Research report · data-report
Self-hosted: a census of 110 listings
30 of 110 listings answer yes for “Self-hosted”. Counted from our own datasheet, with the query that reproduces every figure.
On this page
Self-hosted, across 110 settled listings
Of the 110 Published listings whose datasheet settles “Self-hosted”, the largest group is “No” at 41 (37%). Measured 2026-09-07.
Explore the matching agents
Self-hosted · Datasheet criterion self_hosted · Report counted .
Current listings as of ; this set may differ from the report's original sample. Includes “Yes” and “Qualified” answers only. Read the conditions on qualified answers before treating them as a match. “No”, “Not applicable” and “Not established” answers are excluded.
View 66 current matching records
- AgentBridge Yes
Local-only by design: an npm-installed CLI and a background daemon on your machine, with no hosted component involved.
Source retrieved 2026-09-02
- Agent Development Kit (ADK) Yes
Runs wherever containers run, including Docker or Podman locally and systems with no connection to Google Cloud; GKE and Cloud Run are options, not requirements.
Source retrieved 2026-09-02
- Agentix Labs Qualified
Client-owned deployment is offered but explicitly conditioned on the architecture, so running it on your own infrastructure is not guaranteed per engagement.
Source retrieved 2026-09-09
- AgentsKit.js Yes
Everything runs on your infrastructure by definition: the runtime is process-agnostic and documented for servers, edge workers, terminals, browsers and Electron hosts.
Source retrieved 2026-09-09
- Agno Yes
The runtime always runs in your own cloud; self-hosting the control plane itself is an Enterprise add-on.
Source retrieved 2026-09-02
- Aide Qualified
Not customer-hosted software: the default is Aide's Azure tenancy, with private-cloud provisioning offered only where residency rules force it.
Source retrieved 2026-09-03
- Aider Yes
Runs entirely on your own machine in its own python environment against your local git repo.
Source retrieved 2026-09-02
- Aisera Agent Studio Yes
Unusually for this category, on-premises is offered outright, alongside customer VPCs and hybrid clouds on all three major providers, as part of the TRAPS governance framework.
Source retrieved 2026-09-03
- Augment Code Qualified
Agents can execute on your own machines via daemons, but inference and the control plane stay with Augment.
Source retrieved 2026-09-02
- AutoGen Yes
Self-hosting is the only option: pip-installed packages, a standalone or distributed runtime across your own machines, and Studio backed by your SQLite or PostgreSQL database.
Source retrieved 2026-09-02
- Bitsclan IT Solutions Qualified
One published engagement runs the models on the client’s own machines; self-hosting is a demonstrated build option, not a stated default of the agent service.
Source retrieved 2026-09-02
- Bland Qualified
Your VPC, on-premise and air-gapped deployments exist, but self-hosting is an Enterprise option.
Source retrieved 2026-09-02
- Bolt Qualified
Enterprise BYOK deployment runs Bolt in your own AWS or Azure tenant; standard plans are vendor-hosted only.
Source retrieved 2026-09-02
- Cartesia Qualified
On-premise or VPC model deployment is sales-gated; separately, agent reasoning code can run on your own servers with Cartesia connecting to your URL.
Source retrieved 2026-09-02
- Claude Agent SDK Yes
Self-hosting is the default: your process spawns a claude CLI subprocess with your working directory and disk.
Source retrieved 2026-09-02
- Claude Code Qualified
The CLI executes on your own machine, and cloud sessions can run on your runners in public beta on Team and Enterprise; inference always goes to Anthropic or a cloud provider.
Source retrieved 2026-09-02
- Cline Yes
Client-side execution on your machine, and Enterprise deploys into your VPC, on-prem or air-gapped.
Source retrieved 2026-09-02
- CodeRabbit Qualified
Self-hosting is Enterprise-only and gated on 500+ seats; it ships as a container image you run yourself.
Source retrieved 2026-09-02
- Codex CLI Qualified
The agent binary runs on your machine or runner; inference is OpenAI-hosted unless you route it through your own AWS Bedrock account.
Source retrieved 2026-09-02
- CrewAI Qualified
Own-VPC and own-infrastructure deployment (Factory, Helm-configured) is an Enterprise option.
Source retrieved 2026-09-02
- Cursor Qualified
Only tool execution moves to your machines: Cursor still runs the agent loop, inference and planning, and the Self-Hosted Pool needs Enterprise. The Enterprise FAQ says there is no on-premises deployment.
Source retrieved 2026-09-02
- Deepgram Voice Agent API Qualified
Self-hosted containers are an Enterprise-only path. Deepgram's own self-hosting pages enumerate speech-to-text, text-to-speech and audio intelligence, not voice-agent orchestration, while the Voice Agent product page lists self-hosted as a deployment option. Dedicated single-tenant hosting does explicitly cover agent workloads.
Source retrieved 2026-09-02
- Devin Qualified
Outposts runs session execution on machines you operate, but the agent loop stays in Cognition's cloud.
Source retrieved 2026-09-02
- Devin Desktop Qualified
Enterprise-only, and the deployment docs describe Cognition hosting in a customer-isolated single-tenant VPC reached over AWS PrivateLink rather than customer-run software.
Source retrieved 2026-09-03
- Dify Yes
Community Edition runs anywhere Docker runs; Enterprise adds Kubernetes/Helm, VPC and air-gapped deployment.
Source retrieved 2026-09-02
- ElevenLabs Agents Qualified
The full agent runtime can run in your own VPC or hardware on AWS or GCP, but the documentation is NDA-gated and sold through the account team; Azure is unreleased.
Source retrieved 2026-09-02
- Elicit Qualified
Enterprise offers custom deployments including single-tenancy; Elicit publishes no on-premises, VPC or air-gapped option.
Source retrieved 2026-09-02
- Factory Qualified
The agent always runs on your machine, and a dedicated airgap build exists, but airgap builds are Enterprise-only and not publicly distributed.
Source retrieved 2026-09-02
- Firecrawl Qualified
A pinned Docker Compose release runs scrape, crawl, map and search on your own machine, and the CLI and MCP server can be pointed at it. But the vendor's own table excludes Agent, Interact, screenshots, page actions, the dashboard and the Fire-engine anti-bot path from that stack.
Source retrieved 2026-09-03
- Genspark Qualified
Claw's desktop app runs the agent on your own machine (inference still comes from Genspark's cloud on credits); the workspace itself is SaaS, with a dedicated VPC only on Enterprise.
Source retrieved 2026-09-02
- Glean Qualified
Customer Hosted runs Glean inside your own AWS or GCP account, but as a Glean-operated managed service you may not deploy or patch yourself.
Source retrieved 2026-09-02
- Google Antigravity CLI Qualified
The agy binary runs on your own machine and enterprise inference can run in your Google Cloud project behind VPC Service Controls, but there is no on-premise model serving.
Source retrieved 2026-09-02
- goose Yes
Self-hosting is the only deployment model: it runs on your machine, in Docker, or as goose serve on your own VM. There is no vendor-hosted instance.
Source retrieved 2026-09-02
- GPT Researcher Yes
Self-hosting is the only deployment: pip package, FastAPI server, or docker-compose on your own machine or container platform.
Source retrieved 2026-09-02
- HOL Guard Yes
Enforcement is designed to live on your own machine and works air-gapped; only the optional shared control plane is hosted, and Enterprise offers an on-premises Guard Cloud.
Source retrieved 2026-09-07
- IrisAgent Yes
IrisAgent Private installs the whole stack in your data centre or your own VPC, including an air-gapped variant with no outbound connectivity; it is software only, no appliance.
Source retrieved 2026-09-03
- Juggler Yes
It only runs on your own machine: a single binary, localhost-only by default, with a documented headless Linux and Docker path.
Source retrieved 2026-09-02
- Junie Qualified
Inference can be fully self-hosted (bundled local engine, Ollama/LM Studio/LiteLLM, or a self-hosted proxy); the agent binary itself is JetBrains-distributed.
Source retrieved 2026-09-02
- Kilo Code Qualified
The MIT clients can be forked and self-hosted and inference can be fully local, but Gateway, Cloud Agents and team administration are Kilo-hosted.
Source retrieved 2026-09-02
- LangChain Qualified
Platform self-hosting and hybrid are Enterprise-only; Developer and Plus are Cloud, while the open source packages run wherever you install them.
Source retrieved 2026-09-02
- Langflow Yes
Self-hosting is the only hosting the docs describe: a published Docker image, a Helm chart for production Kubernetes, or the Python package.
Source retrieved 2026-09-03
- LangGraph Yes
The library is MIT-licensed and installed with pip or npm, so it runs wherever you install it; self-hosting the managed Agent Server platform is a separate Enterprise hosting option.
Source retrieved 2026-09-02
- Letta Yes
Local runtime keeps all state on-device with no account, or run the App Server via Docker, Railway or Fly.io.
Source retrieved 2026-09-02
- LiveKit Agents Yes
Agents are containers you can run on your own Kubernetes; the LiveKit media server and SIP service are separately self-hostable. Self-hosted agents still meter WebRTC participant minutes if they connect to LiveKit Cloud.
Source retrieved 2026-09-02
- LlamaIndex Qualified
Full platform deploys into your own Kubernetes/VPC, but only on Enterprise plans.
Source retrieved 2026-09-02
- Mastra Yes
mastra build emits a self-contained Node server that runs on any Node/Bun/Deno host; Enterprise Edition can be deployed on-prem or in your own cloud.
Source retrieved 2026-09-02
- Microsoft Agent Framework Yes
Self-hosting is one of the two documented hosting models; your application owns routing, identity, authorization, storage, deployment and scaling. The self-hosting packages are prerelease in both Python and .NET.
Source retrieved 2026-09-02
- MindStudio Qualified
Self-hosting the platform is a Business-plan deployment option sold by quote; Individual and Free are hosted only. Self-hosted models are separate and available on every plan.
Source retrieved 2026-09-02
- Muse Code Qualified
The binary runs on your own macOS or Linux machine, but inference is Meta's hosted API and requires a Meta credential; no self-hosted serving path is documented.
Source retrieved 2026-09-02
- n8n Yes
Docker, npm or Helm on your own boxes, air-gapped if required; the free Community edition needs no licence key, and Business or Enterprise features unlock with one.
Source retrieved 2026-09-02
- OpenAI Agents SDK (Python) Yes
The SDK runs in your own Python process, local runtime tools execute in your environment, and sandbox sessions can be Unix-local or Docker on your own machine.
Source retrieved 2026-09-02
- OpenClaw Yes
Self-hosting is the only deployment model: one Gateway process on your laptop, a Mac mini, a VPS, Docker, Kubernetes or a Raspberry Pi.
Source retrieved 2026-09-02
- OpenCode Yes
The agent runs entirely on your own machine or runners, talking directly to whichever provider or internal gateway you point it at.
Source retrieved 2026-09-02
- OpenHands Yes
Runs locally for free, and Enterprise installs into your own Kubernetes cluster or VPC with no shared tenancy.
Source retrieved 2026-09-02
- Pydantic AI Yes
It is a library in your own process, so the runtime is yours; the Gateway and Logfire also publish self-hosted deployments if you use them.
Source retrieved 2026-09-02
- Retell AI Qualified
On-prem deployment is an Enterprise offering; on pay-as-you-go the only self-hosted piece is your own LLM behind the custom-LLM WebSocket, or your own SIP telephony.
Source retrieved 2026-09-02
- Rox Qualified
'In VPC Deployment' is checked only in the Enterprise column: the warehouse stays the system of record and data need not leave the customer VPC, but compute runs in Rox's VPC, so this is not a customer-run install.
Source retrieved 2026-09-03
- SafeNet Creations Qualified
Hosting is deliberately left outside the project price and settled per proposal, so running the result on your own infrastructure is negotiable rather than published.
Source retrieved 2026-09-04
- Sim Yes
Docker Compose and Kubernetes deployments are documented and free; the pricing table nevertheless lists 'Self Hosting' only under Enterprise.
Source retrieved 2026-09-03
- TinyFish Qualified
Cloud-hosted by default; VPC deployment is an Enterprise item and on-premises needs a separate written agreement.
Source retrieved 2026-09-03
- UiPath Agent Builder Qualified
Self-hosting is real but doubly bounded: Agent Builder is named as included in Automation Suite on AKS, EKS or OpenShift including air-gapped, while the Automation Suite on Linux flavour has agentic AI only planned, and the Basic plan carries an em dash for deploying in your own environment.
Source retrieved 2026-09-03
- Vapi Qualified
Self-serve is hosted only; on-premise is offered to large enterprises, and the orchestration layer otherwise never leaves Vapi.
Source retrieved 2026-09-03
- Vellum Yes
Run the runtime locally on a Mac, on a GCP VM in your own project, or on any SSH-reachable Linux host, with no platform or machine-tier fee.
Source retrieved 2026-09-03
- Warp Qualified
Enterprise only, and execution plane only: agents run on your machines while orchestration, session management and inference still route through Warp's backend.
Source retrieved 2026-09-03
- Zed Qualified
The editor runs on your machine and inference can run on hardware you control; Zed-hosted models, collaboration and the org dashboard stay on Zed's infrastructure.
Source retrieved 2026-09-03
- zot Yes
Self-hosting is the only mode: a static binary on your machine, using your credentials, in your working directory. There is no vendor-run backend.
Source retrieved 2026-09-03
The population
110 listings in this index are Published. This is the population as of 2026-09-07. 110 of the 110 published listings in this index carry a settled answer for “Self-hosted”. That row asks: Can you run it on your own infrastructure? Every listing in the index is settled on this row, so nothing is left out.
Findings
-
30 of the 110 settled listings answer “Yes” for “Self-hosted”. That is 27% of the settled set.
-
36 of the 110 settled listings answer “Qualified” for “Self-hosted”. That is 33% of the settled set.
-
41 of the 110 settled listings answer “No” for “Self-hosted”. That is 37% of the settled set.
-
3 of the 110 settled listings answer “Not established” for “Self-hosted”. That is 3% of the settled set.
1 of those yes answers is goose, which still answers yes for “Self-hosted”. Its listing carries the stored answer and links to the source recorded for it. This is one worked example, not an independent audit of every source in the census.
What we counted, and how
Each figure above is a count over the “Self-hosted” row of the listing datasheet, taken from the same stored answer the listing page renders. The denominator is the 110 listings whose answer is settled, meaning one of yes, qualified, no, does not apply, not established. On this row that is the whole index, because every listing has a recognised stored answer. Every number here is stored with the read-only query that reproduces it and re-run every sixty seconds against the live corpus, so a figure that stops reproducing surfaces as drift rather than as a stale sentence nobody notices. Our full method covers how a datasheet row is settled in the first place.
Limitations
This counts stored datasheet answers, not independently tested capabilities. A sourced answer can record a vendor statement or our reading of published evidence; this census does not re-fetch those sources. “Not established” means we have not established an answer. That can reflect vendor nondisclosure, blocked evidence, or unfinished research, not a no. An absent or unrecognised answer is excluded rather than treated as a researched finding. The figures are restated when the stored corpus changes; the date above is the count used for this published version, not a new verification of the vendors.
Get the next report
New agents rankings and fresh data reports. One short email, one-click unsubscribe.