LanesGuide · MCP
An MCP server for team context between agents
In shortAn MCP server for team context lets any MCP-capable coding agent ask what teammates' agents are working on, whether a file is being changed right now, and what was said in the team chat. Lanes ships one, lanes mcp, with seven tools. Add it with one line, for example claude mcp add lanes -- node /path/to/lanes.cjs mcp. Its tools are information only, and there is no tool to approve a request, so only a person can.
The gap MCP can fill
A coding agent knows its own conversation and the files it reads. It does not know that a teammate's agent is renaming the function it is about to call, or that someone posted "don't touch the auth module today" in chat. An MCP server can expose that knowledge as tools, and any MCP-capable agent can call them.
What a team context server should offer
A useful set of tools answers the questions an agent should ask before and during a task:
| Question | Lanes tool |
|---|---|
| What are teammates' agents working on right now? | team_brief |
| Is a teammate changing this file, and how? | team_check_file |
| Who is on the team, and what is each agent doing? | team_status, team_members |
| How do I tell the team something? | team_say |
| How do I ask a teammate's agent for help? | team_ask |
| What happened to the requests I sent? | team_requests |
Each answer is short and plain text. team_brief is capped like the brief the hooks add to prompts, so it never floods the context window.
Rules a team context server should follow
Information, never instructions. Text written by a teammate, such as a chat message, reaches another person's agent. It must arrive as quoted data with its author named, never phrased as an order. Otherwise a chat message becomes a prompt injection into your teammate's agent.
No approval through MCP. A request lets one person's agent act in another person's repository. If an agent could approve requests, one prompt injection would be enough to let a stranger's text run with your permissions. Lanes has no approve tool; a person approves in VS Code, the browser app or with lanes approve.
Never block. The tools report, and the agent decides. Nothing in the MCP server refuses an edit.
Add Lanes' MCP server to your agent
Build Lanes once, then add one line to your agent. For Claude Code:
claude mcp add lanes -- node "/path/to/lanes.cjs" mcpFor Codex, Gemini CLI, Copilot CLI, Qwen Code and Amp there is a similar mcp add command. Cursor, Kiro, Antigravity, Factory Droid, Augment, Junie, Warp, opencode, Kilo Code, Continue and Zed take a small JSON or YAML entry. The MCP reference lists each one.
Then tell the agent once, or put it in your project instructions: "Call team_brief when you start a task and team_check_file before you edit a file."
MCP and hooks together
MCP depends on the agent choosing to call a tool. Hooks run on every edit whether the agent thinks of it or not. Lanes uses both: where an agent supports hooks, they queue edits, sync files and add the brief automatically, and the MCP tools add the parts an agent should be able to ask for. For agents with no usable hooks, such as Zed or JetBrains AI Assistant, MCP is the whole integration.
Questions
What is the Model Context Protocol?
MCP is an open protocol that lets an AI agent call tools and read data from local or remote servers. Most coding agents, including Claude Code, Codex, Cursor and Gemini CLI, can use MCP servers.
Is MCP enough on its own?
For context, yes. For preventing overwrites, no: an agent has to choose to call a tool, so it can skip it. Lanes uses hooks where an agent supports them, which run on every edit whether the agent asks or not, and MCP for everything else.
Can an agent approve a teammate's request through MCP?
No. Lanes' MCP server has no approve tool on purpose. Approving a request that lets another person's agent act in your repository stays a human decision.