Stop explaining production to your agent. Let it see for itself.
Polylane gives your coding agent your whole production stack over one MCP server.
claude — ~/checkout-edge > why is checkout slow in prod?
●polylane · searchissues · checkout-edge
●polylane · runToolqueryMetrics · p99 by deploy
●polylane · runTooldeployHistory · checkout-edge
⏺ Deploy 9f3c2a1 shrank the Hyperdrive pool 50 → 5; P99 followed within a minute. Polylane already raised the issue, and the investigation carries the receipts.
The fix is a one-line pool restore.
Want me to open it?
❯ 1. Yes, open the pull request
❯ 2. Yes, and run the regression tests first
❯ 3. No, tell Claude what to do differently (esc)
Bring your own agent. Any MCP client works with the same server.
Your entire stack sits behind one server. Your agent sees all of it at once.
Your clouds
AWS, Cloudflare, Vercel, Fly.io, Render, Kubernetes, PlanetScale, Railway, Supabase, Modal, Convex, ClickHouse, Turso, with every resource and dependency live in the context graph, so your agent sees how production is actually wired.
Your telemetry
Logs, metrics, and traces through Datadog, Honeycomb, Axiom, Grafana Cloud, Sentry, Better Stack, OpenStatus, and Logfire: your agent reads the same data your dashboards chart.
Your code and deploys
GitHub, deploy history and code search, so your agent knows what shipped, when, and what it touched.
Your memory
The workspace keeps notes on past incidents: what worked, and the queries that found the cause. Your agent starts every question already knowing your system.
A handful of tools. Your agent won't drown in five hundred.
of the context window spent on tool schemas, before a single question is asked.
a handful of tools, and a schema only loads when that tool runs.
tools/list The current tool list, with every schema, is the card the server publishes about itself: server-card.json
Prefer the CLI? One command sets it all up.
The setup script installs the CLI, registers the MCP server in every coding agent it finds, and signs you in. The other options do the same thing with the package manager you already have.
curl -fsSL https://polylane.com/setup | bash # macOS / Linux
npm install -g @coreplane/polylane # npm (Node.js 20+)
bun add -g @coreplane/polylane # bun
brew install coreplanelabs/tap/polylane # Homebrew
irm https://polylane.com/setup.ps1 | iex # WindowsTwenty tool calls become one program. That's code mode.
Instead of calling tools one by one, the agent writes a small program that calls them all and returns just the answer.
● queryMetrics → 41 kB of series
● deployHistory → 12 deploys
● queryMetrics → another 41 kB
● queryLogs → 3,000 lines
… 17 more round trips
Every intermediate result lands in the context window.
const deploys = await tools.deployHistory({ service: "checkout-edge" }); const { points } = await tools.queryMetrics({ metric: "latency.p99" }); return correlate(deploys, points);
Intermediate results stay in the sandbox. Only the answer comes back.
Common questions.
Which agents can use it?
Anything that speaks MCP: Claude Code, Cursor, Codex, VS Code, Zed, OpenCode, Pi and the rest. There is no Polylane-specific client.
What does the agent authenticate with?
It signs in with OAuth in a browser the first time it calls the server. CI and headless machines use an API key scoped to one workspace.
Can an agent change my infrastructure?
Not unless you opt in. Write tools stay hidden until your credential carries the write scope, and every write is screened before it runs.
Do I need the CLI as well?
No, the MCP server is a complete surface on its own. The CLI wraps the same API for the terminal and for scripts.
Give your agent production. It gives you back the fix.
Reading this as an agent? https://polylane.com/llms.txt · https://polylane.com/agent.json · https://polylane.com/.well-known/mcp · this page as markdown: https://polylane.com/for-agents.md