Connect your coding agent

If you used the agent setup prompt, your agent has already done this for its own host; this page is the manual route and the reference for other clients.

Once Kibi is installed in the repository, your agent needs a way to call it. Two surfaces expose the same operations:

MCP server (kibi-mcp) — 16 tools such as kb_search, kb_check, and kb_upsert show up in the client's tool list. kb_search also answers plain questions, such as "what governs checkout rounding?", with the requirements that apply. The MCP reference lists every tool.
CLI (kibi) — the same operations as JSON on stdin, plus commands you run yourself, such as kibi report.

Use whichever surface your client can see. You do not configure both unless you want to.

Claude Code

The kibi-claude plugin brings the server, the bundled skills, and hooks that show the agent the linked requirements and tests before it reads or edits code, and before an edit, what those requirements must keep true and the decision behind them. The same context reaches reads and writes the agent makes through shell commands such as cat, sed -i or a script, and an in-place write such as sed -i hears what must stay true before it runs:

bash

claude plugin marketplace add Looted/kibi
claude plugin install kibi-claude@kibi

Without the plugin, register the server for the project. This writes .mcp.json, so the whole team gets it:

bash

claude mcp add --scope project kibi -- npx --no-install kibi-mcp

See Claude Code plugin.

Cursor, Codex, and most other clients

Add this server to the client's MCP config. The working directory must be the repository where you ran kibi init, so the server attaches to that project's .kb/.

json

{
  "mcpServers": {
    "kibi": {
      "command": "npx",
      "args": ["--no-install", "kibi-mcp"]
    }
  }
}

--no-install matters. The client starts the kibi-mcp already in the project. It does not download a different copy at launch.

Plugin pages in the installation guide:

OpenCode

OpenCode takes the command as a list of tokens, in opencode.json:

json

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "kibi": {
      "type": "local",
      "command": ["npx", "--no-install", "kibi-mcp"],
      "enabled": true
    }
  }
}

The optional kibi-opencode plugin adds prompt guidance and background maintenance. It does not replace the server above. See OpenCode MCP.

VS Code

Create .vscode/mcp.json:

json

{
  "servers": {
    "kibi": {
      "type": "stdio",
      "command": "npx",
      "args": ["--no-install", "kibi-mcp"]
    }
  }
}

See VS Code MCP.

pnpm or Yarn

Keep the same shape and swap in that package manager's local runner. For pnpm, the command is pnpm with args ["exec", "kibi-mcp"]. For Yarn, yarn with args ["exec", "kibi-mcp"]. The installation guide has the full table.

No MCP? Use the CLI

If the agent cannot see MCP tools, it can call the same operations through the CLI:

bash

printf '%s\n' '{"query":"login","limit":10}' | npm exec -- kibi search --input -

Agents discover this path from the agent onboarding snippet. You do not configure it beyond installing the packages.

The first prompt

After the tools are visible, ask:

Bootstrap Kibi for this repository.

The agent first asks where product intent already lives outside the code, such as your issue tracker or wiki, and reads those sources through its other connectors. Connect those tools to the agent too if you want them used. It then shows a read-only plan and a hash, and it waits for your approval before writing. After that, ordinary prompts — features, fixes, refactors — keep the model in step with the code.

What stays in your hands

Product calls. If two behaviors would contradict, you pick one. Kibi will not paper over that.
Approval of the plan. Bootstrap and other bulk writes show you the plan before they apply it. A plan compiled from a prompt applies in full or not at all; if the agent is interrupted mid-write, its next write completes or rolls back that plan.
The report. You decide whether a gap is work still to do, or a behavior you no longer want.

The agent writes the model through typed operations. It does not edit .kb/ by hand. A change is not finished while checks are failing or the snapshot is stale.