Coding agents
Your coding agent can see what Polylane sees. Once connected, it can query telemetry across every connected provider, traverse the topology, search code, read issues and their investigations, and save memories. It answers questions like "what errored in production for the service this branch deploys, in the last hour?" without you leaving the editor.
Two surfaces
Supported coding agents
Claude Code, Cursor, and OpenAI Codex install all of this in one step as a plugin: both MCP servers, the skills, and slash commands for the common workflows, from each agent's plugin directory.
Any coding agent that speaks the Model Context Protocol can connect to the MCP server. Platform MCP has per-client install steps.
The CLI works from any agent that can run shell commands, and it ships with a skill that teaches your agent when and how to reach for it. See CLI and Scripting.
Reads by default, writes behind a gate
Agent tools are read-only by default on both the MCP server and the CLI. Telemetry queries, topology traversal, and code search always work. Write tools, the ones that mutate provider resources, stay hidden and denied until you opt in:
- On the MCP server, send the
x-polylane-allow-writes: trueheader with a credential that carries theagent_tools:writescope. - On the CLI, pass
--writetopolylane tools runorpolylane tools codewith a credential that carries the same scope.
Even with writes enabled, a safety model screens every write call and refuses egregious ones. On the MCP server, clients that support confirmation prompts also ask you to approve each write before it runs.
Which surface to pick
- Claude Code, Cursor, or Codex: install the plugin and get the MCP servers and skills together.
- Any other agent living in your editor: use Platform MCP.
- Scripts, continuous integration, or terminal agents: use the CLI. Start with Authentication.
Once connected, try a prompt like this: