codag pro

Debug large logs
without reading them all

Codag Pro hands your agent just the lines that matter, with the surrounding context, each one cited to a real line number, so it can debug a 100k-line window without grepping the whole thing. Same agent, same prompts.

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

See the benchmark See pricing →

01 · how it works

From a 100k-line window to the lines that matter

Group similar lines

→

Rank by severity, rarity, query

→

Keep surrounding lines

→

Return compact text

02 · benchmark

Codag Pro vs raw plus grep vs codag-drain

The same incidents, read three ways, paired at 1k / 10k / 100k lines.

98–99%

incidents answered
at every window size

5.9–7.5×

fewer agent tokens
vs raw + grep

100%

root evidence present
across the run

invalid outputs
0 judge errors

Debugging accuracy by window size

whiskers = 95% bootstrap CI · n=102 paired runs / window (51 incidents × 2 agents) · Codag Pro vs raw p<1e-4

Agent tokens by window size

bar = median, whisker = p90 · linear scale · lower is cheaper

51 incidents × 2 agents · blind cross-provider judges · exact paired sign test · bootstrap-2000 CIs · methodology & artifacts

03 · drop in

One command, no surprises

Tiny logs pass through raw
Errors fall back to deterministic drain
Every kept line cites a real line number

$ codag wrap -- kubectl logs api
any log command, compacted

$ codag setup
Claude Code: hook + MCP, wired

$ codag mcp serve
Codex or any MCP client

$ curl https://api.codag.ai/v1/compact
REST · /v1/compact over HTTPS

Give your agent logs it can actually read.

Start on the free tier. Go Pro when your windows get large.