# Corduroy Labs > An AI Architecture Studio in Truckee, California. We build agent-first foundations on open systems for the people who run the work, the people who use it, and the agents that help them both. Corduroy Labs designs and ships the operational layer that lets small businesses run with agents as first-class collaborators — role-based inboxes, policy-as-code, approval workflows, model routing, and self-hosted infrastructure. Written for the operator, not the platform. ## About - [Home](https://corduroy-labs.ai/): studio overview, positioning, contact. - [About](https://corduroy-labs.ai/about): how we work and why. - [Contact](https://corduroy-labs.ai/contact): how to reach us. ## Feeds - [RSS 2.0](https://corduroy-labs.ai/feed.xml): traditional syndication with content:encoded and media:content extensions. - [JSON Feed 1.1](https://corduroy-labs.ai/feed.json): modern feed format for JS-native aggregators and AI summarizers. ## Updates - [Quiet by design.](https://corduroy-labs.ai/updates/quiet-by-design/) (2026-09-12): We build AI operations for a living. Last week we shipped a website with no AI in it at all — on purpose. Notes on building a portfolio site for a teenage artist who writes everything herself: privacy in the pipeline, two themes on one token system, and why the workshop and the product are different rooms. - [The agent that maintains the agent.](https://corduroy-labs.ai/updates/the-agent-that-maintains-the-agent/) (2026-08-28): Our COO agent's runtime was eleven weeks and four releases stale. Notes from update night: why maintaining an agent is different from maintaining software, the checklist that is the actual update, and the moment we handed the runbook to another agent. - [An agent that shows its work.](https://corduroy-labs.ai/updates/an-agent-that-shows-its-work/) (2026-07-19): Cords, our COO agent, now runs the @corduroy-labs.ai Bluesky feed within audited caps. Notes on giving an agent room to move, one honest failure we caught in real time, and why observability is the substitute for control. - [Deploys that refuse to break the site.](https://corduroy-labs.ai/updates/fail-closed-deploys/) (2026-07-17): We replaced our static-site deploy with a fail-closed script: build, verify, scrub for strings that should never go public, then swap atomically. If any gate fails, the live site never notices. Notes on why rsync --delete is a loaded weapon and what we do instead. - [The studio, machine-readable.](https://corduroy-labs.ai/updates/machine-readable-corduroy/) (2026-07-09): This week we shipped RSS, JSON Feed, an llms.txt, and a small MCP server for the studio. Notes on why a small business site should be legible to agents, why we picked all three, and the pattern the MCP server is really trying to demonstrate. - [Self-improvement, with direction.](https://corduroy-labs.ai/updates/self-improvement-with-direction/) (2026-06-19): We let the agent get better at its job this week, but on a leash. It now checks its own claims against the world before showing them to us, and it proposes its own next reference documents — drafts in a folder until we approve them by email reply. - [The agent becomes a teammate.](https://corduroy-labs.ai/updates/the-agent-becomes-a-teammate/) (2026-06-16): Today we gave the agent a shared workspace and a small cabinet of specialized minds. A shared vault, end-to-end encrypted, syncing on every device. A router that picks the right model for the job. The difference between a tool you operate and a teammate you collaborate with. - [Boundaries make the agent trustworthy.](https://corduroy-labs.ai/updates/boundaries-make-the-agent-trustworthy/) (2026-06-14): Notes from a week of operational hardening: narrowing what the agent can touch, writing the rules down, verifying reality instead of configuration, and giving the agent some real context about the operator. - [Take a breath. You can become agentic-first without the stress.](https://corduroy-labs.ai/updates/agentic-first-business-operating-system/) (2026-06-01): Our first update on building a practical operating layer for people and agents, starting with the workflows that slow work down. ## Public artifacts The studio publishes reusable code and patterns directly from this site: - [corduroy-labs-mcp v0.1.1](https://corduroy-labs.ai/artifacts/corduroy-labs-mcp-v0.1.1.tar.gz): a small Model Context Protocol server that projects our JSON Feed into MCP tools (`list_updates`, `get_update`). MIT-licensed. Reference implementation of "any site with JSON Feed already has an MCP-shaped surface." - [Live MCP endpoint](https://corduroy-labs.ai/mcp): Public and rate-limited, ready for any MCP client (Claude Desktop, Cursor, n8n). ## Optional - [Feed pattern write-up](https://corduroy-labs.ai/updates/machine-readable-corduroy/): the studio's own machine-readable surfaces, described in the operator's voice.