Which AI Agents Can You Self-Host? A Census of 94 Tools

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

Of the 94 AI agent products published in The Agents Index on 27 August 2026, 42 can be run on infrastructure you control and 52 are cloud-only. The Deployment field is filled in for 94 of the 95 published listings, so every count on this page is out of the full index with no unknowns set aside. The corpus also holds one agency listing, Agentix Labs, which sells engineering services rather than software with a deployment mode of its own, so it sits outside this count entirely.

Correction, 11 September 2026. This report originally read “41 can be run on infrastructure you control and 53 are cloud-only” out of 94 listings, and described the Deployment field as filled in for 94 of 94. No collector run in this index’s history holds 53 cloud-only listings. The reading the page was measured against, 19 August 2026, held 94 published listings of which 93 carried a deployment value, split 52 cloud-only, 14 self-host only and 27 both; the 53 was that 52 with Lead Scorer added by hand to reach a denominator of 94. This edition is restated to the single reading that produces every figure on it at once, collector run 221173 of 27 August 2026: 95 published listings, 94 of them carrying a deployment value, split 52 cloud-only, 13 self-host only and 29 both. Three coding agents changed bucket in that window and the closed-source table below gains a tenth row. If you cited the earlier 53-of-94 figure, this is the correction.

That field is easy to skim past because it looks like a formality. It isn’t. If a compliance team needs code and customer data to stay inside a network you control, this one field decides whether a tool is even in the running, regardless of how good its agent is.

The headline count

52 of the 94 tools in The Agents Index (55%) are cloud-only, with no self-hosted or on-premise option at any price. The rest split two ways: 13 of 94 (14%) are self-host only, mostly frameworks and libraries with no managed cloud offering at all, and 29 of 94 (31%) offer both, a cloud-hosted default with a genuine private-deployment path alongside it.

So 42 of the 94 tools (45%) can be run somewhere you control. Which 42 depends heavily on what you are buying.

The category split is the real finding

The table says one thing plainly: what you are buying predicts whether you can self-host it far better than how much you pay. Agent frameworks are unanimous in one direction, support and sales agents unanimous in the other, and no price tier crosses that line.

(Sierra, Botsify and ElevenLabs Conversational AI each belong to two categories and are counted once in each, so the category rows sum to 97 against an index of 94 distinct tools. The same convention applies in our open-source license census.)

Support and sales agents offer nothing at any price

All 17 tools across support agents and sales and marketing agents are managed cloud products only. They are Sierra, Intercom Fin, Decagon, Crescendo, Cresta, Ada, Botsify, Zendesk AI Agents, Auto-Respond, 11x, Artisan, Qualified, Rox, Unify, Warmly, Amplemarket and Lead Scorer. 0 of the 17 offer any self-hosting at any price.

If a support or sales team has a hard data-residency requirement, there is currently nothing in either category on this site that clears it. The constraint has to be solved contractually, through data processing agreements and regional hosting commitments from the vendor, rather than architecturally.

The third all-cloud row, agent tools and infrastructure, holds only two listings, Browserbase and SalesTouch. Two tools is far too few to read as a property of that category. It is in the table for completeness, not as a finding.

Agent frameworks are the mirror image

The index holds 14 agent frameworks. 14 of the 14 agent frameworks in The Agents Index offer some form of self-hosting. That is not a coincidence, since a framework you cannot run yourself is a managed service with extra steps. The split inside the category still matters.

6 of the 14 agent frameworks are self-host only, though two carry a caveat. LangChain, Microsoft AutoGen, the OpenAI Agents SDK and Pydantic AI have no managed layer at all. Agno and AgentsKit sit in the same bucket for different reasons: Agno’s paid AgentOS tier still deploys into the customer’s own cloud rather than a vendor-run one, and AgentsKit’s own AgentsKit OS is enterprise-only, early access, reachable only by booking a sales demo, and discloses no public pricing, so it does not read as a generally available second deployment path yet.

The other eight frameworks pair a self-hosted core with an optional managed service: Microsoft Agent Framework, LlamaIndex, Mastra, Letta, Google Agent Development Kit and Claude Agent SDK, whose managed layers are Foundry Hosted Agents, LlamaCloud, Mastra Cloud, Letta Cloud, Google’s Agent Runtime and Anthropic’s Claude Managed Agents respectively, plus CrewAI, which ships a hosted platform with a visual editor alongside its open-source Python framework, and LangGraph, whose deployment and tracing layer is LangSmith.

Coding agents spread across all three modes; platforms split on licence

Coding agents are the most genuinely flexible category, and the only one with real representation across all three deployment modes rather than clustering at one end: 32% cloud-only, 20% self-host only, 48% both. The index holds 25 coding agents, and 17 of the 25 coding agents (68%) offer some form of self-hosting. That range exists because “coding agent” spans everything from a $0 terminal binary you point at your own model key to a fully managed cloud IDE, and buyers here actually care about the difference. Our coding agent comparison guide covers the rest of that decision.

Agent platforms are the category where workflow-automation buyers with compliance requirements would most expect on-premise options, and a little under half deliver. The index holds 16 agent platforms, and 7 of the 16 offer a self-hosted path. They are n8n, whose own hosting docs cover running it on-premises or in a private cloud, plus Dify, Botpress, Sim, Vellum, MindStudio, whose self-hosting is gated to its Business plan, and OpenClaw, the only agent platform in this index with no vendor-hosted option at all.

Every one of the 9 cloud-only agent platforms is closed-source, and every platform with a private path is open source except MindStudio. In this category the licence and the deployment option move together for all but that one listing, whose self-hosting is a Business-plan feature on a closed-source product.

Among research and data agents, 10 of the 11 are cloud-only and GPT Researcher is the sole exception. The cloud-only ten include Microsoft Fabric Data Agent, which runs only inside a Fabric or Power BI Premium capacity, and Webhound, whose hosted web app and MCP server have no self-run option.

Self-hosting and open source are not the same fact

Read the “any self-host option” column above next to our open-source license census and a pattern breaks. We have determined open-source status for 93 of the 95 published listings, and 33 of those 93 are open source. The two we have not are Agentix Labs, the agency listing this census excludes, and Lead Scorer, which joined the index after our last full research pass and has not been read for it yet.

Of the 42 tools that offer some form of private deployment, 32 are open source and 10 are closed-source products that offer a private-deployment path anyway. Claude Code is why that count reads 10 rather than 9: its deployment field moved to self-host on 24 August 2026 while the product stayed closed-source. MindStudio, the addition the previous edition explained, is still here, its open-source status undetermined then and confirmed closed since. Both are in the table below, and 0 of the 42 tools with a private path are left with an undetermined open-source status.

The reverse gap matters just as much. 32 of the 33 tools in our index recorded as open source have a self-hosted path. 1 of the 33 does not: Cartesia, whose Line voice-agent runtime always runs on Cartesia’s own managed infrastructure even though its Sonic and Ink speech models can be self-hosted into a customer’s environment on an Enterprise contract. So do not assume “open source” means “you can just run it yourself” without checking the deployment column specifically. Our open-source collection and the coding agents self-hosted facet page show how much the work varies, from a single static binary to a real orchestrated deployment.

The ten closed-source tools you can still run yourself

10 of the 42 tools with a private-deployment path are closed-source, and they do it in three genuinely different ways.

The first five are the enterprise pattern: closed-source code, but a paid tier, or in Deepgram’s and MindStudio’s case a specific plan, that lets a regulated buyer keep data inside their own infrastructure. The vendor sells deployment flexibility as a feature rather than the source code.

Augment Code, Windsurf, Junie and Claude Code are a different shape and worth separating out. None of the four describes an on-premise server you deploy for a team. What they offer is execution you site yourself, on a developer’s own machine or in your own cloud account, pointed away from the vendor’s hosted service. That clears this census’s bar, and it does not carry the same compliance weight as the five above. A procurement team should not read these four rows as an on-prem offering.

Claude Agent SDK is a third shape. Self-hosting is its primary, free path, yet it is still counted closed-source here, because Anthropic’s own docs place overall SDK use under its Commercial Terms of Service and the compiled Claude Code CLI binary the wrapper bundles is not published as source, even though the thin Python and TypeScript wrapper packages carry an MIT license. It clears this census’s self-hosting bar while missing the open-source census’s stricter one, which is the clearest proof on this page that the two questions really are independent.

Method

We read the Deployment field on all 94 published product listings in this index on 27 August 2026, pinned to collector run 221173. The field is populated for all 94, so there is no unknown bucket and no count above is a percentage of a subset. The index held 95 published listings that day; the 95th is Agentix Labs, our one agencies listing, which sells engineering services rather than software with a deployment mode of its own, so it is left out rather than counted. Open-source status is a separate field. It is determined for 93 of the 95 published listings; Agentix Labs and Lead Scorer, added to the index after our last full research pass, are currently undetermined.

The judgement the deployment field encodes is this: can the agent runtime execute on infrastructure you control, without a mandatory call to the vendor’s own hosted service? Where the answer is no, the listing reads Cloud even when parts of the stack are self-hostable. Cartesia’s speech models can run in your environment while its Line agent runtime cannot, so Cartesia reads Cloud. Muse Code runs in your terminal but reaches Meta’s Model API by its only access path, so it reads Cloud too. Where the agent can run without the vendor’s service, as with Junie pointed at a local model runtime, the listing reads Both. That test is the reason a handful of tools sit in a different bucket than their marketing implies, in both directions.

Six listings separate this reading from the 19 August one our earlier figures were pinned to, and they account for the whole difference. Aider and OpenCode moved from self-host only to both, so a vendor-hosted option is now recorded for each. Claude Code moved from cloud to self-host on 24 August while staying closed-source. Clay left the index, and Lead Scorer and VoiceAgent joined as cloud-only, which is why Sales & marketing agents now reads 8 of 8 rather than 7 of 7. Cloud-only stands at 52 either side of that window; self-host only falls from 14 to 13 and both rises from 27 to 29.

Counts are a census of this index, not a survey of the market. They describe the 94 tools we have researched and published, and the index grows, so the date above is part of the claim. Corrections are welcome through our methodology page.