What is osFoundry? A guided tour of the desktop AI workspace and its main areas

It is Monday, 9:40, and you have four browser tabs open: one chatbot for questions, a second one because you don't fully trust the first, an editor with a coding agent, and a notes app holding last week's research. Each one knows a piece of your work and none of them knows the rest. By lunch you have copied the same paragraph of context into three different windows.

One app instead of five tabs

osFoundry is a desktop AI workspace: one app where you chat with AI models, hand work to agents and agent teams, keep notes and searchable knowledge, run open models on your own computer, and build small apps your team can use. The models can come from three places: your own machine, cloud providers through your own API keys, or rented cloud GPUs. The app is organised into categories, and each category holds several screens. Chat is where you talk to models (Maestro, Multi chat, Branchspace and others). Workstation is where agents do longer work (Code, Agent, Agent team, Loop, Terminal). Knowledge holds your writing and what the AI remembers. Local Model is for finding, installing and testing open models. Custom App is for internal tools, and Discover is for web research and news. If you have used VS Code, the closest comparison is an editor with an activity bar where every icon opens a different tool, except here the tools are AI surfaces and they share the same projects, memory and models.

What a typical screen looks like

The Chat category is where most people start, so it is a good place to learn how osFoundry screens are put together. This is Multi chat in Council mode, with example data: three models were asked the same question and then scored each other's answers.

• The tab strip at the top (Maestro, Model chat, Multi chat, Research, Wiki, Code) switches between the surfaces of the Chat area: the full Maestro agent, a plain question-and-answer chat with cited sources, several models at once, deep research, code documentation and coding agents.
• The Compare / Council switch picks how the models work together. Compare puts the answers side by side for you to judge. Council has them answer separately, rank each other blind, and then writes a synthesis.
• Each column is one model with a provider tag: here Anthropic, Ollama (running locally) and OpenAI. You can mix cloud and local models on the same screen.
• The right rail shows the run's stages (Answers, Ranking, Synthesis) and the Peer review grid, so you can see who scored what instead of only reading a final answer.
• The Run button says how many model calls the run will make before you press it ("Run · 7 calls" here). Turning peer ranking off lowers that number.

Many osFoundry screens follow this pattern: a list on the left, the work in the middle, and an inspector on the right that shows how a result was produced.

A solo developer's first week

Ken works alone on a small web shop. He wants an AI that knows his codebase, and he would like to keep some of the work on his laptop.

On Monday he opens Local Model, then Explore. The screen reads his machine first (16 GB of memory in this example) and puts a badge on every open model: Fits, Tight or Too big. He installs a model with a green badge and asks it a few questions in Playground, which runs the same prompt on two models side by side and shows how fast each one started answering.

On Tuesday he has to change a price formatting function and is not sure what depends on it. He opens the project in the IDE, right-clicks the function and chooses Show Code Graph.

The panel lists every caller, marks which links were read straight from the code and which are only inferred, and sums up Impact of change: how many symbols, files and tests could be affected. He clicks a caller to jump to that line, and uses Send to chat to bring the graph into a conversation with Maestro.

By Thursday he is handing small tasks to coding agents. In the ADE screen each task gets its own worktree, and the Changes tab shows the diff. He clicks the gutter next to a line he disagrees with, types a note, and sends all unsent notes to the agent in one go. He ends the week with a local model he trusts for quick questions and a habit of checking impact before refactoring.

A small team lead's first week

Aiko leads four people and a release every two weeks. Her problem is not writing code; it is knowing what everyone, human or agent, is doing.

She starts in Message, the team chat. Channels and DMs work the way they do in Slack, with one difference: AI agents are channel members. Someone @mentions an agent in #release, and its reply and run land back in the channel where everyone can see them.

Next she sets up an Agent team in Workstation. A leader agent coordinates specialist members, and the screen keeps Team mail, Handoffs, Sub-agents and Proposals in separate tabs, so she can audit who passed what to whom. For the checks that happen every release, she turns that team into a Loop: a trigger (for example, weekday mornings), limits on rounds and retries, and success criteria. The Loop runs in rounds until the criteria pass and keeps a ledger of which ones passed in each round.

When several agents work on the same repository, she uses Code. Each agent gets its own branch and worktree, and nothing merges by itself. She selects two finished runs, presses Preview combined merge, and sees which files collide before anything touches the main branch.

For line-level review she uses the same notes as Ken: a note pinned to the code, sent to the agent with the file, line and code attached. At the end of the week she has one channel where agent work shows up, a release check that runs itself, and a review step that happens before merges, not after.

A researcher and writer's first week

Lena writes long reports on workplace policy. She needs sources she can check, and she has been burned by confident AI answers before.

For questions where one model's answer is not enough, she uses Multi chat in Council mode, the screen from the tour above. When two reviewers score an answer low, she knows to dig further.

For bigger questions she opens Agents, presses New session and picks the ready-made Deep Research Team. The leader splits the question into parts and gives each to a member, such as a Data Researcher and a Search Specialist.

A separate reviewer checks each step before it reaches her, and a step without a source is sent back and run again. The final answer comes as a table with a source on every row, plus a short note on where sources disagree and what was not covered (example data in the screenshot).

The writing itself happens in Knowledge. In Paper she highlights a sentence in her draft and attaches a thinking space to it, with tabs such as Academic search, Browser or a Maestro session that discusses only that passage. Supporting files and comments live in a Note. In Discover she keeps saved searches in Browser and follows keyword channels in News, so the reading she does for one report is still there for the next.

How the areas hand work to each other

The categories are not separate products; they pass work along. Local Model is the layer underneath: a model you install from Explore appears in the model pickers of Maestro, Multi chat and agent sessions.

Knowledge feeds Chat: Memory shows what Maestro knows about your work, and Indexed Folders lets you choose which folders agents may search. Agent sessions, Agent team runs and Loops report into Message channels. Workstation (the screen with that name) lets you open a channel, a note, a file and an agent session side by side in one layout. When something you built with agents should become a tool for the team, it moves to Custom App, where a Room holds a team's apps next to its own Maestro chat. Keys and providers for all of this sit in Configuration. The platform overview lists every screen if you want the full map.

Questions people ask in their first week

Do I need to use all of this?
No. Most people start with one area, usually Maestro in Chat or Explore in Local Model, and add others when a task needs them. The categories share projects and memory, so nothing you do early has to be redone later.

Does my data leave my computer?
Not when you use local models; they run on your device. Cloud models you connect through your own API keys see what you send them, and that usage is billed by the provider. The local-first page explains how the two are separated.

Is it only for developers?
No. Code, IDE and Terminal are developer tools, but Message, Paper, Note, Research, Multi chat and Discover are built for writers, analysts and team leads as well.

What does it cost?
The desktop app is free to download, and running local models has no osFoundry usage charge. You can get it from the download page.