# Codag > Codag cuts coding-agent costs without changing how developers work. One setup command attaches it to Claude Code and Codex; it teaches the agent to read large tool results efficiently, keeps exact content retrievable on the developer's machine, and passes everything else through unchanged. Codag is a developer tool for teams running AI coding agents. It ships as a signed binary with a one-command setup. After `codag setup`, the developer keeps using Claude Code or Codex exactly as before; Codag changes only how large tool results reach the agent. Codag Inc. is based in San Francisco and is backed by Y Combinator. ## The problem Coding agents spend most of their budget reading, not reasoning. Tool output was designed for humans to skim, but an agent pays for every token: - A test run prints thousands of lines to report one failure. - A search returns a hundred files when the agent will use three. - A log tail buries the error under startup noise. - One oversized result floods the context window and forces expensive re-reads. ## How Codag works 1. Install: `curl -fsSL https://codag.ai/install.sh | sh`. The installer verifies checksums; releases carry provenance and build attestations. 2. Once the team's workspace is active, each developer runs `codag setup`. It signs the developer in (browser by default, `CODAG_API_KEY` for CI), verifies the workspace with a live unbilled reduction check, installs a per-user local service, attaches detected Claude Code and Codex installations, and registers a retrieval-only local tool. Any failure rolls the machine back. 3. From then on, Codag observes each tool call and classifies it. Large eligible results (logs, test/build/lint output, search results, file listings, document reads, agent handoffs) are reduced to the evidence the agent needs. Source code, diffs, configuration, exact structured data, and anything unrecognized pass through unchanged. 4. The exact content a reduction omits stays encrypted on the developer's machine for seven days (1 GiB cap). The agent retrieves exact slices locally by line range, JSON path, or group when it needs them. 5. Everything fails open: timeouts, validation failures, unknown tool types, and exhausted quotas all mean the agent receives the original bytes. Verify and manage with `codag status`, `codag savings`, `codag doctor`, `codag disable`, `codag uninstall`. All configuration changes are reversible. ## Privacy boundary - Model traffic goes from a local proxy on the developer's machine directly to their provider. Provider credentials never pass through Codag's cloud. - Only large eligible tool output plus minimal task context is sent to Codag, processed transiently on Codag-controlled inference endpoints, and not retained. - Codag's cloud keeps contentless accounting metrics only: byte and token counts, estimated costs, latency, decisions, and counters. Never prompts, commands, paths, filenames, task text, or output. - Optional product telemetry can be disabled with `codag config set telemetry off`. ## Pricing Codag is sold to teams; there is no self-serve individual tier. Metered on eligible tool output processed, not seats. Pilots run 14 days on the team's real workload, need no card, and never convert automatically; they are arranged on a demo call. - Team: $499 per month. 200 GB of eligible tool output included, $0.50 per additional GB, priority support with 24h SLA, up to 25 seats, dedicated Slack channel. - Enterprise: custom. Contractual volume, SSO/SAML, VPC peering, client version pinning, custom SLAs. - Caps mean byte-for-byte passthrough, never a broken session. Dollar savings reported by the product are estimates computed from public model prices. See https://codag.ai/pricing ## Benchmarks Codag publishes performance numbers only from accepted benchmark artifacts, each with provenance (artifact hash, product and harness hashes, model versions, price dates) and reproduction instructions. The program measures task success, provider cost per successful task, turns/retries/rereads, and eligible-traffic share against stock agents and version-pinned alternatives. Results for the current product are being finalized; until they land, no numbers are published. ## API and SDKs The action API (base `https://api.codag.ai`, bearer auth with `cdk_` workspace keys) centers on `POST /v1/actions/reduce`: send a tool call, its raw result, and an action kind; receive a `reduced` or `passthrough` decision with evidence, retrieval selectors, and usage. Companion endpoints: `/v1/metrics/batch` (contentless), `/v1/usage/summary|timeseries|breakdown|reliability`, `/v1/trials/report`, `/v1/model-prices`, `/v1/workspace/policy`, `/v1/service/status`. The original is never returned or retained. Official open-source SDKs speak the same API: Python (`pip install codag`), TypeScript (`npm install @codag/sdk`), Go (`go get github.com/codag-megalith/codag-sdk/go`). The OpenAPI spec lives in https://github.com/codag-megalith/codag-sdk ## Open source The production CLI/runtime is a closed-source signed binary; releases with checksums, provenance, and attestations are published at https://github.com/codag-megalith/codag-releases. Open source: the SDKs and OpenAPI spec (codag-sdk), the deterministic log-templating library codag-drain (MIT, Rust), and codag-visualizer for VS Code. See https://codag.ai/open-source ## FAQ Does Codag replace my agent or my model? No. You keep the same agent, model, prompts, and provider account. Codag only changes how large tool results reach the agent. Is my content stored in Codag's cloud? No. Eligible tool output is processed transiently and not retained. Exact omitted content stays encrypted on your machine for seven days with a 1 GiB cap. Billing runs on contentless metrics. What happens if Codag goes down or the quota runs out? The agent gets the original tool output byte for byte. Codag fails open by design. Which agents are supported? Claude Code and Codex at launch. Other tools may be detected and reported, but are never modified. Broader support is coming. ## Links - Overview: https://codag.ai/ - Docs: https://codag.ai/docs - Pricing: https://codag.ai/pricing - Open Source: https://codag.ai/open-source - Vision: https://codag.ai/vision - Blog: https://codag.ai/blog - Support: https://codag.ai/support