Codag
Sign in

Today we’re launching Codag, and the free tier is live. Install it and point your agent at real logs in about a minute, no account required.

Debugging runs through an agent now. Claude Code, Codex, whatever you point at the incident. The moment it reaches for the logs, it hits a wall. Two of them.

Problem one: tokens and context

For an individual, the wall is cost. Every line your agent reads is tokens, and a routine log tail is tens of thousands of them. Your context window fills with noise long before the agent reaches the line that matters, you pay for all of it, and when the window runs out the agent compacts and loses the thread. Reading logs is the most expensive, least rewarding thing you can ask an agent to do.

Codag compresses the logs before the agent ever sees them. A million lines become a compact, schema-valid capsule: the handful of lines that matter, cited, plus a summary of the rest. Same answer, a fraction of the tokens. On the free tier, that is the difference between letting your agent read logs and not being able to afford to.

Problem two: scale

For a team, the wall is bigger. A single service emits millions of lines a day, and a real incident spans the whole fleet at once. That is far more than any agent can hold, so it truncates, samples, and guesses. It is genuinely bad at debugging across that much surface area, because it never sees most of it. I hit this on infrastructure at Okta and on web platform at Shopify, where OpenTelemetry was wired through everything and the answer still came down to a person grinding through logs by hand.

Codag works at that scale. We pull the signal out across services, not one pod at a time, and hand the agent a capsule it can actually reason over: root cause, trigger, and consequence, each cited back to a real line. The agent stops drowning and starts debugging.

Free to start

One command wires Codag into Claude Code and Codex. No account required.

$ curl -fsSL https://codag.ai/install.sh | sh

This is the start

We start with logs because that is where agents drown first. It does not stop there. Anywhere an agent has to understand a system through more output than it can hold, we intend to be the layer underneath.

Michael ZhouFounder, Codag