Last updated: 2026-11-09
Codag is a cost layer for AI coding agents that shrinks large tool results, such as test runs, logs, search hits and file trees, before Claude Code or Codex reads them. Its homepage cites about 75% compression, and any omitted line stays encrypted on your own machine for exact retrieval.
About Codag
Codag Inc. is a San Francisco company from the Y Combinator Summer 2026 batch. It launched around log compression for incident triage, and by November 2026 its homepage pitches a wider job: cutting what coding agents such as Claude Code and OpenAI Codex spend reading their own tool output. A single test run can print 4,212 lines when the agent needs only the one failing assertion, and Codag hands the agent that failure instead, with every omitted line retrievable word for word.
Setup is one codag setup command. It signs you in, runs a live unbilled check, installs a local service and attaches the supported agents it detects; codag disable rolls it back. Model traffic goes through a local proxy straight to your own provider account, so the model, prompts and credentials stay yours. Large eligible results (logs, test, build and lint output, search hits, file listings, document reads) are reduced; source code, diffs and configuration always pass through untouched. If Codag times out, fails validation or runs out of credits, the agent gets the original output byte for byte.
The company also keeps open-source pieces on GitHub: codag-drain, an MIT-licensed Rust log-templating library, the codag-sdk clients for Python, TypeScript and Go, and the codag-visualizer VS Code extension. For how it sits next to other agent-ops tools, see Codag vs Crukx and our map of the three jobs in AI agent observability. Testing an agent before release is a different job, covered by Archal in Archal vs ReasonBlocks.
Who should skip it: a team on an agent other than Claude Code or Codex, since only those two are attached today, and a team whose spend problem is attribution across many providers, where SuperPenguin answers the question of who spent what. If the waste is bad production answers rather than tool-output bloat, Maitai works on the model traffic itself, and Cursor users have nothing to attach yet. Anthropic and OpenAI bill the tokens either way; Codag only changes how many of them a tool result costs.
Pricing
Codag bills in credits at $0.05 each, as of 9 November 2026. Pay as you go starts with 10 free credits (about 1 GB of log output) as a one-time grant that never expires and needs no card, then packs of about 1 GB for $0.50, 10 GB for $5 or 50 GB for $25. Credits burn at different rates per tool: log compaction uses about 1 credit per 100 MB, while inference-based agent-context reduction uses about 1 credit per 100 KB, roughly 1,000 times faster.
Team is $499 a month for 12,500 credits with overage at $0.50 per GB; Enterprise is a custom annual contract. When credits run out, optimization pauses and output passes through unchanged.
| Tier | Monthly price | What it includes |
|---|---|---|
| Pay as you go | Free | Claude Code, Codex, CLI, API and MCP, Three fixed credit packs or a custom pack, bought any time, Optimization pauses when credits run out, no overage bill |
| Team | $499/mo | Inference pipeline for CLI, API and MCP, Per-member usage attribution, Priority support with a 24-hour SLA, Dedicated Slack channel |
| Enterprise | Custom | Contractual volume, SSO / SAML, VPC peering, Client version pinning, Custom SLAs |
Feature Comparison by Tier
| Feature | Pay as you go | Team | Enterprise |
|---|---|---|---|
| Monthly price | $0 + credits | $499 | Custom |
| Included credits | 10 one-time | 12,500 a month | Contractual |
| Inference pipeline | — | ✓ | ✓ |
| Per-member attribution | — | ✓ | ✓ |
| Support | Standard | 24h SLA, Slack | Custom SLAs, Slack |
| SSO / SAML | — | — | ✓ |
Key Features
- Tool-output reduction: Shrinks large test, build, lint, log, search and file-listing results to the failures and signal lines; the homepage cites about 75% compression.
- Exact local retrieval: Omitted bytes stay encrypted on your machine for 7 days under a 1 GiB cap, so the agent can pull any skipped line back word for word.
- One-command setup: `codag setup` attaches Claude Code and Codex, verifies the workspace with an unbilled check, and is undone with `codag disable`.
- Fails open: Codag says timeouts, validation failures, unknown tool types and an empty credit balance all return the original output, across its 4 documented fallback cases.
- Savings analytics: `codag savings` reports tokens avoided and an estimated dollar saving at public model prices, using counts and sizes only.
- SDKs and API: Official clients in Python, TypeScript and Go expose a single reduce call plus usage and workspace policy for CI and custom harnesses.
Pros
- The bet is low-risk: because it fails open, the worst case is paying for nothing on one tool call, never a broken agent run.
- Model calls run through a local proxy to your own provider account; only the eligible tool output goes to Codag, which says it does not retain it.
- Log compaction costs cents per gigabyte, which is cheap next to the model tokens the same output would cost to read raw.
Cons
- Only Claude Code and Codex are attached automatically; other agents are detected and reported but not changed.
- Inference-based context reduction burns credits about 1,000 times faster than log compaction, so the per-GB price depends on what you reduce.
- The step from pay as you go to Team is a large flat monthly fee, with no plan in between.
- No compliance certification is named on its home or pricing pages; regulated buyers are pointed to an Enterprise call.
Product Information
- Cloud
- Yes
- Languages
- English
- Training
- Documentation, SDK docs
Data Handling
- Training-data policy
- Eligible tool output is processed transiently and not retained; omitted content stays encrypted on the user's machine for 7 days (1 GiB cap); billing uses counts and sizes only.
- Data retention
- Zero retention
Frequently Asked Questions
What does Codag cost per month?
Pay as you go has no subscription: you buy credits at $0.05 apiece, and logs spend roughly one credit per 100 MB, so about $0.50 per GB. The $499 Team plan bundles a monthly credit allowance and bills extra use at the same per-GB rate, and Enterprise is quoted on an annual contract.
Can I use Codag for free?
Yes, to start. A new account receives 10 credits at no charge, enough for roughly a gigabyte of logs, with no card and no expiry date. After that you top up by the credit or the pack; when the balance hits zero, optimization simply pauses and your agent reads the original output.
What are the best Codag alternatives in 2026?
It depends on the cost problem. SuperPenguin attributes AI spend across providers rather than reducing it, Maitai sits on production LLM traffic and corrects bad output, and Crukx now optimizes and verifies models for your hardware. None of them shrinks a coding agent's tool output the way Codag does.
Does Codag change my model or prompts?
No. Codag runs a local proxy that sends model traffic straight to your own provider account, and it only changes how large tool results reach the agent. Source code, diffs, configuration and exact structured data always pass through unchanged.
Is my code stored on Codag's servers?
Codag says eligible output is processed transiently and not retained. The exact omitted content is kept encrypted on your own machine for seven days, capped at 1 GiB, and billing uses counts and sizes only, never prompts, paths or output.
Top Alternatives
- SuperPenguin: Pick SuperPenguin if you first need to see where AI spend goes across providers and teams.
- Maitai: Pick Maitai if the waste is wrong production answers rather than bloated tool output.
- Crukx: Choose Crukx when you run your own models and want a smaller verified build for your hardware.
Hokai guides covering Codag
- Codag vs Crukx: Crukx Changed Jobs Again, So Pick by Job: Crukx moved from an agent release gate to model checkpoint storage. Compare it with Codag's $0.05 credits and see which job each tool fits as of November 2026.
Also mentioned in
- Archal vs ReasonBlocks: A Sandbox for Your Agent, or a Check on Its Tests: What is Archal AI? It gives AI agents stateful API sandboxes to test in, while ReasonBlocks audits the evals that score them. Prices, data access and who fits.
- ReasonBlocks vs Touchmark: Two Pivots, Two Different Problems: ReasonBlocks moved from cutting agent tokens to auditing evals; Touchmark prices reserved inference on a 20% capped curve. What each does and costs now.
- The Three Jobs Hiding Inside "AI Agent Observability" in 2026: AI agent observability now means three jobs: testing an agent before launch, checking the evals that score it, and cutting its cost. See which you need.