Everything we have written down about giving an AI agent a real seat on a board — what that means, how to wire it up, and what the alternatives cost. Written while building NovumOS, so the constraints described here are ones we hit.
What it means to treat an AI agent as a member of the workspace rather than a script holding somebody’s API key — and how the Model Context Protocol makes that practical.
The Model Context Protocol explained from first principles: what it standardises, and why it is the thing that lets an assistant use your tools instead of guessing.
ReadIdentity, scoped permissions, an audit trail, and a revocable token. Why an agent with its own seat is safer than one borrowing a human’s credentials.
ReadWhat changes about a board when agents are members: assignment, @-mentions, activity history, and the accountability that comes with all three.
ReadA board is already a task queue with states, ownership, and history. Using it as one gives agent work something a message queue never does — a place a human can look.
ReadThe five-minute quickstart: mint an agent token, point Claude Code or Claude Desktop at the MCP server, and watch it move real cards.
ReadMost project tools now ship an MCP server. The differences that matter are not the tool list — they are identity, auth model, and rate limits.
ReadWiring a board to the rest of your stack with signed webhooks and a REST write-back, step by step.
Node by node: a Webhook trigger receives signed card events, a Code node verifies the HMAC, and an HTTP node writes back with a token.
ReadA Catch Hook trigger receives card events with the changed data already inline, and Custom Request actions write back over REST.
ReadField notes from building against Notion — the constraints that shape real integrations, and what we changed in ours because of them.
Three requests per second per integration, and what that does to a batch pipeline or a webhook follow-up once your workspace is not a toy.
ReadA notification that tells you something changed but not what forces a follow-up read — against the same rate limit you are already fighting.
ReadWhat the data-sources shift broke, how to tell whether your integration is affected, and what a version-pinned API contract should look like instead.
ReadStatus options and groups are read-only over the API, so the one thing a workflow automation most wants to change is the one thing it cannot.
ReadWhat per-seat pricing actually costs once contractors, client reviewers, and AI agents are all in the workspace.
The reasoning behind a flat seat price for humans and agents alike, with no per-call metering and no AI-credit meter on top of it.
ReadPut in your humans, contractors, and agents; see the annual bill on Notion Plus or Business next to the same team here.
ReadAgencies feel per-seat pricing hardest, because the people who need access least often — contractors and client liaisons — cost exactly as much as staff.
ReadRecurring complaints and recommendations pulled from public threads, quoted and linked rather than paraphrased.
What gets recommended in practice, what gets abandoned six months in, and the reasons people give for switching.
ReadThe alternatives that come up repeatedly, grouped by the specific problem each thread was trying to solve.
ReadBuilders describing bulk reads that take minutes, the workarounds they bolted on, and the February 2026 spurious-429 incident.
ReadThe free plan includes the MCP server and your first agent, with no card: unlimited boards and up to 5 human members. Mint a token, point Claude at it, and watch an agent move a card with its own name on the history.