reads
# codag npm test printed 4,212 lines. your agent reads this instead: FAIL src/checkout.test.ts > applies discount codes AssertionError: expected 500 to equal 502 at src/checkout.test.ts:118 the other 4,180 lines were passing suites. skipped, and your agent can pull any of them back word for word.
the problem
Agents are great at reasoning
and terrible at reading
context
Tool output floods the window
Test runs, search results, file trees, logs. One big result can push out everything the agent already learned and force another expensive lap.
spend
You pay for every byte it reads
A session makes hundreds of tool calls, and a handful carry the answer. The rest is billed input the model reads once and throws away.
one command
Set up once
Go back to your agent
No wrapper command, no config files to edit, no new workflow to learn. Setup attaches the agents it detects and rolls everything back if any step fails.
codag setup
Attaches Claude Code and Codex
Sign in through the browser, or set CODAG_API_KEY for CI. Setup verifies your workspace with a live unbilled check, installs a per-user local service, and attaches the supported agents it finds.
Other tools are detected and reported, never modified. Everything is reversible with codag disable.
$ codag setup ✓ signed in via browser ✓ workspace verified, live reduction check passed ✓ local service healthy ✓ attached: Claude Code, Codex setup complete. use your agents normally.
then nothing changes
Same agent, same prompts, same model
Your provider account, your credentials, your gateways. Model traffic goes through a local proxy on your machine directly to your provider.
Codag only changes how large tool results reach the agent. Check it any time with codag status and codag savings.
$ claude > run the test suite and fix what breaks · npm test ran, output reduced to the failures · exact lines retrieved locally when needed · everything else passed through untouched
how it works
The short version up front
The exact bytes on demand
Large eligible results are reduced to what the agent needs. Nothing is lost.
observe and classify
Every result is classified by kind. Unrecognized passes through untouched.
reduce what is large
Large results shrink to the failures, matches, and signal lines.
retrieve exactly, locally
Omitted bytes stay encrypted on your machine, pulled back exactly on demand.
analytics
Every token,
accounted for
What your agents spent, what they would have spent, and the difference.
all metrics are counts, sizes, and decisions. never prompts, commands, paths, filenames, or output.
faq
Questions, answered
What is Codag?
Codag is a developer tool that cuts what coding agents cost to run. It teaches agents like Claude Code and Codex to read large tool results efficiently: big outputs are cut down to the parts that matter before the agent reads them, and the exact content stays retrievable on your machine. Setup is one command.
Does Codag change my model, my agent, or my prompts?
No. You keep the same agent, model, prompts, and provider account. Model traffic goes through a local proxy on your machine directly to your provider. Codag only changes how large tool results reach your agent.
What does Codag reduce?
Large eligible tool results: logs, test and build and lint output, search results, file listings, document reads, and agent handoffs. Source code, diffs, configuration, and exact structured data always pass through unchanged. When the agent needs an exact omitted section, it retrieves it locally by line range, JSON path, or group.
Is my code or data stored in Codag's cloud?
No. Eligible tool output is processed transiently to produce the shorter version and is not retained. Exact omitted content is kept encrypted on your machine for seven days with a 1 GiB cap. Billing runs on contentless metrics: counts and sizes, never prompts, paths, or output.
What happens if Codag goes down or I hit my quota?
Your agent gets the original tool output byte for byte. Codag fails open on timeouts, validation failures, unknown tool types, and exhausted quotas. A hard cap means passthrough, never a broken session.
How much does Codag cost?
Codag is sold to teams. Team is $499 per month with 200 GB of eligible tool output included and $0.50 per additional GB; Enterprise is contractual. Pilots run 14 days on your real workload, need no card, and never convert automatically. See pricing or book a demo.
Your agent works the same
Your bill does not
Book a demo and pilot Codag on your team's real workload. Setup is one command, and your developers keep working exactly as before.
14-day pilot, no card, never auto-charges.