Last updated: 2026-08-19
Codag is a log compression layer that sits between raw production infrastructure logs and an AI coding agent, cutting the token volume of an incident log by up to 95% before it reaches the model. Benchmarked against Drain3 and raw-log baselines on the LogHub-2.0 dataset, it works with Claude, GPT, Gemini, and local models alike.
About Codag
Codag Inc. is a Y Combinator company from the 2026 Summer batch, based in San Francisco. It was built to solve a specific failure mode: when an AI agent has to diagnose a production incident, the raw log file is too large to hand the agent directly, and summarizing it risks losing the one line that actually explains what happened. Codag sits between the log source and the agent, compressing that file down while keeping only the lines that matter and citing exactly where each one came from, without changing the underlying content.
The typical user is a backend developer or SRE who already runs an AI coding agent, such as Claude Code or Codex, against production infrastructure. Instead of pasting a raw dump from a failing Kubernetes pod, an AWS CloudWatch stream, or a Vercel build log straight into the agent's context window, they route it through Codag first. Teams already running Datadog, Sentry, or journalctl pipelines can drop Codag in as a wrapper around the same commands they already use.
The CLI, written in Go, and the codag-drain log-templating library, written in Rust, are both MIT-licensed and published on GitHub under the codag-megalith organization. A separate VS Code extension, codag-visualizer, has passed 500 GitHub stars and adds live workflow-diagram visualization inside the editor.
Pricing
Free: 50 MB of compressed logs per month, no credit card required. Hobby: $19/month with expanded limits. Team: $199/month plus usage-based compute charges on top of the base fee.
Pricing is metered by compute consumed rather than seats. The open-source CLI and codag-drain library run locally at no cost with no usage cap.
| Tier | Monthly price | What it includes |
|---|---|---|
| Free | Free | |
| Hobby | $19/mo | |
| Team | $199/mo |
Key Features
- 8,021x Log Compression: Reduces 1.2 million raw log lines to 3,300 schema-valid JSON tokens, cutting typical per-incident token cost by up to 95%.
- Role-Tagged Evidence: Tags every retained log line as root_cause, trigger, or consequence, turning a flat dump into a causal map an agent can follow straight to the failure.
- MCP Server and CLI: A single `codag setup` command installs the CLI and registers Codag's MCP server with Claude Code, Codex, and Cursor, with no manual config file needed.
- Line-Number Citations: Every compressed line includes a pointer back to its exact position in the raw file, so an agent can double-check evidence instead of inventing context that was never there.
- PII Redaction: Strips personally identifiable information out during preprocessing, before any log content leaves the local system, whether you use the hosted service or run everything offline.
- Multi-Format Log Support: Parses JSON logs, syslog, Kubernetes, Hadoop, Spark, HDFS, and unstructured application logs with no format configuration required.
Pros
- The cost math favors Codag heavily for any team already paying per-token for a coding agent's context: it strips out the vast majority of a log's volume before it ever reaches the model.
- MIT-licensed CLI and drain library mean the compression logic itself is auditable and portable, not a black box you depend on a vendor to keep running.
- Agent-agnostic: because output is just schema-valid JSON, switching the underlying model between Claude, GPT, Gemini, or a local model requires no change to the log pipeline.
Cons
- The free tier is capped tightly enough that an active team can hit a paid plan within their first handful of incidents.
- No SOC 2, HIPAA, or other compliance certification exists yet, since Codag is a one-person, newly founded company; that rules it out for regulated-industry procurement for now.
- There is no web dashboard: everything is CLI and MCP-first, so a team without terminal or MCP-configuration comfort will find onboarding harder than a typical SaaS tool.
Data Handling
- Training-data policy
- PII is redacted during preprocessing before any log content is transmitted to the hosted service. Open-source CLI processes logs locally with no data sent to the server.
Frequently Asked Questions
How much do you pay for Codag?
Codag has three tiers priced by compute rather than seats: a free tier for light use, Hobby at $19 a month, and Team at $199 a month plus usage-based compute charges on top. There are no per-seat fees, so the Team plan scales cost-effectively for larger engineering teams. Overage pricing on the Team plan is not published, so high-volume teams should contact Codag directly for an estimate.
What do you get on Codag's free tier?
Codag's free tier covers 50 MB of compressed logs a month, plenty for a handful of incident investigations before Hobby's $19 tier is worth it, with no credit card needed up front. The open-source CLI and codag-drain library carry no cap at all since they run locally on your own machine rather than through Codag's hosted service. The codag-visualizer VS Code extension is included free as well.
Which tools compete with Codag in 2026?
Compresr, also a 2026 Y Combinator company, is the closest direct alternative, focused on boosting an agent's usable context rather than tagging causal structure inside the log. For broader AI agent observability, teams often pair raw log access with LangSmith or Arize Phoenix for tracing, though neither compresses the log before it reaches the model. Drain3, the open-source parser codag-drain is modeled on, handles log templating but does not output schema-valid JSON or causal role tags.
Codag or Archal: which should you pick?
Archal works upstream, catching agent failures inside a sandboxed pre-production test environment before they ever reach real users. Codag works downstream: once an incident has already happened, it compresses the raw log evidence into something an AI agent can actually read and diagnose, at the extreme compression ratio Codag is built around. Teams that want both often run Archal for pre-release testing and Codag for post-incident triage rather than picking one over the other.
How do you set up Codag?
Run the install script from codag.ai to download the codag-cli Go binary for macOS, Windows, or Linux, then run `codag setup` to authenticate and automatically register Codag's MCP server with Claude Code or Cursor. From there, wrap any existing log command, for example `codag wrap docker logs my-container`, or use the built-in AWS CloudWatch and Vercel log-fetching tools. The free tier activates immediately with no credit card required.
Top Alternatives
- Archal: Archal fits a team that wants failures caught before code ships, running validation checks against agent behavior pre-release. Codag only activates after something has already broken, wrapping the log commands already in daily use and handing back a compressed, schema-valid file instead of a raw dump.
- Openlayer: Openlayer covers broad AI governance and evaluation across a live production system. Codag stays narrow: it exists for the moment an incident already happened and a huge log needs compressing down to something an agent can read.
- ReasonBlocks: ReasonBlocks catches and prevents agent failures mid-run, stopping problems before they compound. Codag steps in after the fact, compressing the resulting log evidence so a debugging agent can diagnose what already went wrong.
HokAI guides covering Codag
- Codag vs Crukx: Why Only One of Them Ships in 2026: Crukx quietly left observability behind for a waitlist-only regression gate in 2026. Here's what Codag, Archal and Maitai actually cost and ship right now.
- ReasonBlocks vs Touchmark: Which One Actually Fixes Your AI Bill?: ReasonBlocks cuts wasted AI agent tokens by up to 52%. Touchmark, after its August 2026 pivot, sells forward contracts locking in token prices instead.
- The Three Jobs Hiding Inside "AI Agent Observability" in 2026: AI agent observability tools split into three jobs in 2026: tracing, mid-run intervention, and outcome-based billing. Here is how to tell which one you need.