The State of AI Agents

The short version

The agent ecosystem is still overwhelmingly a public-repo ecosystem: 86% of the 247 tools we track are open source, and commercial products are a thin 14% layer on top. Python leads TypeScript 1.8:1 overall — but that headline hides the real story, which is territorial: TypeScript already owns the categories where the thing being automated is itself a web app. The most crowded arena is Coding Agents, at 37 tools, while the two categories closest to a paying customer — support and sales — hold just 9 open records between them.

What the numbers say

1. 01

   The ecosystem is 86% open source — the centre of gravity is still the repo, not the SaaS product.

   212 of 247 tools ship as open source; only 17 are paid products and 18 freemium. For a category with this much venture attention, that ratio is unusual — and it means the practical constraint on adopting an agent stack is rarely licensing.

2. 02

   Python leads TypeScript 1.8:1 — but the split is territorial, not global.

   Across 209 open-source projects that declare a primary language, Python takes 103 and TypeScript 57. Yet TypeScript is ahead in Coding Agents, Workflow & Productivity, Agent Infrastructure — the categories whose subject matter is itself JavaScript: editors, browsers, web workflows. MCP servers are split exactly evenly (10 each), which is what a genuinely new, still-unclaimed category looks like.

3. 03

   The most crowded category is Coding Agents — 37 tools.

   Then Agent Frameworks (31), Voice Agents (29), MCP Servers (28). Crowding is not the same as maturity: it marks where the problem is legible enough that everyone attempts it, which is also where differentiation is hardest to find.

4. 04

   The most fragmented arena is Voice Agents — 29 tools and no dominant project.

   Its most-starred entry, xiaozhi-esp32-server, has 10k stars — against 199k for n8n in Workflow & Productivity. When a category is busy but flat, nobody has found the shape of the answer yet; expect consolidation rather than a new entrant.

5. 05

   Customer-facing agents are where open source is thinnest — 9 records across support and sales.

   These are the categories closest to a budget line, and they are the ones commercial vendors hold. The pattern repeats across the corpus: the further you get from a developer’s own machine, the more the open ecosystem thins out.

6. 06

   Momentum is concentrated in coding and web-data tools — jcode added 3,620 stars in 7 days.

   Over the 27 July – 3 August 2026 refresh window, the fastest-growing records were jcode, firecrawl, deer-flow, n8n. Two sub-areas keep producing the movers: agents that write code, and the tooling that feeds agents the web.

Jjcode15k stars▲ 0Ffirecrawl160k stars▲ 0Ddeer-flow79k stars▲ 0Nn8n199k stars▲ 0SScrapling72k stars▲ 0Ddify151k stars▲ 0Cchrome-devtools-mcp48k stars▲ 0Rruflo67k stars▲ 0Bbrowser-use108k stars▲ 0Aagent-governance-toolkit5.6k stars▲ 0

Δ stars over 7 days · refreshed every Monday

How this was counted

Agent Search Engine is a hand-curated index of AI agents, frameworks, and MCP servers — every record reviewed and categorised by a person, not assembled by keyword scrape. Adoption figures are live GitHub stars, refreshed weekly by an automated pass; category and editorial judgments are ours. Records without a public repository carry no star signal and are never ranked as if they did. Rankings are not for sale — no placement in this report or anywhere on the site can be purchased. The full method is on the methodology page, and the entire corpus is machine-readable at /llms-full.txt.

Every figure above is computed from the live index when this page is served, not typed in by hand — so the report ages with the ecosystem instead of freezing at publication. The date under the headline is the last refresh it reflects.