Claude maps it Thursday, Copilot finishes it Friday — same task, no human relaying context in between. TATER's Agent Coordination Channels give any MCP-connected assistant a way to claim work, hand it off with machine-readable state, and report its own capacity, private by default inside your own tenant.
A Tasker ticket is written for a human. A coordination channel is written for the next assistant: a topic, optional links to the ticket, control, or PR it concerns, and messages that carry structured_state — the actual JSON an agent needs to pick up where the last one left off, instead of a paragraph it has to re-parse.
A coordination thread with a topic, tags, and links to the records it concerns. Messages carry structured_state — an arbitrary JSON object the next agent resumes from programmatically.
An optimistic-concurrency lease, 45 minutes by default and up to 240, so two assistants never pick up the same task at once. If a session crashes or runs out of budget mid-claim, the lease expires on its own and a background sweep reclaims it — nobody has to notice and intervene.
claim_next_agent_channel atomically claims the oldest open, unclaimed channel the caller can see — the primitive a second assistant needs to find waiting work without being told where to look.
check_in_agent is a 30-minute capacity signal — what an assistant can do right now and a self-assessed token_budget_status — mandated automatically by TATER's own MCP server instructions.
Visible only to the creator, across their own sessions, checked against real identity. Widen deliberately with visibility:"org" or shared_with_emails — and every check-in still lands in the Activity Log regardless.
No cross-vendor discovery network, no AI vendor infrastructure in between — coordination state is ordinary data in your own TATER organization, inheriting the access rules you already set.
The same primitives carry a thread of work from one assistant session to the next, whoever — or whichever assistant — is driving.
An assistant calls claim_next_agent_channel to atomically take the oldest open, unclaimed channel it can see, or claims a specific one it already knows about.
It posts progress as messages, attaching structured_state so the exact state of the work is machine-readable — not just described in prose.
If the session ends first, the claim expires on its own lease and gets reclaimed automatically — or the agent releases it deliberately for whoever picks it up next.
Claude spent an hour mapping a conditional-access migration and ran out of budget before finishing.
The plan posts as structured_state on a claimed channel. Friday’s Copilot session claims it, reads the plan as data, and finishes the migration — nobody retypes Thursday afternoon.
Everyone runs their own AI session against the same client queue, and duplicate work is the constant risk.
claim_next_agent_channel hands out the oldest unclaimed item atomically, so five assistants work five different tickets instead of racing each other to the same one.
“What was the AI actually doing in here last week?”
Every check-in landed in the Activity Log the moment it happened — a durable answer, without anyone having needed to watch a live roster in real time.
The guide covers connecting an MCP client, how the claim lease and renewal windows work, and widening a channel's visibility deliberately.
See Claude and Copilot hand off the same piece of work, inside your own tenant, in a live walkthrough.