber "Not saved yet" shows while what you typed differs from what is stored, and the last line of the list says to press it. Settings that do not apply (a Logseq graph folder while "database graph" is on) are greyed out.

Values stay on this computer. Secret ones are stored encrypted with the same key that protects your conversations, are never shown back, and go only to that tool. A tool missing a required setting says so and is not started; the "this computer or your own network only" rule for a tool's address is checked on the address as it is used. Change a tool's settings and the next message uses them: the chat opens a fresh tools session instead of carrying on with tools set up the old way.

Giving it to an AI

Tools are chosen per AI - the "which AI uses it" line of the Set up list, or the Tools section when you edit the AI. An AI carries only the tools you gave it.

How an AI uses it

• In a project, the tools are attached to the session; the AI uses them like any other action - each step shows in the working rail with its real result, and every action goes through your approve step, the same as a file edit. Your records keep all of it.
• In chat, an AI that carries tools runs its turns through the same machinery behind the scenes, with the same steps and approvals in the thread. Nothing needs a folder, and the workspace behind a tools chat is not a project: it never shows as an open folder elsewhere, and its replies are not offered to a folder's notes.

Each listing carries guidance for the AI - how to use that tool well (Blender: change the open scene through the live connection, work in small steps, never save unless asked). It rides into every session the tool is part of. A tools session is also told where a tool's things live: the folder it runs in is empty scratch space, notes and scenes exist only inside the tool, and their paths are the tool's own, as listed.

An AI that carries a tool keeps its documents: a chat that runs through a tools session, and a project session, are given the passages of the AI's own documents that match the question, the same as an ordinary chat, and the Sources button under the reply lists them.

Tools only when a message needs them

An AI that carries a tool does not run every message through a tools session. A message goes to a session when it names the tool, sounds like what the tool does, or the conversation has already used a tool; anything else - "thanks" - is an ordinary answer, faster to start and, on an online model, fewer paid tokens. The check runs on this computer, with no model call, on the phrases the tool listed or its description. When it gets it wrong, the reply has Answer again with tools. Without the small helper download nothing changes: every message still goes to a session.

The chip beside the model chip

In chat, a small chip beside the model chip shows what the AI carries - an icon and a count each for tools and skills, nothing when it carries nothing. Obsidian, Logseq and Blender show their own logos, monochrome like every other glyph; a tool you added can fetch its site's icon on request. The chip opens one panel:

• Every tool and skill installed on this computer, each with an on/off switch. The ones this AI carries are on; switching another on gives it to the AI right there. Off keeps a tool on the AI - it just sits out. All tools off and the AI answers directly. A tool that is not ready opens its Set up list from here.
• Approvals for tools - Ask every time, Auto, or Approve everything, the same three levels as projects. The choice is remembered per AI and applied to the open session at once.
• The Projects install when the helper is missing: tools run on the Projects helper, a free ~50 MB download that shows its progress on the panel.

Changes apply from your next message. Settings › Appearance can hide the chip; tools and skills stay on the AI and can be changed from its form.

What a tools turn looks like

Chat with tools has its own shape, and it looks like chat. The page changes on the keystroke: your message and a "Thinking.." line go up before the app decides whether the message needs a tool and before the session opens; the line then says "Getting your tools ready.." and "Looking through documents.." as it goes. While the turn runs, each action is its own collapsed row with its glyph and name. When it is done, one grey line such as "Read 2 files, ran 1 command" sits above the reply and opens to the rows, and the AI's words are plain chat text. There is no box and no Simple or Detailed control; those belong to projects.

A turn that used no tool says where its answer came from: "Answered from its documents" when the AI's own documents rode along, or "Answered without using its tools", with Try again with the tool - one click sends the question again with the AI told to use its tools first.

Your first message waits for the tools: a chat that carries tools opens once every tool has finished starting, so the AI never answers a first message without them. A tool that never starts cannot hold the chat up: after 45 seconds the chat opens anyway and says which tool did not start.

A session that carries tools asks for reading room before it starts. The tool descriptions alone take about 10k tokens, so a local model running with a small window is reloaded at a bigger one when your computer can afford it, and told plainly when it cannot - naming a reading room pinned in Fine-tune as the cause when that is what it is. A picture a tool returns (a Blender viewport shot, say) does not move the session onto a vision model; the AI checks its work through the tool's text outputs instead.

What tools need from your model

Tool use needs a model that is good at calling tools. Online models are; offline, that means the agentic tier - Ornith, Qwen 3.8, GLM, Nemotron 3.5, the coders - on hardware that runs them well. A tools session picks quality first: the most capable model that runs on this computer, by capability, then size, then fit, and a weaker model already loaded gives way, one reload per session.

When a small model struggles - a tools turn that never got a tool call to work, or two or more steps that ended red, on an AI kept offline - the turn says so in plain words, once per session, and names the way up: online models for someone with a plan, the Online Models page for someone without, or the stronger model on this computer for an AI pinned to a smaller one. Never on a fault of the app's own.

When a tool does not start

If a tool's server fails to start, the turn says so - "blender didn't start, so its tools are not available this session" - with the reason from its log. A tool that cannot join a session for another reason - a setting not filled in, an address that is not on your computer or network - says so on the turn too, with the reason, instead of being left out without a word. The AI is never left to improvise around a missing tool.

Sharing a tool of your own: Sharing add-ons.