# Codag

Aiolum story: https://aiolum.com/story/codag/ · added 2026-11-12 · updated 2026-11-12

**In short.** Codag (Tool-output compression for coding agents, United States): Self-serve paid plans live, November 2026. Evidence grade E: no revenue figure was checked. Team: 1 founder; Y Combinator lists a team size of 1 in November 2026. Launched November 2025; still running.

Tool-output compression for coding agents · United States · AI-native · solo founder · still running · group: Developer tools and coding agents

- Outcome: Self-serve paid plans live, November 2026 (evidence grade of the headline number: E; no revenue figure)
- Launch: November 2025; months from launch to the outcome: 12
- Team: 1 founder; Y Combinator lists a team size of 1 in November 2026
- Outside funding: Y Combinator Summer 2026; no round amount found
- Main factor: Domain expertise
- Meters, 1 to 4: evidence strength 1, copyable pattern 2, capital needed 2
- Tags: AI-native, Solo founder, Funded, Running, B2B, Developer tools

Links:

- Visit codag.ai: https://codag.ai/

**The product.** A service that shrinks the output of tools such as test runs, searches and log tails before a coding agent reads it. The agent keeps its own model account. Codag sends a model-bound request through a local proxy on the engineer's machine straight to the model provider. The company says the omitted lines stay on the machine and can be fetched back exactly.

**Why it worked.** Codag sells savings on the tokens a coding agent spends reading tool output. The founder's profile lists infrastructure work at Okta and Shopify, and his launch post says he met the log problem there. The company's pricing page charges US$0.05 a credit, and the page says a credit covers about 100 MB of log output. The founder says the product shrinks tool results by 75% on average.

Evidence grades: A = disclosed by a public company, a filing or a payment processor, or verified by one; B = two or more independent reports agree, or a journalist confirmed it; C = a founder claim that no one else verified; D = reported, not verified: one founder claim or a third-party estimate with no second source; E = no revenue figure found, so the headline is a non-money fact and the other facts carry their own grades; calc = computed on this page from graded facts.

## Summary

Codag is a gateway for coding agents, programs such as Claude Code and Codex that write and test code by calling tools. It cuts down the tool output, the text that test runs, searches and logs return, before the agent reads it. A model token is the unit a provider bills for, so less text to read means a smaller bill. Codag also reports spend by tool and by agent.

Michael Zhou is listed by Y Combinator as the solo founder and CEO of Codag. The oldest product on the same domain is an open-source viewer of AI workflows, whose code repository was created on 14 November 2025. The company's log-compression launch post is dated 23 May 2026. Y Combinator lists the company with a team size of 1 in the Summer 2026 batch.

No revenue figure was found, so the story is grade E. The headline is a fact about the offer: a price list, billing terms and a free start with no card. The founder wrote in a Y Combinator launch post that more than 90 organizations use Codag. That is his own statement, and no source names a customer.

## Why did Codag work?

### It shrinks large tool results before the agent reads them

The company says each result is classified by kind and large ones are cut to failures, matches and signal lines. The home page shows a test run of 4,212 lines reduced to a few lines. Omitted bytes stay encrypted on the engineer's machine. These are company claims.

### It leaves the customer's model account and traffic alone

The privacy policy says model traffic goes from a local proxy to the customer's own provider. It never passes through Codag's cloud. The terms bar reselling or proxying the Codag API. No page read here offers or pools a provider's access.

### It bills by volume reduced and fails open

The pricing page says optimization pauses when credits run out and the agent gets the original output. The company says log compaction uses no model, while agent-context reduction runs a frontier model and costs far more per byte.

## How it was built

These details come from the company's privacy policy, documentation and open-source page. They are grade C unless stated.

- The privacy policy names Modal for the managed infrastructure on which Codag runs its own inference.
- The same policy names Railway for the API and database, and Vercel for the web console.
- It names Stripe for payments, Google OAuth for sign-in, Resend for email and PostHog for analytics.
- The docs say setup installs a per-user local service on macOS, Linux or Windows and attaches Claude Code and Codex.
- The open-source page lists an SDK in Python, TypeScript and Go, and codag-drain, a Rust log grouper under the MIT license.

Sources: [3, 16, 11]

## Who tried this before Codag, and why did they fail?

Four companies in our database of failed startups built developer tools and closed. The values below are from that database, were read on 12 November 2026 and are not independently verified.

The failed startups come from our database of dead startups, retold here in our own words; Its entries are AI-assisted summaries of public sources, so they carry no evidence grade.

### CodeSee (2020 to 2024; funding US$10.0M; cause: Business model)

- Promise: Automatic visual maps of code architecture and dependencies, to help engineers understand large codebases.
- Why it died: The record says developers liked the diagrams in demos but could not state the return on investment to their managers.
- Record: 

### Fig DevTool (2020 to 2024; funding US$2.4M; cause: Monetization)

- Promise: Autocomplete and visual command builders for the terminal, for more than 300 command-line tools.
- Why it died: The record names free alternatives and shell scripts that did similar work, and a weak willingness to pay for terminal tools.
- Record: 

### Kite DevTool (2014 to 2024; funding US$21.0M; cause: Competition)

- Promise: AI code completion inside the editor, with a free tier and a paid Pro version.
- Why it died: The record says GitHub Copilot, launched in June 2021, gave a large step up in suggestion quality.
- Record: 

### Sider (2012 to 2024; funding US$7.0M; cause: Competition)

- Promise: Automated code review and static analysis for software teams.
- Why it died: The record says free open-source tools and larger security platforms covered the same need.
- Record: 

## What changed

The four records share one pattern: a developer tool that engineers liked but that was hard to charge for. CodeSee and Fig named weak willingness to pay. Kite and Sider named larger rivals that bundled the same work. Codag's first product, a free viewer of AI workflows in code, is close to CodeSee's code diagrams.

Codag's 2026 product charges by the credit for a cost that the customer already pays in tokens. The pricing page says each credit is meant to return more than it costs. The sources do not show how many customers pay. Model makers could also add tool-output reduction to their own agents.

## Can one person copy Codag now?

**Verdict: The pricing idea can be copied; the results are not shown.**

A small team can sell a tool that lowers a bill the customer already pays, and charge a small share of the saving. The sources show a price list, billing terms and a free start. They show no revenue and no named customer. A reader should treat this as a pattern, not as proof that it earns.

1. Pick one cost that engineers already see in a bill, such as tokens spent on tool output.
2. Keep the customer's own provider account, so no model cost lands on the company.
3. Give a free start with no card, then sell credits in small packs.
4. Fail open: when credits run out, pass the original data through.

**What is still open.** Other agent costs that repeat in every session, such as file reads, web fetches and long conversation histories.

## How much revenue does Codag make? The numbers, graded

A solo founder's company that moved from a free viewer to a paid cost tool in 2026, with a price list and no revenue figure

| Fact | Value | Grade | Note | Sources |
|---|---|---|---|---|
| What is sold | Reduction of coding-agent tool output, billed by the credit | C | The home page says Codag cuts what coding agents cost by teaching them to read large tool results efficiently. The company's pricing page lists the credit price and three plans. | [2, 1] |
| Start | 14 November 2025 | B | The code repository of the earlier Codag product, a VS Code viewer for AI workflows, shows a creation date of 14 November 2025. Y Combinator's page says Founded: 2026. The earlier date is used. | [12, 10] |
| Earlier product, same name | Open-source viewer of AI workflows, listed with 677 stars | B | An Internet Archive copy of codag.ai dated 5 February 2026 offers this free VS Code extension. The 2026 product has a different purpose and is paid. | [13, 12] |
| Company | Codag Inc., San Francisco | C | The footer of every page reads Codag Inc. and San Francisco. The terms of service, last updated in August 2026, say California law governs. | [1, 4] |
| Founder | Michael Zhou, founder and CEO | C | Y Combinator's page lists him as the active founder. The vision page and the launch post of 23 May 2026 are signed by Michael Zhou, Founder, Codag. | [10, 7, 8] |
| Team size | 1 per Y Combinator | C | Y Combinator lists Team Size: 1 and the founder's profile reads Solo. The support page says the company is a small team. No headcount was published. | [10, 14] |
| Accelerator | Y Combinator, Summer 2026 | C | The Y Combinator page lists the batch and the status Active. No funding amount or investor was found in a page read here. | [10] |
| Prices | US$0.05 a credit; Team US$499 a month plus usage | C | The company's pricing page lists these prices and states that Pay as you go starts at US$0. | [1] |
| How a plan is bought | Start free; packs bought by card; Stripe takes payments | C | The pricing page says No card until you buy and lists credit packs. The terms say paid plans renew monthly. The privacy policy names Stripe for payment processing. | [1, 4, 3] |
| Where model traffic goes | Local proxy, then the customer's own provider | C | The docs say provider credentials are never sent to or stored by Codag. The home page says the customer keeps the provider account. The terms bar reselling the Codag API. | [6, 2, 4] |
| Users | More than 90 organizations, teams of up to 35 people | D | The founder wrote this in the Y Combinator launch post dated 16 August 2026. No customer is named and no second source was found. | [10] |
| Size of reduction | 75% smaller on average, and 20-30% less token spend | D | The founder claims the first figure in the launch post dated 16 August 2026. The company description on Y Combinator's page states the second. Neither figure was tested here. The pricing page says dollar savings are estimates. | [10, 1] |

How to read this table. Grade C means the company or a directory states the value. Grade D is a single founder statement with one date. Grade B means public code shows it. No value here is grade A. No value is calculated on this page.

## Pricing

Codag prices read from the company's pricing page on 12 November 2026 (grade C)

| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Pay as you go |  | 0 base, then 0.05 a credit | 10 free credits, about 1 GB of logs; packs of about 1 GB for 0.50, 10 GB for 5, 50 GB for 25 |
| Team |  | 499 plus usage | 12,500 credits a month, about 1.25 TB; overage at 0.50 a GB; per-member usage attribution; 24-hour support SLA |
| Enterprise |  | custom, annual contract | SSO, VPC peering, client version pinning, custom SLAs; book a demo |

Pay as you go is billed by the credit. The company says one credit covers about 100 MB of log output, and its agent-context reduction uses credits about 1,000 times faster. Team adds a monthly bundle. Enterprise has no published price. The first 10 credits are free, with no card.

## When did Codag launch, and how did it grow? Timeline

From an open-source workflow viewer in November 2025 to a priced cost tool for coding agents in 2026.

- **14 November 2025** (grade B [12]): The repository of the first Codag product, a free viewer of AI workflows for VS Code, is created.
- **5 February 2026** (grade C [13]): An archived copy of codag.ai describes the open-source viewer and offers it on the VS Code Marketplace.
- **23 May 2026** (grade C [8, 9]): The founder launches Codag as a log compressor for coding agents, free to start. The company later marked the post as history.
- **13 August 2026** (grade C [7]): The vision page is revised. It says the company serves agents first, and it calls the agent a reader that pays by the token.
- **16 August 2026** (grade C [10]): The Y Combinator launch post appears. The founder says that more than 90 organizations use Codag.
- **12 November 2026** (grade C [1]): The pricing page lists Pay as you go at US$0.05 a credit and Team at US$499 a month.

## What could go wrong for Codag?

- **What is not verified.** No revenue figure was found, so the story is grade E. The pricing page, home page, terms, privacy policy, documentation, vision page, blog and Y Combinator page were read. A reader could look for a press interview or a payment-processor statement. No source states that a customer has paid. The founder's statement of more than 90 organizations has no second source. The support page says Codag is sold to teams through a demo and a 14-day pilot. The pricing page offers a free start and credit packs without a demo.
- **The AI part of the product is partly model-based.** The pricing page says log compaction is deterministic and uses no model. Only agent-context reduction runs a frontier model. The product is aimed at AI coding agents, and the company calls itself an AI gateway on Y Combinator's page. A reader may judge AI to be a smaller part than the label suggests.
- **The product changed a lot in nine months.** The first Codag was a free workflow viewer. The company's launch post of 23 May 2026 describes a log compressor, and the company marks that post as outdated. Commands, quotas and prices in it no longer apply. The open-source page shows 500+ stars for the viewer, and the repository showed 677 on the day it was read.
- **Savings figures are the founder's claims.** The founder claims 75% smaller results on average, and Y Combinator's profile says 20-30% less token spend. No test was read. The pricing page says dollar savings in the product are estimates from public model prices. The home page shows example outputs, not customer data.

## Sources

1. Codag pricing page — https://codag.ai/pricing (grade C). Read in full on 12 November 2026. Source for the credit price, the plans, the free credits, the packs and the Book a demo buttons.
2. Codag home page — https://codag.ai/ (grade C). Read in full on 12 November 2026. Source for the product description, the example outputs, the local proxy and the FAQ.
3. Codag privacy policy, last updated August 2026 — https://codag.ai/privacy (grade C). Read in full on 12 November 2026. Source for the sub-processors, the data flow and the company name.
4. Codag terms of service, last updated August 2026 — https://codag.ai/terms (grade C). Read in full on 12 November 2026. Source for monthly billing, the bar on reselling the API and the governing law.
5. Codag documentation: Supported agents — https://codag.ai/docs/agent-setup (grade C). Read in full on 12 November 2026. Source for the local proxy, the supported agents and the fail-open behavior.
6. Codag documentation: Privacy and data flow — https://codag.ai/docs/privacy-data-flow (grade C). Read in full on 12 November 2026. Source for the statement that provider credentials are never sent to Codag.
7. Codag vision page, May 2026, revised 13 August 2026 — https://codag.ai/vision (grade C). Read in full on 12 November 2026. Signed by Michael Zhou, Founder, Codag.
8. Codag blog, 23 May 2026, by Michael Zhou: Hello Codag — https://codag.ai/blog/hello-from-codag (grade C). Read in full on 12 November 2026. Founder's launch post for the log compressor, marked as history by the company.
9. Codag blog index — https://codag.ai/blog (grade C). Read in full on 12 November 2026. Source for the date of the launch post and the company's note on it.
10. Y Combinator company page: Codag — https://www.ycombinator.com/companies/codag (grade C). Read in full on 12 November 2026. Source for the founder, the batch, the year, the team size, the founder's launch post dated 16 August 2026 and its claims.
11. Codag open-source page — https://codag.ai/open-source (grade C). Read in full on 12 November 2026. Source for the SDK, codag-drain and the viewer with 500+ stars.
12. GitHub repository codag-megalith/codag-visualizer — https://github.com/codag-megalith/codag-visualizer (grade B). Read the repository page and GitHub's public data record of the repository on 12 November 2026. Source for the creation date of 14 November 2025, the MIT license and 677 stars.
13. Internet Archive copy of codag.ai, 5 February 2026 — https://web.archive.org/web/20260205114439/https://codag.ai/ (grade C). Read in full. Source for the earlier product, the open-source viewer.
14. Codag support page — https://codag.ai/support (grade C). Read in full on 12 November 2026. Source for the statement that the company is a small team.
15. Database of failed startups, local export — /method/#dead (grade C). Record pages read on the website on 12 November 2026. AI-written summaries, not independently verified. Source for the failed predecessors.
16. Codag documentation: Quickstart — https://codag.ai/docs (grade C). Read in full on 12 November 2026. Source for the setup steps and the systems supported.

Contact: aiolum@livyden.com. Run Codag? Send corrections or a verified figure to aiolum@livyden.com (subject: Correction: Codag).