AI assistants are getting better at doing things, not only answering questions. The Model Context Protocol, or MCP, is an open standard that lets an assistant connect to a tool and use it. KanbanWP includes an MCP server, so an agent such as Claude, or any other MCP-compatible client, can work with the boards on your site.
This post explains what an agent can and cannot do, how access is controlled, and a few ways to use it safely.
What the server exposes
The server is an HTTP endpoint on your site. An agent calls it with JSON-RPC and gets back the result. It offers seven tools:
| Tool | What it does |
|---|---|
list_boards | Lists your boards |
get_board | Returns a board with its stages and tasks |
list_tasks | Lists tasks on a board, optionally for one stage |
create_task | Creates a task with title, description, stage, priority, assignee and due date |
update_task | Changes a task’s title, description, priority, assignee or due date |
move_task | Moves a task to another stage and position |
add_comment | Adds a comment to a task |
If you create a task without naming a stage, it lands in the board’s first stage.
What it cannot do
There is no delete tool. An agent cannot remove tasks or boards through MCP, so destructive actions stay in wp-admin or WP-CLI. That is a deliberate guard against a mistaken instruction wiping out work.
Turning it on
MCP is off until you enable it. In the plugin settings you enable MCP and generate a token. The agent sends that token as a bearer token in the request header. Some browser-based connectors cannot set headers, so the token can also be passed inside the request itself. The step-by-step guide is in connecting an agent, and the overview is in the MCP documentation.
Treat the token like an admin key
This is the most important point. The MCP token grants read and write access to every board. Anyone holding it can create, edit and move tasks. So:
- Store it in your agent’s secret storage, never in a shared document or a public repository.
- Rotate it straight away if it leaks or if a team member who knew it leaves.
- Use HTTPS only.
- Give each agent a clear name (see below) so you can tell them apart in the log.
Audit trail
Every tool call is written to the board’s activity log with the agent’s name. You identify your agent with a header, or a value in the request; if you do not, it is recorded as plain “mcp”. Comments added through MCP are attributed to the agent in that log. Writes also trigger the same events as normal changes, so notifications and webhooks behave as usual.
Ideas for using an agent
- Triage: have an agent read an inbox export or meeting notes and create tasks with priorities in the first stage, for you to review.
- Standup digest: ask an agent to read a board and tell you what is blocked or overdue.
- Housekeeping: ask it to move cards whose checklist items are complete, or to add a comment asking for an update on stale ones.
- Coding assistants: let a coding tool open a card for each bug it finds, and comment when it has a fix.
Guardrails that work
- Start with a test board, not your live client board.
- Ask agents to create cards, not rearrange everything, until you trust the result.
- Review the activity log after the first few sessions.
- Remember the agent has access to all boards through the token. If a board holds sensitive material, do not connect an agent you do not fully control.
- Remember that an agent reading your boards sends their contents to whichever AI service it runs on. Check that service’s terms.
MCP or the built-in AI?
They solve different problems. The built-in features, covered in Bring your own AI key, run inside the plugin with caps and logging, and are triggered by your team. MCP lets an external agent drive the board from outside. You can use either or both.
Developers who prefer scripting can also use the REST API or WP-CLI.
See how this fits the bigger picture in AI project management, and how other tools compare in best AI project management tools in 2027.