One codebase for 12 coding agents: oar's 490 stars, with the biggest wave from DeepSeek
What it is An npm library that lets people building agent workbenches and dispatch tools use one interface to start sessions, receive events, interject, resume, and spawn sub-agents, without writing a separate integration for each coding agent.
Why it took off The original author of Kimi CLI and several RisingWave engineers posted on Chinese developer X, opening with "I don't like ACP" to spark debate. In Oct 28, DeepSeek Harness used it as the base for subagents, which someone spotted in the code and spread. Then the launch of the team's new product Kanso brought another wave.
CHow you could follow Don't build another TypeScript unification layer. First have an agent product that's really in use, then open-source the foundation, filling what oar doesn't do (other languages, one process managing multiple sessions, goal support), and proactively send PRs to a growing agent project to get embedded.
- Data (updated daily) · Repo created 2026-08-05 · 493 stars · 7 days +128
GitHub · stars Data 2026-10-09 → 2026-11-08 (31 days, 31 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
botiverse/oar is a TypeScript library that uses one interface to drive 12 coding agents, including Claude Code, Codex, Cursor, Kimi Code, Qwen Code and DeepSeek Harness. It went public in late September and reached 6 490 stars by November. Each of the three star surges lines up with an event: the team posted on X, DeepSeek's official agent tool used it as a core dependency, and the team's new product was labeled "Powered by oar." Its reputation has only spread within Chinese developer X circles, and the star count isn't large. What's worth watching is how it earned a chance to be embedded in a big project. If you want to copy that path, first you need a big project that wants to use you.
Someone found it in DeepSeek's code
In 2026 Oct 28, at 1pm Beijing time, Chinese developer @jakevin (卡比卡比) posted:
"I just found OAR in Deepseek Harness!! The DSH code is actually trying to use OAR to build subagents"
He was referring to DeepSeek's official coding agent tool, DeepSeek Harness (DSH, about 270k stars on GitHub). Its v0.2.2-alpha.2 release that day added an experimental feature called subagent_oar, which directly depends on @botiverse/oar: it lets DSH hand subtasks to other companies' agents such as Claude Code and Codex, and oar is the connector. The post has 31k views (X, 11-06). Afterward, @teortaxesTex (about 80k followers), a major DeepSeek fan account, tried the feature, and QiNiu Cloud, Juejin and several AI daily newsletters also wrote about it.
The repo gained over 100 stars in those two days: Trendshift's daily table shows +126 on 10-28; by GitHub's daily star counts, it went from 257 to 343 between 10-28 and 10-29. In the full month before that, it gained only 0 to 6 stars a day.
Some replies threw cold water. @rya_nooo said: "If you actually use it, you'll find it's just as painful as acp... no goal support; codex app is an app-server that can manage multiple sessions, but acp/oar is one-to-one, and resource usage rises noticeably."
What it is
An npm package under the Apache-2.0 license. People building agent workbenches, remote controls and multi-agent dispatchers used to write a separate integration for each CLI. Now one oar interface covers opening sessions, receiving the event stream, steering mid-run, queuing, aborting, resuming sessions, sub-agents, passing the host's MCP tools to the agent, and even installation, login, usage and model listing. It supports 12 agents, with Chinese ones especially well covered: Kimi, Qwen, Tencent CodeBuddy, Alibaba Qoder, and DSH.
It was split out of Raft. Raft is a multi-agent collaboration workbench built by the botiverse company, originally named Slock, and oar was always its foundation. The team's CEO stdrc is the original author of Kimi CLI, and the main developer xxchan and contributor TennyZhuang are both engineers who came from RisingWave, a database company.
oar has been developed very quickly: 56 days saw 527 PRs merged (Trendshift), over a dozen releases in a week, and it reached v0.92.0 by 11-06. The second-most-prolific contributor is a bot called stamp-build[bot], with 223 commits. Raft's other product is called Stamp, with the slogan "let agents ship code under their own names," so most of this library's code was probably written by agents (inferred).
"A library, not a protocol"
Plenty of people are working on the same problem. The best known is ACP (Agent Client Protocol), led by the Zed editor: it defines a protocol for conversations between editors and agents, which each agent then implements. oar's launch post was aimed squarely at it. On Sep 30, xxchan wrote:
"Don't like ACP, but want to call various agents (Claude Code, Codex, Kimi Code, Pi, ...)? Try oar... The idea is similar to @xuanwo's OpenDAL. Using one protocol layer (S3-compatible) to connect to many different services always runs into lots of pitfalls, so it's better to build a library and do more with it."
The next day TennyZhuang added: "ACP takes the intersection of all harnesses' capabilities, which means it also drops the stronger features of SOTA harnesses." The point: a protocol can only use the small slice everyone supports, while a library can adapt to each vendor's native interface, so extra features from any vendor are usable, and anything a vendor can't do is reported up front when the session opens, not silently dropped.
Opening with "ACP, which everyone dislikes" got the discussion going, and the replies quickly turned into a debate over "library vs protocol." This was the first wave: 10-01 gained 126 stars and ranked #23 on Trendshift's TypeScript chart.
Where the stars came from
| Date | Stars gained that day | What happened |
|---|---|---|
| 9-30 ~ 10-01 | +10 / +126 | xxchan and stdrc posted in Chinese. stdrc's post read: "Oar, extracted from the Raft foundation, welcome to try it." |
| 10-02 ~ 10-27 | 0–6 per day | No new events, almost no growth, flat at 207 |
| 10-28 ~ 10-29 | +126 / +25 | DSH embeds oar, discovered by @jakevin and spread |
| 11-02 ~ 11-04 | +11 / +73 / +32 | stdrc launched their new product Kanso, with the post ending "Powered by github.com/botiverse/oar" |
| 11-05 ~ 11-06 | +5 / +3 | Decline |
(Daily star counts from Trendshift and GitHub, 11-06)
Kanso is a Mac disk cleaner. You hand it to your own coding agent, which finds out what's eating your space, then deletes only what you confirm. One scan takes about 10 minutes and runs on your own agent quota. The launch tweet got 124 likes and 9,700 views, and over the next two days oar picked up more than 100 stars.
All three waves started with tweets from these few people, plus one endorsement from DeepSeek. The English-speaking world barely picked it up: the Hacker News launch post hit 1 points and 0 comments, it never appeared on GitHub Trending, and we found no sign of paid acquisition or star farming.
Real usage is far bigger than the star count. @botiverse/oar got 160k downloads on npm in the week of 10-30 to 11-05, and 295k in the 30 days after that (official npm stats). Raft's own background processes, two DSH subagent packages and one third-party package depend on it. The number includes inflation from CI reinstalls, mirror syncs and several releases a day, so treat it as an order of magnitude (estimate).
Who else is doing the same thing
| Project | Scale (early November) | vs. oar |
|---|---|---|
| ACP (led by Zed) | Main repo 4,583 stars, claude-agent-acp 2,730 stars | A protocol each vendor has to implement itself, so you only get the shared subset of features |
| HarnessRouter / UHP | 4,259 stars, +450 in the last 7 days | Almost the same positioning, but built as an HTTP service on an open standard, with a paid hosted tier. It has more than 8 times oar's stars |
| coder/agentapi | 1,496 stars, no longer maintained | Wraps a terminal as an HTTP API. Doesn't handle native events or subagents |
| Claude Agent SDK / Codex SDK | claude-agent-sdk-typescript, 1,861 stars | Official, but only for its own agent |
| vercel-labs/durable-harness | 293 stars (created 10-16) | Leans toward the Vercel ecosystem, supports few agents |
| MiniMax-AI/OpenAgentCore | 319 stars (created 9-21) | A self-hosted service, not a library |
| cocofhu/dsh-aio-agent | 7 likes in the DSH forum | A DSH plugin that uses ACP to connect 6 agents. Head-on competition with subagent_oar |
(Star counts from GitHub, 11-06)
oar's own repo contains earlier research that goes through Synara, Paseo, Lody, Orca, Multica, T3 Code and others one by one. The team clearly knows this space is crowded. On stars it isn't ahead. Its edge is being picked by DeepSeek. Whether DSH chose it for the technology or through personal connections, we couldn't find out. Also, oar has a DSH runtime inside it, so the two are wired into each other.
If it were you
Bottom line: There's something to learn from this space, but building another "one interface for every agent" library isn't a good bet. oar's stars come from the team's reputation and DeepSeek embedding it. Followers don't get that. Similar projects already exist, such as HarnessRouter at 4k stars and dozens of workbenches. Outside the three weeks with event-driven pushes, oar gained only 0 to 6 stars a day.
Who it suits: Teams already building an agent workbench, remote control or scheduler, with a product people actually use, who can open-source the core while they're at it. Or people who can talk to the maintainers of a major agent project.
Who it doesn't suit: Solo developers who want to write yet another "unified agent layer" from scratch to collect stars. The users of this category are developers who build tools, there aren't many of them, and whichever one a big project picks is the one that wins.
How to do it:
- Have the product first, then open-source the core. oar only came out after it ran in Raft, and each agent comes with a documented live test. That is why DeepSeek could use it.
- Build the part oar hasn't done well, not the part it has already done. Gaps the replies pointed out: no goal support, one process per session, and no way to manage multiple sessions in one process the way Codex app-server does. It's also TypeScript only, so the equivalent libraries for Python, Go and Rust are still open (inferred; oar has said it plans multi-language support).
- Pick a growing agent project, treat its plugin or subagent feature as your first user, and send it a PR directly. Getting embedded once is worth more than posting ten tweets.
- One page per agent called "native protocol live test", re-tested every time the agent updates. These docs are the thing most likely to earn trust, and the hardest to copy.
How to get users: Launch with a comparison against an approach people already dislike ("if you don't like ACP, try this"). That gets arguments going. Chinese-speaking developers on X drive more traffic than HN. Add a "Powered by" line on every new product you ship. And let others "discover" you in their big-project code. That's more convincing than announcing yourself.
When the window closes: Agent CLIs from every vendor are shipping one after another, so demand for a unifying layer is growing. Once ACP or some standard settles, or a big company ships its own cross-agent SDK, the window shuts. Roughly six months left (inferred).
What would show this read is wrong: If DSH's subagent_oar stays experimental or gets replaced, and oar's daily gains drop back to single digits within a month, then being embedded was a one-off and the category can't sustain stars. The other way round: if HarnessRouter and oar both keep growing steadily, and a second major project starts depending on them, the demand is real, and building a rival in another language is worth trying.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one Agent Client Protocol(ACP) Open source 2025-06 Main repo 4,583 stars, claude-agent-acp 2,730 stars ACP is a wire protocol each vendor must implement itself, so you only get the common-denominator features. OAR is a library that adapts directly to each vendor's native interface and keeps the full feature set, including steer, subagents and rawEvents HarnessRouter / UHP Open source 2026-08-09 4,259 stars, +450 in the last 7 days, plus a hosted version at harnessrouter.ai Almost the same positioning, but built as an HTTP service + open standard + commercial hosting. OAR is an in-process library with broader coverage of Chinese agents, and DSH embeds it coder/agentapi Open source 2025-04 1,496 stars, stalled Wraps a terminal as an HTTP API. Doesn't handle the native event stream, subagents or recovery Claude Agent SDK / OpenAI Codex SDK Big-company features 2025 claude-agent-sdk-typescript 1,861 star Each only handles its own agent. OAR has one API across 12 agents vercel-labs/durable-harness Open source 2026-10-16 293 star Durable runs mainly for the Vercel Workflow ecosystem, covering fewer agents MiniMax-AI/OpenAgentCore Open source 2026-09-21 319 star A self-hosted service compatible with the OpenAI Agents API. It's a service, not a library cocofhu/dsh-aio-agent Other 2026-10-31 7 likes in the DSH forum A DSH plugin that brings 6 kinds of agents into the DSH interface over ACP / JSON streams. OAR is a general-purpose library, and DSH picked it as the base for subagent_oar Nasiko-Labs/nasiko Open source 2026-02 11,820 star Only the name is similar ('Open Runtime for AI Agents'). It's actually an agent deployment platform, not a cross-harness programming interface - Where users come from
- Verified founding team's Chinese X posts: xxchan 9-30 tweeted 'don't like ACP... maybe try oar', and stdrc (original author of Kimi CLI, 14k followers) said 'Oar, extracted from a Raft base'; 10-01 + 126 stars, ranked #23 on the Trendshift TS chart
- Verified Built into DeepSeek Harness officially, and spotted by a blogger: DSH v0.2.2-alpha.2 (10-28) added subagent_oar, which depends on @botiverse/oar 0.67.0. @jakevin7 tweeted "I found OAR inside Deepseek Harness". Qiniu Cloud, Juejin, dailyinference, and aibriefingdaily covered it. +126 stars that day.
- Verified own product uses "Powered by": stdrc 11-02 launched Kanso (124 likes, 9.7k views), tweet reads 'Powered by github.com/botiverse/oar'; 11-03, 11-04 +105 stars in total
- Verified Hacker News: xxchan22's launch post got 1 points and 0 comments. Our HN library has no entries with ≥5 points.
- Inferred npm downloads from Raft's own dependencies: @slock-ai/daemon and @botiverse/raft depend on oar. npm weekly downloads 160k
Why botiverse/oar (Open Agent Runtime) took off
Research day 2026-11-06. Sources: our GitHub database (repo_profile / repo_day), the GitHub API, Trendshift, the npm downloads API, plus web pages and X mirror pages found through Exa / Jina.
One line
OAR is a TypeScript library (@botiverse/oar) that drives 12 coding agent CLIs through one API: Claude Code, Codex, Cursor, Kimi Code, Qwen Code, Grok Build, OpenCode, Pi, Antigravity, CodeBuddy, Qoder and DeepSeek Harness. Sessions, the event stream, steer / queue / abort, recovery, subagents, host MCP tools, login, usage and model listing are all covered. It was split out and open-sourced from Raft (formerly Slock, a multi-agent collaboration workbench).
Conclusion: Its stars didn't come from organic virality. Three things drove them:
- The founding team's personal influence in Chinese-speaking developer circles on X (original author of Kimi CLI, RisingWave kernel engineer).
- DeepSeek's official DeepSeek Harness (270k stars) embedded OAR as the base of
subagent_oarin the 10-28 release, and it spread after a Chinese blogger spotted it. - The team's own new product Kanso (11-02 launch) states "Powered by oar", which sends traffic back.
For real usage, look at npm, not stars: about 160k downloads a week, mostly from dependency chains such as Raft's own daemon and DSH.
1. Growth data: three waves, each matching one event
| Date | Stars gained that day | what happened |
|---|---|---|
| 2026-09-30 → 10-01 | +10 / +126 | 9-30 16:49 UTC, xxchan (@xiaoxxchan) posted a Chinese tweet: "Don't like ACP, but still want to call various agents... try oar", comparing it to OpenDAL. CEO stdrc posted "Oar, extracted from the Raft base, welcome to try it" at the same time. 10-01 It ranked #23 on Trendshift's TypeScript chart |
| 10-02 ~ 10-27 | 0–6 per day on average | Organic traffic, barely moved (207 stars) |
| 10-28 / 10-29 | +126 / +25 (Trendshift); our repo 10-28→10-29 +86 | DeepSeek Harness v0.2.2-alpha.2 released on 10-28, adding subagent_oar. @jakevin7 tweeted "I found OAR inside Deepseek Harness!!". Qiniu Cloud, Juejin, Yunzhan Community and dailyinference followed with coverage |
| 11-02 ~ 11-04 | +11 / +73 / +32 | 11-02 16:39 UTC stdrc posted Kanso (an agent-driven Mac disk cleaner, 124 likes, 9.7k views), with the tweet ending "Powered by github.com/botiverse/oar" |
| 11-05 / 11-06 | +5 / +3 | Decline |
Source: Trendshift daily table (trendshift.io/repositories/272615); our GitHub repo_day (10-28 257 → 11-06 490); tweet times converted from the status id.
- The repo was created on 2026-08-05, but Trendshift shows zero stars until 9-30, so it was probably made public in late September. As of 11-06 it had 490–493 stars and 14 forks, and it never made GitHub trending. According to why.md, trending only admits repos about 60 days old, and its daily growth wasn't high enough anyway.
- Development speed is off the charts: 56 days, 527 PRs merged (per Trendshift), versions went from v0.79.0 (10-31) to v0.92.0 (11-06). The second-biggest contributor is
stamp-build[bot], with 223 commits. Raft has a product called Stamp with the tagline "Agents ship code under their own names," so most of this repo's code was written by agents. - npm downloads are higher than the star count suggests:
@botiverse/oargot 159,802 downloads in the last week (10-30 ~ 11-05), and 30 on 295,401. Daily downloads jumped from about 3 thousand to 8–12 thousand around 10-22, which lines up with when DSH integrated it; the single-day peak was 52,512 (11-05). Packages that depend on it:@slock-ai/daemon,@botiverse/raft(Raft's own),@deepseek-ai/dsh-experimental-subagent-oarandtool-subagent-oar(DSH, pinned to version 0.67.0), and@rien7/agent-center-host(third party). Download counts include CI and mirror traffic, and multiple releases a day inflate them too, so treat them as an order of magnitude only (real install count unverified).
II. Competitors: plenty of people are doing the same thing, and some are bigger
| Name | Where | Launched | Scale | How OAR differs |
|---|---|---|---|---|
| Agent Client Protocol (ACP, led by Zed) | Open protocols | 2025-06 | agent-client-protocol 4,583 star;claude-agent-acp 2,730 star | ACP is a wire protocol between editor and agent, and each vendor has to implement it. OAR is a library that talks directly to each vendor's native interface, instead of going through the lowest common denominator. xxchan's launch tweet opened with "if you don't like ACP," |
| HarnessRouter / UHP | Open source + hosted (harnessrouter.ai) | 2026-08-09 | 4,259 stars, +450 in the last 7 days | Nearly the same positioning ("unified interface for agent harnesses," supporting Codex, Claude Code, Hermes, PI, DSH), but built as an HTTP service plus an open standard plus a commercial hosted offering. It has more than 8× OAR's stars, and its Show HN scored only 10 points. |
| coder/agentapi | Open source | 2025-04 | 1,496 stars, stalled | Wraps the terminal as an HTTP API, doesn't handle native events, subagents, or resume |
| Claude Agent SDK / @openai/codex-sdk | Big-company features | 2025 | claude-agent-sdk-typescript 1,861 star | Each covers only its own vendor; OAR covers 12 |
| vercel-labs/durable-harness | Open source | 2026-10-16 | 293 star | Runs Claude Code, Codex, OpenCode, and Pi on the Workflow SDK for persistence, Vercel-ecosystem oriented |
| MiniMax-AI/OpenAgentCore | Open source | 2026-09-21 | 319 star | Self-hosted OpenAI Agents API that can run Codex, Claude Code, and MiniMax Code |
| cocofhu/dsh-aio-agent | DSH plugin | 2026-10-31 | 7 likes in the DSH forum | Runs 6 agents inside DSH over ACP / JSON streams, going head-to-head with subagent_oar within the DSH ecosystem |
| Nasiko-Labs/nasiko("The Open Runtime for AI Agents") | Open source | 2026-02 | 11,820 star | Only the name is similar; it's an agent deployment platform, not a cross-harness programming interface. The "11.8k competitors" in the recall card refers to this one, and it's not really the same category. |
| vibe-kanban, claudecodeui, etc. | Open-source apps | 2025 | 1.4–29k stars | These are end-user apps; OAR is the underlying library for people building apps like these. |
OAR's own repo has docs/prior-arts, which surveys Synara, Paseo, Lody, One Works, Orca, Herdr, Multica, T3 Code one by one, so the team clearly knows this space is crowded.
3. Growth tactics (evidence)
- Founders' personal accounts + the Chinese-language developer X circle (confirmed). CEO stdrc (@istdrc, about 14k followers, "former author of Kimi CLI @Kimi_Moonshot, ex-database kernel engineer @RisingWaveLabs") and lead developer xxchan (@xiaoxxchan, RisingWave background) both posted in Chinese. Contributors include TennyZhuang (also from RisingWave) and stdrc himself. The first wave of 126 stars came from this.
- Embedded in a big-company product, then "discovered" by third parties (confirmed). DeepSeek Harness 0.2.2-alpha.2 embeds
@botiverse/oar, and its official docs list 7 runtime ids from the OAR registry. @jakevin7's "discovery" tweet, and the tweet by @teortaxesTex (a major DeepSeek fan account, about 80k followers) testing this feature, along with coverage from Qiniu Cloud, Juejin, dailyinference, and aibriefingdaily, all brought OAR into view. This was the second wave, and the most significant endorsement. I couldn't find how DSH chose OAR (unverified). OAR has a runtime called "DeepSeek Harness (dsh)," which suggests the integration goes both ways. - Own product carries a "Powered by" line (confirmed). Kanso (kanso.botiverse.build) ended its launch tweet with an OAR link, and stars rose by a further +105 over the next two days.
- HN (failed). xxchan's 22 launch post got 1 points and 0 comments (from the recall card), but our HN database has no OAR entry (it never cleared the ≥5-point threshold). It never took off in English-speaking communities.
- No signs of ads, influencers, or star-buying: fork/star ratio is about 2.9%, on the low side but normal, and it wasn't flagged on the suspect list.
4. Why it made the list
- It rode the timing of the "agent harness" boom: in the second half of 2026 every major player released its own coding agent CLI (Grok Build, Kimi Code, Qwen Code, DSH, Antigravity, and more). People building workspaces, remote controls, and multi-agent schedulers all need "one codebase that hooks into every CLI."
- Its endorsement comes from real adoption, not marketing: DeepSeek's official harness depends on it, which is rare among open-source libraries of this kind for being picked by a leading product. HarnessRouter has stars, but I found no comparable embedding.
- It was a production foundation split out and open-sourced: it ran inside Raft first, then went open source. Each runtime has measured docs on its native protocol (docs/runtimes, experiments/runtime-version-checks), and Chinese-speaking commenters praise design choices like rawEvents.
- The team brings its own audience: the original author of Kimi CLI and several RisingWave engineers have credibility on the Chinese developer X circle, and all three growth waves started from these people's tweets.
5. Differentiation
- A library, not a protocol. The idea is the same as OpenDAL: instead of asking every vendor to implement a lowest-common-denominator protocol (ACP / UHP), it adapts to each vendor's native interface itself, and when a capability is missing it "refuses explicitly up front instead of silently dropping it." The host gets full capabilities (steer, compact, subagent ownership via agentPath, raw rawEvents, resume), not a subset of ACP.
- Broad coverage, with a tilt toward Chinese vendors: Kimi, Qwen, CodeBuddy (Tencent), Qoder (Alibaba), and DSH are all integrated, which English-speaking alternatives often lack. This directly explains why Chinese projects like DSH would choose it.
- It handles more than sessions: installation, updates, login (without touching tokens), account usage, models, inventory, and settings forms are all covered, which suits hosts like consoles and setup wizards.
- Are these differences the reason it took off? Partly. "A library rather than a protocol" plus "full coverage of Chinese vendors" is a reasonable explanation for why DSH picked it (inference; DSH's reasons are unverified). But star growth still depends mainly on personal networks and the DSH endorsement; on product strength alone, growth in mid-to-late October would be only 0–6 stars a day.
VI. Visibility
It's circulating in a small circle. It has only appeared on the Chinese developer X circle and in secondhand coverage within the DSH ecosystem; HN scored 1 points; it never made GitHub trending; and no major English-language account wrote about it specifically (teortaxes wrote about DSH's feature). It has no revenue and isn't seeking any. It's an open-source foundation meant to bring traffic to Raft and build a technical brand.
VII. Copycats
I found no sign of anyone copying OAR directly. HarnessRouter, in the same space and at the same time, was open-sourced earlier and has more stars, so it's not a follower; dsh-aio-agent is a plugin in the DSH ecosystem doing the same kind of thing at the same time (over ACP). On npm there's an unrelated @oar-sh/oar, which just shares the name (unverified).
VIII. Takeaways
- Master it inside your own product first, then open-source the foundation. Use your own app's "Powered by" line to drive traffic, and each new product you ship brings another wave of stars to the foundation.
- Aim to be embedded by a leading project. One official DeepSeek dependency is worth many rounds of self-promotion; having a third party "discover" you is more credible than announcing it yourself.
- On the Chinese developer X circle, open with a comparison to a widely disliked approach ("if you don't like ACP, try this"). That spreads better than a generic feature pitch.
- Cover domestic agents (Kimi / Qwen / Qoder / CodeBuddy / DSH). English-language competitors generally don't bother, and this is a natural advantage for Chinese teams.
9. Uncertain
- Why DSH chose OAR (personal networks or technical evaluation), and whether it will become a formal feature: it's still marked experimental in 0.2.2-rc.1.
- How much of the 160k weekly npm downloads are real installs: CI, mirrors, and frequent releases all inflate the number.
- Whether it was a private repo before late September (inferred from Trendshift).
- How many users Raft itself has; not checked.
- The stargazer detail API returned 404, so star growth can't be lined up hour by hour against the tweets.