Companion
Companion keeps an editable memory of your projects, drafts from the documents you choose, shows where every line came from, and plans then carries out multi-step work.
Plan and build
The first call to the model has no tools. The numbered plan is saved and shown before any step runs.
Each step is shown live: memory searched, file read, table analysed, report written.
A board beside the conversation, reports to open and the actions waiting for your approval.
A run is limited to 10 turns and 30 tool calls, and the Stop button cancels it at once.
Boards
Headline figures, lines, bars, waterfalls and tables from a CSV or Excel file. The model chooses what to show, Companion does the computing.
Choose a figure or a bar to see the rows behind it, with their total.
Compare up to three what-ifs with the base. Your files are never changed.
A projection of 1 to 12 periods, with a 95% range and the error measured on recent periods. It is statistics, not a trained model.
Join your tables on a shared column and define your calculations once: margin, growth, running totals, shares.
Recomputed each time they open, laid out in sections, exportable as a PDF, an image or CSV.
Memory
Agents
Connections
Safe by construction
Screenshots of the application, taken with a fictional example.
Keys are never written to a file, a log or an export.
There is no shell, file-system access or browser control for a model to call.
Your provider receives only the selected context and its own key. It sees no other file or credential.
Runs, approvals, agent changes and MCP requests are recorded, chained and exportable.
Encrypted backup, staged restore and a way back to the previous workspace.
Setup and the workspace exist in 24 languages, 22 of them as machine translations awaiting review.
Where Companion stands
Everything described on this page is developed and tested locally. No build is signed or certified for macOS, Windows or Linux yet, and behaviour with real models is still being qualified.
Companion is becoming a product companies can deploy and configure. Here is what does not exist yet:
No. Companion connects the provider you choose: any OpenAI-compatible endpoint, Anthropic or a local model. You choose it and pay it directly.
No. There is no public installer. We deploy it case by case, as part of a project.
Memory and documents stay on your computer. Only the context you select is sent to the AI provider you chose.
Yes, through the MCP servers you connect, and only after you approve each action.
Yes. Companion exposes its own MCP server: Claude, Cursor or VS Code can read a team’s boards, metrics and what-ifs.
OpenOS is the server-side foundation, for the work of several agents across an organisation. Companion is one person’s application, on their computer. Both follow the same rule: the model proposes, the human decides.
The fastest way to understand Open Intelligence is to see it answer on your own documents. Describe your need: we’ll come back to you with a tailored solution and a costed quote.
