Skip to content
Platform

AI that works over your real workspace data

The assistant answers questions about live projects, tasks, and time — and can act on them. It can only do what you can do, and you decide up front whether it can write at all.

  • Scoped to your own permissions
  • Read-only unless you grant more
  • Connect Claude, ChatGPT, or Cursor
Claude·Northwind workspaceRead-only

Which projects are slipping, and why?

list_projectsCompleted

{ "status": "in_progress" }

7 projects

delivery_confidenceCompleted

{ "project_id": 42 }

score 58 · band at_risk · 5 factors

blocker_radarCompleted

{ "project_id": 42 }

3 blocked · 2 stale · 1 stuck in review

explain_delayCompleted

{ "project_id": 42 }

NWD-012 waiting on NWD-009 since Aug 8

Northwind Digital is the one to watch — confidence 58 and falling. The critical path is NWD-009: it is 35% done and NWD-012 has been sitting in review behind it for four days. Two more items are overdue and one high-priority item has no owner.

update_taskRefused

{ "id": 912, "assigned_to_id": 7 }

Refused — this connection is read-only

How it works

Automation you can put in front of a security review

The governance is in the product, not in a policy document about the product.

  1. 01

    Ask about the work

    Query real projects, tasks, and tracked time in plain language, and get a morning brief or a drafted standup.

  2. 02

    Inherit your permissions

    The assistant can only reach what your account can reach — it is not a back door around access control.

  3. 03

    Decide read or write up front

    Write access executes immediately over MCP, with no confirmation prompt — so grant read-only unless an assistant genuinely needs to change work.

  4. 04

    Connect your own host

    Point Claude, ChatGPT, Cursor, or another MCP host at the workspace and work from the tool you already use.

Assistant, reporter, and an MCP server

Three separate surfaces, each one modular and each one logged.

AI Assistant

Ask about your work, draft standups, and get a morning brief over live data.

AI Client Reporter

Weekly client status drafted from what actually happened, for you to review before it sends.

MCP server

Claude, ChatGPT, Cursor and other hosts connect directly to the workspace.

Voice input

Dictate tasks, notes, comments, and instructions to the agent — read it back before saving.

Credit tracking

AI usage draws on workspace credits, visible before and after a run.

Revocable access

Review connected AI hosts and revoke any of them from settings.

How the governance actually works

Every one of these is a product behaviour you can demonstrate, not a promise in a contract.

Permission inheritanceAI Assistant
The assistant operates strictly inside the permissions of the person asking. It cannot read a project you cannot read.
Confirmation stepAI Assistant
In the app, an action that changes data shows you what will happen before it runs. Over a connected MCP host there is no such step — a write tool executes when called.
Pending AI actionsAI Assistant
Action types can be held for a human to approve before they execute.
Supported hostsMCP Server
Claude, ChatGPT, Cursor and other MCP-compatible hosts, with per-platform setup walkthroughs.
Connection limitsMCP Server
Each connection has defined access and limits, and can be reviewed or revoked at any time.
Report reviewAI Client Reporter
Drafts are reviewed before they reach a client — the reporter proposes, a person sends.
Call and action logsActivity log
MCP calls and AI actions are recorded against the workspace with full context.

Common questions

Can the assistant see things a user should not?

No. It operates inside the permissions of the person asking. If you cannot open a project, neither can the assistant on your behalf.

Will it change data without being asked?

In the app, no — an action that changes data shows you what will happen before it runs. Over a connected MCP host it can: a write tool executes the moment it is called, with no confirmation prompt. That is why connections default to read-only, and why you should only grant write access to a host that genuinely needs it.

Which AI hosts can connect over MCP?

Claude, ChatGPT, Cursor and other MCP-compatible hosts. Each has its own setup walkthrough, and every connection can be reviewed or revoked from settings.

How is AI usage billed?

Usage draws on workspace AI credits. Consumption is visible in the workspace, so spend is something you can watch rather than discover.

Governed agents meant we could finally say yes to automation without losing the audit trail our compliance team needs.
Sam WhitfieldDirector of PMO, Alderly Partners
0
bespoke integrations to connect Claude
100%
of AI actions logged
Your own
permissions, inherited exactly
Get started

Stop finding out about risk in the retro

Score every project, catch scope creep while it is still cheap, and give clients a seat at the table — without another status meeting.

90 days
Free, full platform
No card
Nothing to enter up front
20 min
Guided walkthrough on request