A task tracker your coding agent can actually use.

Every workspace has its own MCP server at https://<your-workspace>.kaiku.tech/mcp: streamable HTTP with a token in the Authorization: Bearer header. Nothing to install and nothing to run next to it.

Settings → Connect over MCP shows your exact address, makes the token and gives you the lines to paste, with the token already in them. For Claude Code that is one command:

claude mcp add --transport http kaiku https://<your-workspace>.kaiku.tech/mcp \
  --header "Authorization: Bearer <token>"

Cursor, Claude Desktop and every client configured by a file take the same address under mcpServers.

• Issues: search, read, create, edit, move to another project; comments; links between issues; a project’s own columns.
• The wiki: search, read, create, edit and delete pages; comments; tie a page to the issue it documents.
• Questions: put a question with answer options to one person, and read the answer when it comes — instead of guessing.
• Files: attach to an issue or a page and read them back.
• Agent runs: record what the run spent on an issue — tokens by kind and machine time.

An agent can read and write the same things a person can. The team sees its work on the same board, in the same threads, with the same notifications — not in a log somewhere else.

• The server acts as the owner of the token, with the permissions that person has.
• A token limited to one project limits every tool to that project.
• A read-only token may call only the reading tools.
• Its search refuses a query it cannot fully read instead of quietly widening it — an agent that asked for one project’s bugs does not get every project’s.

Kaiku speaks on the wire the REST API of the tracker and the wiki most companies already run. So the widely used open-source MCP server written for them works against Kaiku unchanged — both halves with one token, pointed at your workspace address. If your agent already has it configured, the only change is the address.

A hosted server cannot read your disk, and an agent should not transcribe a screenshot into base64 by hand. pm-files is a small local MCP server that attaches a file by its path and saves an attachment back to a file. One command installs it and adds it to Claude Code:

curl -fsSL https://<your-workspace>.kaiku.tech/api/mcp/local-files/install.sh | bash

The script takes the token from your clipboard and checks it before writing anything; the assistant never sees it. It reads only the folders you give it.

Every agent run is recorded against the issue it worked on: tokens by kind, machine time, who asked. A week of automation has a number on it while there is still something to do about it — per issue, not as one bill at the end of the month.

Kaiku also brings agents of its own. You call one in a comment, by name, in the thread where the question already is. It reads the issue, answers in the same thread and lists what it read to get there.

• It runs as the person who called it, with that person’s access at that moment, and only inside the issue’s project.
• Its tools are reads only. The comment it answers with is written by Kaiku from its answer — the model never writes to your workspace itself.
• It never decides. A proposal is a question addressed to you, with no default answer, so the issue keeps who chose and what they chose.
• Your own agent can call one over MCP, with ask_agent, and read the answer when it comes.

The built-in agents are part of the Team plans — on our AI, or on your company’s own key. The free trial has everything else, but no AI.