Reos runs a remote MCP server, so an AI assistant can work inside your research workspace: read observations and transcripts, search across projects, start an analysis flow, and write results back. Claude, ChatGPT, and any other MCP-capable client connect the same way.
The connection is scoped to your own account. An assistant sees exactly what you see, with the same role and the same project access - never more.
Connect
The server URL is:
https://www.reos.ai/api/mcp
In Claude (web, desktop and mobile): open Settings → Connectors. If Reos is listed in the connector directory, choose it and click Connect. Otherwise click Add custom connector, paste the URL, and click Connect. Either way Claude opens a Reos sign-in page. Claude registers itself automatically - there is no client ID or secret to copy. On a Team or Enterprise plan, an owner may need to add the connector for the organisation first.
In Claude Code: run claude mcp add --transport http reos https://www.reos.ai/api/mcp, then type /mcp in a session and choose Authenticate. Your browser opens the same Reos sign-in page.
In ChatGPT: if Reos is listed in ChatGPT's apps and plugins directory, add it from there. Otherwise open Settings → Connectors → Advanced → Developer mode and add a connector with the same URL. Developer mode is available on ChatGPT Pro, Team, Enterprise and Edu plans.
In another client: point it at the same URL. The server speaks Streamable HTTP and authenticates with OAuth 2.1, including dynamic client registration, so any spec-compliant client can complete the handshake on its own.
Approving the connection
After you sign in, Reos shows a consent screen before anything is shared. It names the app that is asking and the site it will send you back to (for example claude.ai or chatgpt.com), lists what the app will be able to do, and asks you to tick the organisations it may use. Organisations you leave unticked stay out of reach of that connection. Nothing is shared until you click Allow. If the screen says the app returns to localhost, that is expected for a tool running on your own computer, such as Claude Code - only allow it if you just started that connection yourself.
Try it
Once connected, ask the assistant something and it will pick the tools it needs. For example:
- "What are the top insights from my onboarding research?"
- "Find what participants said about pricing, with quotes."
- "Summarise the transcript of my latest interview and save it as a document in Reos."
- "Which projects have new raw data that hasn't been analysed yet? Run the observations step on them."
Deep research and cited answers
ChatGPT can also treat Reos as a research source, so a deep research run searches your workspace alongside everything else it consults and cites what it used.
Two tools make that work. search takes a plain question and returns matching items, each with a link. fetch takes one of those items and returns it in full. Because every result carries a link back to the item in Reos, an answer that leans on your research can point at the exact observation or insight behind each claim - and anyone reading the answer can open it, provided they have access to that project.
Both tools are read-only and act on the connection's default organisation, the first one you ticked in the list. To search a different organisation, connect it separately.
What the assistant can do
The tools follow the DECODE funnel - raw data, then observations, then insights, then solutions - and split cleanly into reading and writing.
Reading - search your research and read any result in full; list projects, flows, entities, documents and tags; read observations (including everything a single participant said); pull a transcript or a still frame from a recording; read the timecoded comments and markers on a resource; check the status and results of a flow run; and list the organisations the connection can reach.
Writing - create a project, add raw data by pasting text or importing a public URL, start a flow run, create a document, and edit or delete entities.
Every write tool is annotated so your assistant asks before it runs. Read-only tools can run without a prompt each time, which is what makes a research conversation feel fluid rather than click-heavy.
Permissions
Your organisation role decides which tools exist at all:
| Role | What the assistant gets |
|---|---|
| Owner, admin, member | Every tool, reading and writing |
| Viewer | Read-only tools |
| Stakeholder | No tools |
On top of that, every call re-checks project access, so an assistant can only reach projects you can already open. Writing to a project needs write access to that project specifically.
Working across organisations
If you tick more than one organisation on the consent screen, you do not need a second connection. Ask the assistant to list your organisations, then name the one you want - each tool accepts an organisation id per call, and your role is evaluated separately in each one. The assistant can only reach the organisations you ticked; to add or remove one, connect again and change the selection.
The two deep-research tools, search and fetch, are the exception: their shape is fixed by ChatGPT and takes no organisation argument, so they always act on the connection's default organisation.
Managing and revoking access
Connected assistants are listed under AI assistants in your profile menu, where you can revoke any of them. Revoking takes effect immediately: the assistant's access token stops working and it cannot obtain a new one.
Access tokens last one hour and are refreshed silently in the background. A connection you never use expires on its own after 90 days.
Privacy and data handling
The server reads and writes only your Reos workspace data, through Reos's own API, under your identity and role. It does not read your assistant's memory, chat history, or files, and it does not collect your conversations: it only receives what the assistant sends to a tool, such as a search query.
What a tool returns goes to the assistant's provider (Anthropic for Claude, OpenAI for ChatGPT) and is handled under their terms and your settings with them. That can include research content and the participant data inside it - names, quotes, transcripts - so check that your consent forms and policies allow sharing research with that provider before you connect an organisation that holds participant data. Section 7.5c of the Reos privacy policy covers exactly what is shared, what we keep and for how long.
Troubleshooting
- "Couldn't connect" or a sign-in loop: remove the connector in your assistant, then add it again. If you belong to no organisation yet, create or join one first - there is nothing to connect until you do.
- The assistant says it needs to reconnect: a connection that has not been used for 90 days expires, and revoking it under AI assistants ends it immediately. Connect again to get a fresh one.
- "Too many Reos requests": one account can make up to 300 tool calls a minute. Wait for the time the message names, then carry on.
- A tool is missing: your role decides which tools exist (see Permissions). Viewers only get reading tools.
Getting help
Something not working, or a tool behaving oddly? Email support@reos.ai and we will take a look.
