The State of Open Source Agentic AI

01 · THE TERRITORY

The map has 25 named regions.

We read each repository's README, file tree, and source, and wrote down what it actually builds. Those descriptions were embedded and clustered, so two repositories sit near each other when they do similar things. Twenty-five regions emerged, built from 178 finer sub-clusters, running from coding-agent harnesses to voice cloning.

No taxonomy was imposed; the categories are the ecosystem's own structure.

Hover a region to light it up; click to pin it and unfold its sub-clusters. Hover any point to identify it; click for the full card. Explore this view →

02 · HOW WE GATHERED THIS DATA

Seven gathering passes. Every repository read.

Curated lists, topic and code searches, ecosystem sweeps, dependency and network hops, in seven passes over one territory: coding agents, agent infrastructure, and the tools agents wield.

Every repository was then read (README, file tree, one or two source files) and classified by a frontier model.

Full methodology & dataset notes in the appendix. Every number on this page regenerates from audited scripts.

03 · ONE YEAR OLD

73% of the territory didn’t exist last July.

14,890 of these repositories were created in the twelve months before our snapshot, and 89.6% since the start of 2024.

Lit on the map: repositories created since July 2025. Click a bar to light a single year.

04 · THE AUTHORSHIP FLIP

Two-thirds of 2026’s repos are AI-assisted builds.

We classified each repo's construction from its bones: tree shape, tests, CI, the fingerprints agent workflows leave. In repos created in 2023, 1 in 20 showed AI-assisted construction. In 2026: 2 in 3.

The boundary is soft and we drew it with an error band, but the direction is not: agentic software is now mostly built with the agents it is built for.

Lit on the map: every repository our read classified as an AI-assisted build.

05 · A LOW-STAR WORLD

Nearly half the map sits under ten stars.

Stars measure attention, and attention is concentrated: the 1,173 repositories above 10,000 stars, 5.8% of the corpus, hold 77.7% of its 47.9 million stars. The median repository has 16. Most writing about this ecosystem covers the starred end; most of what is being attempted sits below ten stars.

Lit on the map: repositories under ten stars, 45% of the corpus. Click a band to light it instead.

06 · THE HARNESS EXPLOSION

1,250 people built their own coding agent.

The largest specialized region on the map is the coding-agent harness itself: terminal agents, runtimes, and CLIs. Median star count: 6. Building your own coding agent is 2026's rite of passage, the way a generation once wrote its own static site generator, except this artifact can rebuild itself.

The camera is parked on the harness region; its three big neighborhoods are Rust runtimes (503), Go runtimes (403), and DIY terminal CLIs (142).

07 · SYSTEMS LANGUAGES

Rust and Go are eating the agent stack.

Python wrote the model era; it is not writing the harness era. Among repos born in 2026, Rust and Go together pass 30%, roughly double their 2023 share, while Python fell from half the field to under a third. Single binaries, fast startup, no dependency hell: the qualities you want in a tool an agent installs a thousand times a day.

On the map, points are recolored by language:

08 · THE MCP ECONOMY

One in eight real repos is an MCP server.

The Model Context Protocol shipped in November 2024. Twenty months later, 2,364 repositories here are MCP servers, 12.6% of everything real, and last quarter they appeared at nine per day.

Much of it is already a ghost town. The median MCP server has 9 stars, the MCP-server region carries one of the highest dormancy rates of the 25 regions, and MCP vaporware is the largest single genre of slop.

Lit on the map: every MCP server, threading through every region. Click a bar to light a single quarter's arrivals.

09 · THE SLOP STRATUM

7.9% of the territory is slop, in 63 genres.

1,608 repositories make claims their code does not back. We mapped them rather than deleting them, and they sort into genres:

One caution: our reader over-calls slop on real code hiding behind deceptive READMEs. In a blind audit, roughly 40% of sampled slop was real code presented deceptively; treat the label as triage, not a verdict.

Lit on the map: the slop stratum, mapped in its own space and projected in as the violet fringe. Click a genre bar to isolate it.

10 · THE DECEPTIVE BUMP

Slop falls as stars rise, except in one band.

Between 50 and 199 stars the slop rate nearly doubles: the zone where deceptive presentation buys real credibility.

At the top of the range sits a “coding-agent” rebrand with a stub source file and 194,000 stars. Stars are the industry's default quality signal; they can be manufactured at six-figure scale.

Lit on the map: slop repositories sitting in the 50–199-star band. Click any band to see its slop instead.

11 · CATEGORIES BEING BORN

Whole regions of this map didn’t exist a year ago.

We can timestamp them: the median repository in each of these regions was created in Feb–Mar 2026. Code-graph indexes that let agents stop re-reading codebases. Penetration-testing agents. Worktree managers for running five agents on one repo. Payment rails that let an agent spend money with a budget. And the largest newborn, skills: instruction files agents write for other agents. It is the most AI-built region on the whole map, and the thinnest; much of it is folders of markdown rather than software.

All five newborn categories are lit; hover a chip to isolate one.

12 · THE VERTICAL WAVE

Everyone is attempting “agents for X”.

The vertical-agent future the industry keeps predicting is being attempted across the map: six hundred and ten mostly unnoticed tries at trading desks, tutors, health coaches, legal drafters, trip planners. Median star count: 3. Seventy-one percent AI-assisted builds.

This is what early adoption looks like: professionals who know what a coding agent is, spinning one up against their own field and publishing the result. What they're building:

The camera is parked on the vertical & domain agents region. Hover a sub-category to isolate it.

13 · THE PROBLEM LENS

What problems does the ecosystem fixate on?

Every repo also told us why it exists, and the whys cluster differently than the whats. The biggest fixation: agents lose all memory between sessions. 790 repositories exist to fix that. Read the count with some care: memory systems are also the easiest agentic infrastructure to attempt, the problem is obvious, and the models themselves gravitate to it when asked what to build.

Then the integration burden, the cost of frontier APIs, and the fact every practitioner already knows: AI writes code and reports it as correct when it isn't. 407 repositories exist to catch that.

The map is recolored by problem: each hue is one of 120 shared reasons to build. Click a bar to isolate one.

14 · ONE PROBLEM, TWENTY SHAPES

The mechanism space is still speciating.

“Getting an agent to production” is one sharply-felt problem solved by 16+ unrelated kinds of object: runtimes, k8s operators, fleet platforms, frameworks. Nobody has found the shape yet. Memory, by contrast, is converging on one mechanism family. Here is every problem we measured, most-scattered first:

Watching which problems scatter and which converge is watching the ecosystem decide what its products will be. The scatter you see on the map is a fair picture: fragmentation is measured in the full-dimensional space, and the 2-D view spreads these repos out much as the measurement does.

Lit on the map: the 302 production-infrastructure repos, colored by which mechanism region each belongs to: sixteen different answers to one question.

15 · AGENTS IN OPERATIONS

Agents are moving into the operations jobs.

Follow the problem statements past the development loop and the biggest cluster is security: 254 repositories building agents that run penetration tests. Behind it, agents doing incident root-cause, production observability, and telemetry queries; 477 repositories in all point agents at software that is already running.

Alongside them, 195 repositories exist to measure agent performance itself: teams wanting to see how well agents actually do, graded on their own tasks and data rather than public benchmark suites.

Lit on the map: the five post-deploy problem clusters. Click a bar to isolate one.

16 · THE NAMED AGENTS

Claude Code is named nearly twice as often as Codex.

When a repository's own description names a coding agent, it marks what the project is built for, on, or around. Claude Code appears in 3,694 descriptions, 18% of everything we read and nearly double Codex. Read the smaller counts loosely: further down the table, agents are often named only in passing lists (“works with Claude Code, Codex, and others”).

Lit on the map: every repository whose description names one of these five agents. Click a row to isolate one.

17 · ROUTING AROUND LOCK-IN

626 repos exist to undo vendor lock-in.

Four problem clusters share one purpose: opening what vendors have closed. The largest, 257 repositories, unlocks one harness from its maker's backend; the rest multiplex accounts and quotas, swap providers behind a single interface, and let any model drive any CLI. Together: 626 repos.

Whatever structural lock a vendor ships, the builders treat as a puzzle. These are only the attempts published in the open.

Lit on the map: the four lock-in/routing problem clusters, 626 repos. Click a bar to isolate one.

18 · THE REDUNDANCY CENSUS

549 repositories have a near-duplicate.

Compare every capability description against every other and 2.9% of the real corpus has a true near-duplicate somewhere on the map. They are not the junk tail: the median duplicated repo has 122 stars against 21 for everything else, and 98% of duplicate pairs connect different owners. Independent reinvention, mostly of what is already popular. The truly one-of-a-kind, with no meaningful neighbor anywhere on the map, number just 323.

Lit on the map: the 549 repositories with a true near-duplicate. Click a region bar to see where the copying concentrates.

19 · MORTALITY

Younger repos should be more alive. Since 2023, they aren't.

A repo born in 2024 has had two fewer years to run out of steam than one born in 2022, so dormancy should fall with every younger cohort. It doesn't. The class of 2023 is 43% dormant or archived and 2024 sits at 36%, both far above 2022's 26%.

The break coincides with the arrival of AI coding, and the rise is composition: same-age AI-assisted repos go quiet at 26.5% against 20% for human-built (hold 2022's authorship mix and 2025 would sit at 21%), and slop goes quiet fastest of all: 69% of slop created in 2025 is already dormant or archived. Human-built 2025 is the lowest-mortality cohort on record. The class of 2026 is not in the chart; our pulse measure needs months of inactivity, a bar no repo created this year can meet yet.

The map is recolored by pulse. Click a bar to light one cohort's dormant & archived repos.

20 · DRIVE-BY REPOS

15.7% of 2025’s repositories shipped once and stopped.

A drive-by repo's entire observed life, last commit minus creation, fits inside seven days. Under 2% of repos born through 2022 were drive-bys; the class of 2025 is at 15.7%, and its already-dead repos have a median life of 7 days, down from 782 for 2021's. The rate splits by construction: 5.6% of human-built repos, 14.7% of AI-assisted, over half of slop.

The class of 2026 is on the same path: 4,024 of its repos are already drive-bys, a fifth of this entire corpus, and 76% of them still read as active, because a young repo's last commit is always recent. Drive-by share is the leading indicator; the mortality chart above is where it lands.

Lit on the map: every judged drive-by repo. Click a bar to light a single cohort.