Early access Enterprise-grade AI governance, sized for small firms.
We're building a tamper-evident record of what your AI agents change.
For the day you have to show what an agent did, and when: we're building One Agent Harness to keep that record for your firm.
Available now: Claude reads and updates your team's Microsoft To Do and Planner tasks through one governed connection, with updates only where you switch them on.
Join the waitlistThe problem
Your staff already let AI agents change tasks, notes and messages on company data.
When a client or an auditor asks what an agent did, and when, nobody can show it.
The tools that record and control this are built and priced for large enterprises.
How it works — setup takes minutes, not a project
Sign in
Use your Microsoft work account.
Connect Microsoft 365
Choose your services, and read or write for each account.
Point Claude at ONE MCP
With one connector URL or one command.
No admin consent is needed to start: each person approves only what they connect. A tenant administrator can approve once for everyone from the setup page. Nothing is installed on each machine for claude.ai or Cowork.
What we're building
Three additions will attach to the same connection. They are on our roadmap, without a date.
Policy
We're building ready-made policy templates that will decide what your agents may read, write and send in Microsoft 365.
On our roadmap
Marks
We're building visible and machine-readable marks for the tasks, emails and documents your AI agents create.
On our roadmap
Record
We're building a tamper-evident record of what your AI agents change, for the day you have to show what an agent did, and when.
On our roadmap
What works today, and what is on the roadmap
One Agent Harness is in early access. The table below is the whole list; nothing else is promised.
| Capability | Status |
|---|---|
| Ready-made policy templates that decide what agents may do (policy as code) | On the roadmap |
| Marks on what agents write | On the roadmap |
| A tamper-evident record of agent activity | On the roadmap |
| Microsoft To Do: read tasks | Available now |
| Microsoft To Do: create, update and reopen tasks | Available now |
| Microsoft Planner: read tasks | Available now |
| Microsoft Planner: create and update tasks and notes | Available now |
| Several Microsoft accounts per person | Available now |
| Works in Claude Code, claude.ai and Cowork | Available now |
| Microsoft 365 files (OneDrive and SharePoint) | On the roadmap |
| Microsoft 365 calendar | On the roadmap |
| Microsoft 365 mail | On the roadmap |
| Gmail (Google Workspace mail) | On the roadmap |
| GitHub (issues and projects) | On the roadmap |
Security, in plain terms
No passwords People sign in with their Microsoft work account.
Encrypted credentials Microsoft credentials are stored under HSM-backed key wrapping.
Untrusted input Task text from Microsoft 365 is treated as data, never as instructions.
Writes you switch on Each write is audited with field names, never values.
Separate tokens The token Claude holds is never the one Microsoft issued.
Hosted in the EU In Azure Sweden Central.
The service's own policies: Service privacy notice · Terms · Cookies
Questions
Claude already has a Microsoft 365 connector and admin switches — why this?
Those switches turn a connector on or off inside one vendor's product. We're building something different: a tamper-evident record of what agents change, rules written as policy, and marks on what agents write, across every AI tool your staff use rather than inside one vendor. Today One Agent Harness already sends every request Claude makes to Microsoft 365 through one gate that checks what the connection may do; the record, the rules and the marks are on our roadmap.
Will a tool take care of our obligations under EU rules on AI and data?
No tool can do that for you. We're building One Agent Harness to help you keep and show the records and controls those rules ask of you: what an agent changed, when, and under which rule.
Do you have SOC 2 or ISO 27001?
No. We are in early access and hold no certification. Instead, the security section above lists the controls that are in place today, the service is hosted in the EU, and we answer security questions in writing.
Does it need Microsoft admin consent?
No. Each person signs in with their Microsoft work account and approves only what the connected services need. If you prefer, a tenant administrator can grant consent for the whole organisation once, from the setup page, so nobody sees the prompt.
Where is data stored, and for how long?
In the EU, in Azure Sweden Central. For the waitlist we keep your email, company and team size until access opens or for 365 days, whichever is sooner, and then delete them. The privacy notice below has the details.
Will you record what my staff do?
The record we're building will cover what EU rules require an organisation to account for, and no more. Whether staff's personal AI use is included will be the organisation owner's choice, and staff will be told.
What does joining the waitlist mean?
We store your email, company and team size to contact you about access to One Agent Harness. Nothing else happens: joining grants no access and signs you up for nothing else.
Which AI tools does it work with today?
Claude: Claude Code, claude.ai and Cowork. Other assistants are on our roadmap.
Join the waitlist
Join the waitlist to shape it. We'll contact you about early access; nothing else happens.