Collie

---

What it is

Collie is a workbench for systematic reviews and meta-analyses: protocol, search, deduplication, screening, data extraction, risk of bias, meta-analysis, GRADE and PRISMA, in one place. This kit connects the AI agent you already use to your Collie account over the Model Context Protocol (MCP). The AI runs on your own AI subscription, not on Collie. Your agent works through the same tools you see in the app: it writes suggestions and drafts, and you accept or dismiss them in Collie.

Requirements

• A Collie account. Every plan can connect an agent, Free included. Sign up.
• Node.js 20 or newer.
• An MCP client: Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, Perplexity (Mac app) or any client that runs stdio MCP servers.

Install in Claude Code (recommended)

Claude Code gets the full kit: the MCP server, skills, subagents, hooks and the live Collie pane.

/plugin marketplace add leandromedeirosfelix/collie-agent-kit
/plugin install collie@collie-agent-kit

1. Restart Claude Code.
2. Ask: "Connect to Collie." The agent shows a link and a short code.
3. Open app.colliesr.com/connect, sign in and type the code to approve the agent. A link alone never approves it.
4. Type /collie to open the Collie pane.

Details: docs/install-claude-code.md.

Manual install (other MCP clients)

The server runs with npx -y github:leandromedeirosfelix/collie-agent-kit. Add it to your client's config, restart the client, then ask the agent to connect to Collie.

Settings › Developer › Edit Config opens claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "collie": {
      "command": "npx",
      "args": ["-y", "github:leandromedeirosfelix/collie-agent-kit"]
    }
  }
}

Quit Claude Desktop completely and open it again.

~/.codex/config.toml (shared by the Codex CLI, app and IDE extension):

[mcp_servers.collie]
command = "npx"
args = ["-y", "github:leandromedeirosfelix/collie-agent-kit"]
startup_timeout_sec = 60
tool_timeout_sec = 300

Or: codex mcp add collie -- npx -y github:leandromedeirosfelix/collie-agent-kit. The longer timeouts cover the first download and analyses that take a few minutes.

~/.gemini/settings.json (or .gemini/settings.json in a project):

{
  "mcpServers": {
    "collie": {
      "command": "npx",
      "args": ["-y", "github:leandromedeirosfelix/collie-agent-kit"]
    }
  }
}

~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project):

{
  "mcpServers": {
    "collie": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "github:leandromedeirosfelix/collie-agent-kit"]
    }
  }
}

Perplexity runs local MCP servers in its Mac app only.

1. Open Plugins and install the helper app PerplexityXPC.
2. Plugins › Add custom › Simple: server name collie, command npx -y github:leandromedeirosfelix/collie-agent-kit.
3. Wait until the connector shows Running, then turn on collie under Sources.

Field	Value
Command	npx
Arguments	-y github:leandromedeirosfelix/collie-agent-kit
Transport	stdio
Environment	optional, see below

Windows: if your client cannot start npx directly, use "command": "cmd" and "args": ["/c", "npx", "-y", "github:leandromedeirosfelix/collie-agent-kit"].

Optional environment variables

Variable	Default	Use
MSR_TOKEN	none	A personal access token from Collie › Settings › AI agents (MCP). Replaces the code sign-in.
MSR_URL	https://app.colliesr.com	The Collie address. Leave it unset.

All clients, step by step: docs/install-manual.md.

What you get

Part	What it does	Where
MCP server	Collie's tools for your agent: reviews, protocol, search, duplicates, screening, full text, extraction, risk of bias, analysis, GRADE, PRISMA	Every client
13 skills	One per stage of a review, plus help. Start with collie-workflow	Claude Code
8 subagents	Methodologist, search strategist, screener, extractor, risk-of-bias assessor, biostatistician, manuscript writer, peer reviewer	Claude Code
Live pane	/collie: connection, the active review's stages with progress bars, AI suggestions waiting for you, your agent's recent Collie calls. The status line shows the review and screening progress	Claude Code
Guardrail hooks	Ask you before the agent casts a real screening vote or installs R packages; check the syntax of .R files it edits; report the Collie connection at session start	Claude Code

Skills and subagents are short stubs. Each one asks Collie for its full instructions with the collie_guide tool, which answers only for a signed-in account.

Using it

Ask in plain words, for example:

• "Where is my review and what's next?"
• "Is a meta-analysis on early mobilisation after hip fracture surgery viable?"
• "Pre-screen the next 50 records against my criteria."
• "Extract the outcome data for the included studies."
• "Draft the risk-of-bias assessments for the included trials."

The agent writes suggestions and drafts. You accept, edit or dismiss them in Collie. A screening vote by the agent is a real reviewer vote, so Claude Code asks you first.

Privacy and security

• The kit talks only to app.colliesr.com (or your MSR_URL), with your token. Nothing else is sent anywhere.
• Your AI client sends what the agent reads to your AI provider, under your agreement with that provider.
• The token is stored on your computer: %APPDATA%\collie\token.json on Windows, ~/.config/collie/token.json on macOS and Linux (readable by your user only on macOS and Linux).
• Disconnect this computer: ask the agent to run collie_disconnect. Revoke the token itself: Collie › Settings › AI agents (MCP).
• No token can delete anything in Collie.
• The kit contains no review method. Collie serves its guides to signed-in users.

More: docs/security.md.

Updating

• Claude Code: /plugin marketplace update collie-agent-kit, then restart Claude Code.
• Other clients: npx fetches the kit from GitHub when the client starts the server. Restart the client. If an older version keeps running, see troubleshooting.

Uninstall

• Claude Code: /plugin uninstall collie@collie-agent-kit, then /plugin marketplace remove collie-agent-kit.
• Other clients: remove the collie entry from the client's config.
• Then delete the token file above (or run collie_disconnect first) and revoke the token in Collie › Settings › AI agents (MCP).

Troubleshooting

Problem	What to do
"Not connected to Collie"	Ask the agent to connect, then type the code at app.colliesr.com/connect.
The code expired, or the token was revoked	Ask the agent to run collie_connect again.
npx not found	Install Node.js 20+, restart the client. On Windows, try cmd /c npx.
Behind a corporate proxy	See troubleshooting.

Full list: docs/troubleshooting.md.

For developers

The developer documentation covers how the kit works and how to build on it:

• Architecture: the server, the plugin layers, what runs where, and the human-in-the-loop guarantees.
• Authentication: the code sign-in, tokens, scopes and revoking.
• Tools reference: every MCP tool with its parameters, generated from the server.
• Skills and subagents, hooks and the Collie pane.
• Integrating other clients: any stdio MCP client, or your own TypeScript or Python code.
• Versioning.

Contributions are welcome: see CONTRIBUTING.md. Report vulnerabilities privately as SECURITY.md describes.

Support

Write to [email protected].

License

The agent kit is open source under the MIT License. Copyright (c) 2026 CollieSR Company. Open-source components in the server bundle: THIRD-PARTY-NOTICES.md.

The Collie service (app.colliesr.com) and the method guides it serves through collie_guide are not part of the kit. They are governed by Collie's Terms of Use.