The State of AI Agents 2026: A Census of 93 Researched Tools

<a href="https://theagentsindex.com/blog/state-of-ai-agents-2026/" target="_blank" rel="nofollow noopener"><img src="https://theagentsindex.com/charts/state-of-ai-agents-2026.png" alt="AI agent deployment model across the index — The Agents Index" title="AI agent deployment model across the index — The Agents Index" width="640" loading="lazy" style="max-width:100%;height:auto" /></a>

We read the datasheet of every AI agent published in The Agents Index on 26 August 2026. There are 93 of them across 8 categories. This report records who publishes source, who exposes an API, where the software runs, and what the cheapest paid tier costs.

Every figure below names the number of listings it was derived from. Where we could not determine a field, the tool is left out of that count rather than counted as a no.

What the census found:

• 33 of the 93 tools are open source. All 14 tools that run only on infrastructure you control are among those 33, and exactly one open-source tool is cloud-only.
• Open source concentrates in developer tooling. 13 of the 14 agent frameworks publish source. None of the 9 support agents do, and none of the 7 sales and marketing agents do.
• 71 of the 80 tools whose API status we could determine expose a public API. We could not determine it for 13 of the 93, and those 13 are absent from the count rather than counted as a no.
• 52 of 93 run only in the vendor’s cloud. 27 offer a hosted and a private option, and 14 are self-host only.
• Freemium is the most common headline meter, on 42 of 93, ahead of usage-based pricing on 24.

What this census counts, and what it could not determine

The corpus is every listing published in The Agents Index at the moment of measurement. Each listing carries a structured datasheet, and these figures are read from those fields rather than from vendor marketing copy.

Four fields are recorded for all 93 tools: license, deployment model, interface and pricing model. API availability is the one field with real gaps. We determined it for 80 tools and could not for 13, and those 13 sit across five of the eight categories. They represent research we have not finished, rather than a category where the question makes no sense.

That distinction decides every percentage on this page. A tool we have not checked is not a tool without an API. Where a denominator below reads 80 rather than 93, the 13 undetermined tools are the difference.

The landscape: 93 tools across 8 categories

Coding agents is the deepest category at 25 tools, roughly two-thirds larger than the next one down.

Three listings sit in two categories each — Botsify and Sierra (Agent platforms plus Support agents, unchanged from the prior refresh), plus ElevenLabs Conversational AI, which picked up Agent platforms alongside Voice agents as of this refresh — so the Tools column sums to 96 against a corpus of 93. Agent tools & infrastructure is a new and small category, and we draw no conclusions from two rows.

Open source concentrates in the tools developers build with

33 of the 93 tools publish source. That share collapses to nothing at the product end of the market.

Agent frameworks are open source almost without exception, at 13 of 14. LangChain, CrewAI, LlamaIndex and Pydantic AI all publish theirs. Coding agents split closer to even, at 11 of 25. Then it stops. Not one of the 9 support agents in our index publishes source, and not one of the 7 sales and marketing agents does either.

What separates them is who signs the cheque. A framework is adopted by the engineer who will read its source, so publishing the source is part of the product. A support agent is bought by a support director who will never open it, and the vendor keeps the source closed.

For a buyer with a hard open-source requirement, this makes the shortlist a category decision before it is a product decision.

Self-hosting tracks the license almost exactly

An earlier version of this report argued that open source and self-hostable were separate axes that do not move together. Measured against the corpus on 26 August 2026, they move together almost perfectly, and we are correcting that reading here.

All 14 self-host-only tools are open source. 32 of the 33 open-source tools offer either a private deployment or both options, and Cartesia is the one open-source tool we record as cloud-only.

The closed-source row is the half that decides a procurement question. No closed-source tool in our index is self-host only, and 9 of the 60 offer a private deployment alongside their cloud. Those 9 are the complete set of options if you need software you cannot read to run on hardware you control: Augment Code, Bland AI, Claude Agent SDK, CodeRabbit, Deepgram Voice Agent API, Factory, Junie, MindStudio and Windsurf (now Devin Desktop).

An API is the baseline for 71 of the 80 we could determine

71 of the 80 tools whose API status we determined expose a conventional public API on some tier. 9 do not.

The 9 without one are Bolt.new, Crescendo, Decagon, Gemini Code Assist, Genspark, Replit Agent, Rox, Undermind and Unify. All 9 are closed source. Seven are web products and two are editor plugins, which is to say the interface is the product in every one of the 9.

An API is close to table stakes in this category, so it separates almost nothing. The deployment row above does more work for an integration decision, because an API into a cloud you cannot run yourself still leaves your data in the vendor’s cloud.

Deployment: 52 of 93 are cloud only

A majority of the index runs only in the vendor’s cloud. The 41 tools that can run on infrastructure you control are the 27 offering both plus the 14 that are self-host only, and that set of 41 is where a data-residency or cost-at-scale requirement has to shop.

Interface: 37 web apps against 36 developer surfaces

The web app is the single most common interface at 37 of 93. The tools that live in a library, a terminal or an editor come to 36 between them, which almost exactly matches it. This category is split down the middle between software you point a browser at and software you import, and the split is a good proxy for who each tool was built for.

Pricing: freemium leads at 42 of 93

42 of the 93 tools lead with a freemium plan and 24 meter by usage. Usage-based pricing is common enough here that converting the index into a single monthly figure would misrepresent every tool that charges per token, per minute or per resolution. The spec table below prints each tool’s entry price verbatim for that reason. This edition’s totals carry forward the prior bucket for the one row that swapped in the last refresh (see Method): Clay’s own page is gone, so its pricing-model bucket cannot be re-checked against source.

Capability tags: what the index can be filtered on

Tags are assigned by our editors and counted out of all 93 listings.

Two of these tags are worth checking against the datasheet, because they duplicate a field. The open-source tag sits on 33 listings and the datasheet records 33 open-source tools, with no listing carrying one and not the other. The self-hosted tag sits on 41, matching the 41 tools whose deployment field reads self-host or both, again with no disagreement in either direction.

Claude and GPT support are tagged on 51 listings each, which is a coincidence rather than a finding. Browser use is the rarest tag at 8 of 93. As with the pricing-model table above, these counts carry the prior tag set forward for the one swapped row rather than reconcile it against a page that no longer exists.

Full spec table (93 tools)

A tick means yes and a cross means no. A question mark in the API column means we could not determine it, and a question mark is not a no.

Method

• Measured 26 August 2026 against every listing published in The Agents Index at that moment, which was 93 listings.
• License, deployment, interface and pricing model are read from each listing’s structured datasheet, never inferred from prose. All four are recorded for all 93 tools.
• API availability is read from the structured datasheet, which records a conventional REST API on some tier. It is determined for 80 of the 93 and undetermined for 13.
• Capability tags are counted from the tag field, which our editors assign separately from the datasheet.
• Entry price is each tool’s own lowest-priced paid tier, printed verbatim in the tool’s own units. A tool with no paid tier reads Free. A tool that publishes no figure reads Custom.
• Three listings hold two categories each this edition: Botsify and Sierra (Agent platforms plus Support agents, unchanged from the prior refresh), and, newly, ElevenLabs Conversational AI (Voice agents plus Agent platforms, previously Voice agents only). Category counts now sum to 96 against a corpus of 93, up from 95, and Agent platforms moves from 15 to 16 tools as a result.
• One listing swapped in the prior refresh: Clay is no longer a Published listing, and VoiceAgent joined Voice agents in its place. VoiceAgent’s license, API, deployment, interface and entry price were verified against its own listing at that time and are reflected in every table above. Clay’s page returns a 404, so its pricing-model bucket and capability tags could not be checked against source at the time of that swap; the pricing-model and capability-tag tables still carry the prior totals forward rather than reconcile that one row, pending a full re-pull of those two tables specifically.

Citing this census

These figures describe our index as it stood on 26 August 2026, and the corpus grows. If you use a number from this page, cite The Agents Index and carry the measurement date with it, so the figure stays checkable after the next listing lands.

See also

• The Agents Index directory — the sourced entry for each of the 93 tools counted here, with pros, cons and verdicts.
• Open-source license census — which license each open-source agent actually ships under.
• API census — the API field in full, including what we count as an API and what we do not.
• Deployment census — the self-host question on its own, category by category.
• How The Agents Index researches AI agents — the sourcing standard behind every listing in this report.