Skip to content

AI assistant & agents

Put an agent on the team

ConfOS ships an assistant in the dashboard and an MCP server for Claude and ChatGPT. Both work the same audited, permission-checked actions as your staff — writes run on organiser approval: confirmed per change, or auto-approved under a per-run cap when you flip the switch. MCP writes always go through the queue.

Claude

connected to ConfOS via MCP

How's registration for DevConf 2027, and which speakers still haven't confirmed?

events_registration_stats
speakers_list
412 registered, up 38 this week. Three speakers are still unconfirmed: Alex Rivera, Maya Patel and Sarah Chen.

Create a task for each one to chase their confirmation before Friday.

tasks_createproposed

[Requires organiser approval — this proposes a change]

tasks_createproposed
tasks_createproposed
app.confos.ai/events/devconf-2027/approvals

Approvals

3 pending
Approve all

tasks_create

proposed by Claude (MCP token: "ops-laptop")

Pending

tasks_create

proposed by Claude (MCP token: "ops-laptop")

Pending

tasks_create

proposed by Claude (MCP token: "ops-laptop")

Pending

13 tools today

8 read, 5 write: events_list, attendees_search, schedule_list and friends. Writes like sessions_upsert always propose — they never execute.

Named, scoped tokens

Bearer tokens are minted per client, scoped, revocable, and tracked with last-used timestamps under Settings → AI & API.

Read-only on Free

Every plan gets the read tools. Writes through the approval queue are Pro.

Human in the loop

Nothing changes without a person

Agent and MCP writes never execute directly. They enqueue. Only a human organiser can execute one, the action re-runs as the approving user, and the audit log records both actors.

No bypass flag

There is no code path around the queue. A custom ESLint rule fails the build on any escape hatch.

Runs as the approver

Approved actions re-run with the approving organiser's own permissions — an agent can never out-rank a human.

Both actors logged

Every agent write records the agent that proposed it and the human who approved it.

In-app assistant

An assistant that knows what page you're on

It streams in a drawer beside your work, knows the current event, page and selection, and is scoped to your role — it can never do more than you can.

What the assistant won’t do

  • Touch billing, finance or credentials
  • Run bulk-destructive operations
  • Do anything your role can't do in the UI
  • Blow past the auto-approve cap — 8 changes per run, then it asks again

Model calls run through OpenRouter with zero-data-retention flags. Telemetry is operational only — tokens, cost, latency — never prompts.

ConfOS Assistant

Private in this organization

DevConf 2027 · Speakers

Which speakers haven't confirmed yet?

Checked speakers list

Speakers

3 of 14 still unconfirmed

Tool result
  • ARAlex RiveraEngineer · NimbusInvited
  • MPMaya PatelFounder · LoopworkInvited
  • SCSarah ChenCTO · HexleyInvited
Open speakers

Add a task to chase each of them before Friday.

Confirm this change

Create 3 tasks

tasks
3
due
Friday
assignee
unassigned
DenyConfirm

Ask ConfOS

Ask

Writes only run after you press Confirm.

Why this works

The AI surface can never drift

Assistant and MCP tool schemas are generated from the same action registry the dashboard runs on. There is no second implementation to fall out of date — an agent sees exactly what the product can do, nothing more.

Give your agents something real to do.

Join the waitlist and we’ll email you when founding onboarding opens.