Skip to content
repository radar

a repository radar special report

OSS Vector

Summer 2026 · the state of open-source software 📡

Open source is the fastest mechanism software has for evolving its own stack: millions of projects compete for adoption, and developers and users collectively pick the winners. This is where that selection pointed in the summer of 2026, read across 242 high-signal repos we hand-picked for importance, momentum, and novelty, then tracked across stars, commits and downloads while they moved.

242
repos hand-tracked
8.3M
stars under watch
914k
forks
67M
package installs / week
14k
commits / week shipped
11/100
of GitHub's all-time top

deep dive · reading 1

🤖 It's all AI now

The "state of open source" is really the state of open-source AI. Across all 242 repos, nearly every category is a different face of the same build-out.

category · repos covered

27 11%
21 9%
🔄 Other
17 7%
4 2%

the one-line read

~9 in 10
repos touch AI

Strip out a thin tail of pure-infra and security tools and the field is, functionally, open-source AI. The non-AI repo is the exception now, and even the plumbing (gateways, memory, inference) exists to serve models and agents.

Read the split on the left as one thing wearing nine hats: nearly every bucket is a different face of the same AI build-out.

deep dive · reading 2

🕵️ Agents are the field

Not one category among many. Agents & Orchestration is the single biggest bucket, and everything else increasingly serves it.

the biggest bucket, by far

92
agent repos
38%
of the field

Agents & Orchestration is bigger than Coding, Infra and RAG combined, and the boundary keeps expanding. Memory, gateways, voice and inference increasingly exist to serve agents, not to stand alone.

"Agentic" stopped being a category this year and became the field's organising principle: the thing most other repos are built to feed.

deep dive · reading 3

🧠 Memory splits from RAG

For two years "RAG" was the catch-all for anything that fed context to a model. This year it split: persistent agent memory and code-as-a-graph broke out as their own category, distinct from plain document retrieval.

breaking out of the RAG bucket · by stars

claude-mem context across sessions 87k ★
Understand-Anything code → knowledge graph 74k ★
codegraph code as a graph 60k ★
agentmemory persistent agent memory 25k ★
cocoindex incremental live index 11k ★

why it's its own thing now

Retrieval answers "what is relevant to this query." Memory answers "what does this agent already know." Once agents began running continuously rather than one-shot, the second question stopped being a RAG setting and became its own product.

Code-as-a-graph is the same move for engineering: instead of chunking files for retrieval, the whole codebase is indexed as a structured graph the agent can traverse. The fastest-rising shape of it, codegraph, added +43k stars a month.

the shift

Memory is no longer a feature inside RAG, it's a category with its own breakout repos.

deep dive · reading 4

📦 Libraries → products

Open source used to mean a library you import. The breakout repos now are applications you run: desktop apps, CLIs, self-hosted web apps and native binaries, aimed at end users, not just developers.

39%
ship a runnable build
19%
are still import-only libraries
6/10
biggest repos are products, not libraries

products, not just libraries · by stars

🦀 OpenClaw AI assistant · any-OS app 383k
💬 Open WebUI LLM UI · self-host 145k
📷 Immich photos · self-host 108k
🎤 meetily notetaker · desktop 24k
🐋 orca agent IDE · desktop 20k
📝 Reor notes · desktop 9k

library vs application

The distinction is concrete: a library is code other developers import (npm, PyPI); an application is a finished tool an end user runs (desktop app, CLI, self-hosted web app, native binary). Counted by what each repo actually ships: 39% now ship a runnable build, just 19% are import-only libraries, and the rest run from source.

That resets the competition: an open-source project is no longer "a package on npm," it's a product going head-to-head with a funded SaaS, on distribution, polish and trust, not just code.

the shift

Open source went consumer. The unit is the app, not the import.

deep dive · reading 5

🚀 From cold start to six figures, in months

A GitHub project once took years to reach six-figure stars. Now it takes months. Of the 14 repos we track above 100k, nine are less than two years old, and the steepest went roughly 100× in half a year.

the 100k+ club · from → now · time since covered

OpenClaw 4k → 383k 6 mo
hermes-agent 27k → 215k 3 mo
OpenCode 33k → 186k 8 mo
Langflow 60k → 152k 14 mo
Firecrawl 23k → 151k 17 mo
Open WebUI 89k → 145k 15 mo
Claude Code 5k → 138k 16 mo
agency-agents 73k → 132k 3 mo
spec-kit 29k → 121k 9 mo
llama.cpp 90k → 120k 8 mo
cc switch 5k → 117k 8 mo
Immich 82k → 108k 8 mo
Gemini CLI 52k → 106k 12 mo
BrowserUse 29k → 105k 17 mo

the steepest curve in the field

400k200k0 4k · PR#25 383k +62k / mo at peak JanMarMayJul ’26

OpenClaw: 4k → 383k in six months, a hundredfold rise and the steepest curve we've tracked. Cold-start to viral is now a matter of weeks, and eleven of these are among the 100 most-starred repositories in GitHub's history.

deep dive · reading 6

🔐 The coding agent became a platform

A whole stack has grown on top of the coding agent, four layers, none of which existed in this form a year ago. The agent itself is now the least interesting part.

🔧 the harness layer

Run, switch, multiplex the agents

Once the coding agent is a commodity, value moves to orchestrating several at once, on the accounts you already pay for.

💰 the token-cost economy

Tools that exist to cut the bill

A cluster that did not exist a year ago, built purely to shrink what agents cost to run. When they run continuously, the bill becomes the product.

🔌 the MCP layer

How agents reach tools and data

The connective tissue has standardised, and it is now the most-installed corner of the whole field. FastMCP alone pulls 19M downloads a week, on a fraction of the stars.

🧩 skills & marketplaces

What the agents can actually do

The top of the stack: packaged capabilities and the marketplaces distributing them, the app store forming around the agent.

deep dive · reading 7

🛠️ The measure-and-run layer grew up

Once agents shipped to production, two support layers matured fast: the tools to measure them (evaluation, tracing) and the tools to run them reliably (gateways, serving). Both are now infrastructure, not afterthoughts.

📏 measure · eval & observability

MLflow eval + experiment tracking 10M dl/wk
Deepeval LLM eval framework 1.2M dl/wk
TensorZero LLMOps platform 12k ★

🔀 run · gateways & serving

LiteLLM one API → 100+ models 54k ★
OmniRoute free AI gateway 18k ★
FlashInfer LLM serving kernels 1.1M dl/wk

The tell that a field has grown up: "did it work?" and "can it scale?" become products of their own. Evaluation is now a downloads heavyweight (MLflow alone pulls 10M installs a week), and the gateway/serving layer is the plumbing every production agent quietly runs on.

deep dive · reading 8

Stars are the headline, not the story

A star is a bookmark: it measures attention, not use. Rank the field by what people actually install or download instead, and the order barely holds.

libraries · by weekly installs (npm / PyPI)

repoinstalls/wkdl
LiteLLM 170M #1 #14
FastMCP 18M #2 #31
BrowserUse 11M #3 #6
MLflow 11M #4 #29
OpenAI Agents 8.1M #5 #27
Playwright MCP 5.9M #6 #22

products · by binary downloads (GitHub, total)

repodownloadsdl
Voicebox 1.2M #1 #24
OpenCode 1.1M #2 #3
cc switch 650k #3 #6
onlook 435k #4 #39
RTK 416k #5 #14
KittenTTS 230k #6 #56

what the two signals say

Two usage signals, one story: weekly installs rank the libraries, binary downloads rank the products, and neither tracks stars. LiteLLM takes 170M installs a week at #14 on stars (automated CI builds inflate the raw count, but the direction is clear); Voicebox leads product downloads at #24. A star is a bookmark, not proof anyone runs it.

8 of 12
most-starred repos are products, with no package to download

Usage numbers exist only where they're public: weekly installs for the npm / PyPI packages, cumulative binary downloads for the repos that ship them. Source-only and self-host apps stay uncounted. Stars measure what's bookmarked, installs and downloads what's used, commits what's maintained. We track all three because no single one is the story.

deep dive · reading 9

👤 Right now, one or two people carry each project

Behind the star counts, the recent work clusters on very few hands. Across the 200 repos with enough history to judge, over the last six months one author writes a median 44% of the commits, and 42% had effectively a single committer in that window.

top author's share of commits, last 6 months · across 200 repos

< 25%
38 repos
25–50%
79 repos
50–75%
48 repos
75–90%
21 repos
> 90% · solo
14 repos

Median project: 44% of the last six months' commits from one author. For 42%, that's one contributor away from stalling its current pace.

even the famous run on a few recent hands

TradingAgents 90% one author 93k ★
Claude Code 85% one author 138k ★
Open WebUI 65% one author 145k ★
OpenClaw 49% one author 383k ★
The healthy exception, Immich: 108k stars, and even lately just 8% of commits from its top author, across 900+ lifetime contributors. Proof that big hype and a broad base can coexist, it's just rare.

part II · the field, quantified

What it's built in, how it's licensed, who's behind it

The descriptive base under the readings: what the field is built in, how openly it's licensed, and who is actually behind it, by capital and by origin.

💻 language · share of field

Python
44%
TypeScript
30%
Rust
9%
Go
5%
everything else
12%

Python for the models, TypeScript for the product: three quarters of the field. Rust is the latency layer underneath, 9% and climbing.

📜 license · is it really open?

permissive 73 8 other 15
permissive 73% · MIT, Apache: use it in anything, including a closed, paid product copyleft 8% · AGPL, GPL: you can use it, but your changes must stay open (the anti-hyperscaler move) source-available 4% · BSL, SSPL, Elastic, Functional Source: read the code, but you can’t freely run or compete with it other / custom 15% · unrecognised or bespoke terms GitHub can’t classify

Most of the field is genuinely open. The tell is the 4% that has gone source-available: the Terraform / Redis / Elastic move (read the code, but you can't run a rival on it), now arriving at the frontier. It's the one licence shift that can kill adoption overnight.

💰 who's behind it · by capital

independent 44 VC-backed 29 company 17 10
independent · 44% VC-backed · 29% company · 17% foundation · 10%

Only 10% is foundation-backed: the model with the longest half-life. Nearly half (44%) rides on one person or a small unfunded team, and VC money is already the second force at 29%. This is the durability question the star count never answers.

👷 and by origin · two engines

US big tech

Microsoft 5 · Google 3 · OpenAI 2 · Vercel 2

a land-grab for mindshare

Playwright MCP, Gemini CLI, Codex, Claude Code

China-origin labs

HKUDS 7 · ByteDance 2 · OpenBMB 2 · DeepSeek 2

agent tooling at cadence

DeerFlow, UI-TARS, Qwen3 · HKUDS is the most-featured builder overall

part II · leaderboards

What actually moved, four ways

The same field ranked four ways: stars gained (monthly), code shipped, package installs (weekly) and product downloads (total). The top of one board is rarely the top of the next.

⚡ buzz · stars / month

1 OpenClaw +62k
2 hermes-agent +57k
3 codegraph +43k
4 headroom +38k
5 taste-skill +23k
6 OpenCode +20k
7 agency-agents +18k
9 RTK +16k
10 pi-mono +15k

attention, not adoption: what the timeline is talking about.

🚀 output · commits / week

1 ELIZA 2,657
2 OpenClaw 2,381
3 hermes-agent 761
4 hermes-webui 430
6 orca 368
7 CopilotKit 317
8 LiteLLM 314
9 codewhale 289
10 Mastra 268

how hard it's shipping: commits merged, week in, week out.

📦 libraries · installs / week

1 FastMCP 19M
2 BrowserUse 11M
3 MLflow 10M
4 OpenAI Agents 7.6M
6 OpenClaw 2.1M
8 FastAPI-MCP 1.9M
9 Deepeval 1.2M
10 FlashInfer 1.1M

what's actually installed: npm + PyPI pulls per week.

🖥️ products · binary downloads

1 Voicebox 1.2M
2 OpenCode 1.1M
3 cc switch 650k
4 onlook 435k
5 RTK 416k
6 KittenTTS 230k
7 OpenTofu 207k
8 Jan 165k
9 worldmonitor 149k
10 meetily 127k

runnable builds pulled from GitHub releases (total).

part III · where this goes

Where the vector points next

Fold the nine readings together and one arc runs through them, then three calls we're willing to be wrong about.

the through-line

Open source’s frontier is now AI, agent-shaped and shipped as products, often maintained by one or two people, and judged by usage and cost, not stars. The contest moved from "can we build an agent" to "can we run fleets of them, cheaply, on our own machines." And the arrow points one way: AI and agents, coding most of all, are only accelerating, so every signal here matters more each cycle, not less.

Open source is AI's distribution layer

What open source distributes has changed: not just packages developers import, but finished products people actually run, still built and reviewed in the open, community way. The path from launch to a user's machine now runs through a public repo, and the lag is collapsing. Cold start to six-figure adoption fell from years to months.

watch Open WebUI · dyad

Value moves to the scaffolding

As models commoditise, the margin shifts to the layers around them: the harness that runs and multiplexes agents, and the cost tooling that keeps them cheap. The next platform contest is fought here, in the open.

watch headroom · cc switch

Good models get small enough to run local

The frontier isn't only bigger models. Capable models keep getting smaller, and that's what puts real use cases on local hardware: good-enough intelligence on a laptop or a phone, with no data leaving the machine. That is decisive in privacy-bound verticals (health, legal, gov) the funded-SaaS grid can't easily follow.

watch Jan · Unsloth

Step back far enough and this is just open source doing what it does: the abstractions that won this cycle are already turning up in everyone's stack. What we tracked is that selection in fast-forward.

oss vector · summer 2026 · figures frozen as July 16, 2026