AgentDock

1.7k
Dock Agent Api
Claude Code

Claude Code

Use Claude Code with the Dock workspace. MCP for documents, citations, suggestions, and publishing. CLI for login and one-shot commands. SKILL.md so the agent knows the house style before it writes.

If your work happens inside Claude Code, the Dock workspace is the long-term memory the agent can read and write. Add the MCP server once, log in once, and Claude Code can browse your library, draft documents, propose tracked changes, manage citations, and publish to /p/slug.

This page covers both halves: the MCP server (the agent uses Dock as tools) and the CLI (you drive Dock from the terminal alongside Claude Code).

Claude Code already speaks MCP. The Dock workspace is exposed as an MCP server over streamable HTTP. There is no local process to babysit, no stdio bridge, no Docker. Add the server config to ~/.claude.json and the next session sees 38 tools across documents, citations, suggestions, folders, versions, prompts, settings, and publishing.

The CLI is the second path. agentdock runs from your shell, ships through npm, and is useful when you want to script things outside the Claude Code loop (cron, CI, a quick keystroke).

Screenshot
Claude Code session listing the AgentDock MCP tools after first connection.

Open ~/.claude.json and add an entry under mcpServers:

{
  "mcpServers": {
    "agentdock": {
      "type": "http",
      "url": "https://app.agentdock.ai/api/mcp",
      "headers": {
        "x-api-key": "ak_usr_YOUR_TOKEN"
      }
    }
  }
}

Restart Claude Code. The tools appear in the next session.

If you would rather not edit JSON by hand, install the CLI first (next section), then run:

agentdock mcp-setup

That prints the exact JSON above with your token already filled in. Paste it in.

Screenshot
agentdock mcp-setup terminal output showing the JSON with a real token.

npm install -g agentdockai
agentdock login

login opens the browser to authorize the device. Once you approve, the CLI stores an ak_usr_* API key under ~/.agentdock/config.json. The same key powers the MCP header above.

Verify with:

agentdock whoami
agentdock docs prompts
Screenshot
agentdock login terminal flow, browser approval screen, success message.

Ask Claude Code in plain English. The agent picks the right tool. Examples that work today:

Pull a prompt and write from it

Use the AgentDock prompt called "blog post outline" to draft a 1200-word
piece on RAG eval methodologies. Save it as a new document called
"RAG eval methodologies draft".

Under the hood: get_user_prompt -> create_document -> write_document. The agent fills the template, drafts the prose, and saves it. Title and folder default to your settings unless you specify.

Add citations as you go

Add citations for the three papers I mentioned: the GPT-4 technical report,
the original RAG paper by Lewis et al., and the LangChain documentation page
on retrievers. Use APA style.

The agent calls add_citation per source (DOI, URL, PMID, or arXiv ID), then update_document_settings to lock the style.

Propose tracked changes instead of overwriting

Read my "Q3 board memo" draft and propose tracked changes that tighten
the financial summary. Don't overwrite. I want to accept or reject.

read_document -> suggest_edits (find-and-replace operations) -> list_suggestions. You review in the editor and accept or reject each one.

Publish when it is ready

The "RAG eval methodologies draft" is ready. Publish it with the SEO
title "How to Evaluate Your RAG Pipeline in 2026" and a meta description
about RAG eval methodologies for engineers.

publish_document writes the page to /p/slug with the SEO metadata you dictated.

Video
60-second screencast of the four workflows above stitched together. Title: Claude Code + AgentDock: draft, cite, suggest, publish in one session.

Every MCP call needs the x-api-key header today; add it and the document, citation, suggestion, folder, version, and publishing tools light up. Coming soon: a keyless tier where the agent can browse the public prompt library without a key.

The CLI uses the same API key. agentdock login writes it; agentdock logout clears it. Tokens are user-scoped, never org-scoped, so what the agent can touch is exactly what you can touch.

See Authentication for token types and revocation.

Tools do not appear after editing ~/.claude.json. Restart Claude Code. The MCP server list is loaded at session start.

401 Unauthorized from MCP calls. The x-api-key is missing or expired. Run agentdock login again, then agentdock mcp-setup and re-paste.

Claude Code drafts off-style. The MCP server ships a SKILL.md the agent should read first. If your agent skips it, add this to your project's CLAUDE.md: Before writing to any AgentDock document, call list_user_prompts and read the SKILL.md exposed by the server.

AuthenticationCLI Reference