Trackers / Coding agent scoreboard

Open-Source Coding Agents: China vs the West

The coding agents shipped by Chinese and Western AI labs, with the license each repository actually declares and the star, fork and open-issue counts behind the headline, read straight from the GitHub API.

As of 6 September 2026. All counts pulled from the GitHub API at 6 September 2026, 18:08 UTC. Machine-readable version:agent-frameworks.json.

On 13 August 2026 DeepSeek published a coding agent called Harness under the MIT license. Four days later it had more GitHub stars than any comparable tool from any Western lab, including one that had been accumulating them since February 2025. That is a real and checkable fact, and it is also the kind of fact that is easy to over-read. This page puts the number next to the other numbers, for every serious coding agent in the category, pulled in one pass so the rows are comparable.

One caveat governs the whole table. A GitHub star is a bookmark, not a user. It is one click from one account, it is lagging, and it is gameable. It does not count installs, daily active users, or a single line of code shipped in production. We publish forks and open issues alongside it because both are closer to hands-on use, forking being what you do to modify something and filing an issue being what you do when you run it and it breaks. Read the three together, and treat none of them as adoption.

The scoreboard

Coding agents with a public repository, by stars, highest first. Counts were read from the GitHub REST API in a single pass at 6 September 2026, 18:08 UTC, so they are comparable with each other and will have drifted since. License is the SPDX identifier the repository itself declares. Last push is the date of the last push to any branch, a staleness signal and not a release date. Open issues + PRs is a single combined count, because that is the only form GitHub publishes it in.

Scroll the table sideways to see every column.

How to read this

The license column is the most durable thing here. Star counts move hourly; a license is a decision. 8 of the 9 repositories in this table declare a permissive open-source license, MIT or Apache-2.0. The exception is anthropics/claude-code, for which the GitHub API returns no license at all. That is a fact about the repository rather than a claim about how the product is sold, but it is the cleanest available line between the projects you can fork and ship and the ones you cannot. All four Chinese entries are permissive: DeepSeek and ByteDance chose MIT, Alibaba and Moonshot chose Apache-2.0.

Chinese labs now ship the agent, not just the model. 4 of the 9 agents here come from Chinese labs: DeepSeek, Alibaba, ByteDance and Moonshot. That is a step up the stack. The established pattern in Chinese AI has been to open-weight the model and let someone else build the tooling, which is the story our frontier model tracker follows. An agent is the layer where developer habits actually form, and giving it away under MIT is the same standards-over-margin play, moved one level closer to the user.

The quiet repositories are the ones the star column hides. 4 of the 9 repositories here pushed code on the day of this pull. Two did not come close. Aider’s last push to any branch was 107 days earlier, and Trae Agent’s was 213 days earlier, its default branch last moving on 5 Feb 2026 while it continued to hold 12,073 stars. A star count keeps whatever it has already collected whether or not anyone is still writing the software, which is the failure mode this column exists to catch. It is a signal and not a verdict: a push date says where public work is happening, and a project can be maintained somewhere the API cannot see, so we report the date and stop there.

Age is the missing column in most coverage of this. Aider’s repository has been public since May 2023 and has 48,791 stars. Claude Code has been public since February 2025 and has 144,249. Harness has been public since 13 August 2026 and has 213,903. Whatever explains the third number, it is not the same process that produced the first two, and a table that lists stars without listing repository age invites exactly the wrong inference.

What the Harness number does and does not show

DeepSeek Harness passing every Western coding agent on stars within four days is the headline, and it is accurate. Three details sat underneath it on the first pull of 17 August 2026, all readable from the same API, and all of them cut against treating the number as evidence of use. This page promised to re-pull them. On the fourth pull, 6 September 2026, two of the three still hold and the third has now moved the other way three times running.

The public issue tracker is still switched off. The GitHub API reports has_issues: false for deepseek-ai/deepseek-harness. Its zero open threads therefore means "no public bug tracker", not "no known bugs". Every other repository in this table accepts public issues, and the two largest Western agents carry 13,387 and 15,574 open issues and pull requests respectively. The usual public evidence that a lot of people are running the software and hitting its edges is simply not being collected here.

The software is still pre-1.0. The most recent tagged release at the time of this snapshot was dsh-v0.1.3-alpha.1, published on 4 September 2026 and flagged by the repository as a prerelease. The tag has moved twice since the last pull, through dsh-v0.1.2-alpha.5 and dsh-v0.1.2-rc.1, but it has moved sideways rather than up: every release tagged since the first snapshot is a prerelease and none of them is a 1.0. That is a version number that asks for testers, not one that claims production readiness.

Forks no longer trail in the way they did. Harness sits at about 11.8 percent forks-to-stars, still the lowest ratio among the four largest projects in the table, against roughly 16 percent for Claude Code and 15.4 percent for Codex. Bookmarking is one click; forking is what you do when you intend to change the code. On the first snapshot that ratio was about 10.2 percent, and it has risen at every re-pull since, to about 11.1, then 11.5, and now 11.8, because forks have grown faster than stars in all three windows: 37 against 26 percent to 23 August, 11.2 against 7.6 percent to 28 August, then 8.6 against 6.0 percent to 6 September 2026. Harness took the highest absolute fork count in the table on the previous pull and has widened that lead since, to 25,166 against 23,044. Three consecutive windows is a series rather than a blip, though the level still has not changed rank, so this remains a direction rather than a verdict.

None of that makes the tool bad, and none of it is evidence of manipulation, which we have not seen established and do not allege. It does mean the honest reading of 213,903 stars in 24 days is that a very large number of developers noticed a DeepSeek release, which is itself a real finding about DeepSeek’s pull. Whether they stay was the open question, and three re-pulls give a partial answer: the attention has not evaporated, and the one measure here that tracks intent to modify code has grown faster than the one that tracks noticing in every window so far. We will keep pulling it.

Common questions

Sources and method

Every count on this page is a primary reading from the GitHub REST API endpoint api.github.com/repos/<owner>/<name>, taken for all 9 repositories in a single pass at 6 September 2026, 18:08 UTC (2026-09-06T18:08:02Z). Pulling them together matters: these counters move continuously, and rows collected hours apart are not comparable. Nothing on this page is filled in from memory or rounded.

Six handling rules keep the table honest. Stars are labelled as attention, never as adoption, in the prose, the table caption and the JSON payload. License is the SPDX identifier the repository declares via the API’s license.spdx_id; "None declared" means the API returned no license object, which describes the repository and not the commercial terms of the product. Thread counts are omitted, not zeroed, where a repository has disabled public issues, because a zero would misread as an absence of bugs. Repository creation dates are floors, not birthdays: a project may have existed privately before its repository went public, so the column is labelled "repo created" rather than "released". Last push is a staleness signal, not a verdict: pushed_at records the last push to any branch of the public repository, and a quiet repository may still be developed in a private fork or shipped through a separate channel, so we publish the date and do not call any project abandoned. Where an old date carries weight in the prose we corroborate it against the default branch's most recent commit first.

The sixth rule is a correction. From 17 to 28 August 2026 this page labelled one column "Public issues" and reported the GitHub API's open_issues_count under it. That field counts open issues and open pull requests together, and the API does not expose the two separately on this endpoint, so the label overstated the issue count on every repository carrying a lot of open PRs. The column is now labelled "Open issues + PRs", which is what the number has always been. We check the size of the error against GitHub's search API at every pull rather than assuming it is small: on this one, anthropics/claude-code's 13,387 splits into 12,682 issues and 705 pull requests, and Aider-AI/aider's 1,853 into 1,368 and 485, both adding up exactly. That split is deliberately not published. The search API is a second index that lags this one, and rows drawn from two indexes would break the single-pass comparability the rest of the table depends on.

Inclusion is limited to coding agents with a public repository and meaningful traction, covering the four Chinese lab entries, the three Western lab entries, and the two largest independent projects. It is not a complete census of the category, which has dozens of smaller entries. Where a row links to a CAD issue, that issue carries the underlying reporting.

This is a living document. The counts will be refreshed from the daily newsletter as the story moves, and the pull timestamp above always says how fresh the current numbers are.