Self-hosted AI agents keep the agent runtime on infrastructure you operate: the loop that decides the next step, the workers that run tools, the sandboxes those tools sit in, the secrets vault, the session store, and the identity console, while the model can still sit outside that perimeter. Hosting the runtime is not the same as hosting the weights, and it is not a promise that every token stays on your network.
Key facts
Definition
Self-hosted AI agents keep the agent runtime on infrastructure you operate: the loop, tool workers, sandboxes, secrets, session history and identity, while the model can still sit outside that perimeter.
How Bigcircle does it
- On Solarpunk, Bigcircle keeps credentials in an encrypted vault on the user's machine and injects them at the call, so the model is a reasoning service and never holds the keys.
- On Skionis, Bigcircle treats the agent as untrusted and reaches the provider only through purpose-built tools, with no general shell and no path from a plan step to an arbitrary API call.
- On the insurance voice agent, Bigcircle puts permission checks in the tool contract, so the model's output is a request to call a tool and never an action by itself.
Comparison
A self-hosted LLM is about where the model runs, while self-hosted AI agents are about where the loop, workers, sandboxes, secrets and sessions run.
Key number
On Solarpunk, plans completed with no person rose from about 55% to 80% once local retrieval and planner memory grounded the agent, in our own measurement during the build.
We already draw that line on the agents we ship. On Solarpunk, a desktop agent that reaches more than a hundred tools, credentials, the tool registry, retrieval and execution live on the user's machine, and the model is called as a reasoning service that never holds the keys.
On Skionis, the DevOps agent on a live AWS account, the agent is untrusted: it reaches the provider only through purpose-built tools, with no general shell and no path from a plan step to an arbitrary API call.
The question buyers ask after a vendor ships a self-hosted option is usually the wrong one. They ask whether the agent is "on prem". The useful split is which pieces you run, which pieces still leave on each model call, and whether you are ready to operate a server, a database and a worker fleet per trust boundary. Where the model itself runs is a separate decision, covered on self-hosted models, and this guide stays on the runtime.
This guide starts with what a self-hosted AI agent actually means and the three deployment shapes. It then covers what still leaves when the model is external, what you have to run yourself, when self-hosting is worth it against a managed agent platform, and how to secure the runtime. It ends with how Bigcircle already places the vault, the tool boundary and the trace on the agents it builds.
What are self-hosted AI agents?
Self-hosted AI agents are agents whose control plane runs on infrastructure you operate. You host the API that your apps call, the loop that asks the model for the next step, the store for agents and sessions, the workers that execute shell and file tools, and usually the console that uses your identity provider. You may still send every model call to a hosted API, to a named provider such as Bedrock, Google Cloud or Microsoft Foundry, or to a gateway you already run.
That is different from a self-hosted LLM. Self-hosting the model answers where the weights sit and which GPUs answer tokens. Self-hosting the agent answers where the loop, the tool workers, the secrets and the session history sit. You can do either, both, or neither, and mixing the two phrases is how teams buy the wrong project. Three shapes cover most of what vendors now offer, and the figure is the framework rather than a product diagram.
| Shape | What you host | What the vendor still runs | What still leaves |
|---|---|---|---|
| Vendor-hosted | Usually nothing of the control plane | Loop, sandboxes, sessions, credentials | Prompts, tool I/O and history under the vendor's retention terms |
| Self-hosted sandboxes only | Tool workers and the containers or VMs they run in | The loop and the session store | Observations still return to the vendor loop so it can choose the next step |
| Self-hosted runtime | Loop, database, workers, sandboxes, IdP console | Often only the model, if you leave it external | Prior turns, tool arguments and tool returns, on each remote model call |
Treat that table as a map of who runs what, not as a ranking. Keep the three rows on Anthropic's self-hosted overview, checked 16 November 2026, and the private-beta label on the 12 November 2026 article. Use that last row when you want the self-hosted runtime: you operate the service and the sandboxes, and you still pick where the model lives, at Anthropic or at a provider you choose. Read those two pages for the rows and the beta label.
What data still leaves your network when the model is external?
Self-hosting the runtime does not stop the model call. Treat each hop as a copy of the work to date: prompt text, tool arguments and tool returns, all of it going to the inference provider you pointed at. Anthropic's data-and-access page for the self-hosted product, checked 16 November 2026, lists that payload on every call.
If the Claude API answers the model, the request also carries session, agent and environment identifiers your server generated. They hold no person's name and no transcript text. The identifier still carries a creation time, which is enough for a usage log even when the payload looks empty of content.
Ask which disk holds the transcript before you treat zero data retention as the residency answer. If your organisation already has that arrangement with the provider, it applies to these hops as it applies to the rest of your Messages API traffic. The prompt still left. On a self-hosted runtime the session record sits in your database. When the vendor runs the control plane, that record sits on the vendor's disk. Read the provider's current retention terms for the endpoint you actually call.
The server's allowlist is not the sandbox's. Put the egress rule on the worker host, not only on the API server. A shell command inherits the worker process's network, so a sandbox you left open can call hosts the control plane never listed, and a result that looked local is already on someone else's wire.
If the data in the prompt or in a tool result must never leave, the model has to run where that data already sits. That is the self-hosted models decision, measured against a hosted baseline on cost, latency and quality before anyone commits GPUs. Do not pretend a self-hosted runtime alone solved residency.
What do you have to run and operate yourself?
Staff it as a product you run, not as a flag you flip. Budget PostgreSQL 16 or later for agents and session history, plus worker processes that pull jobs and execute shell and file tools inside sandboxes you provide. The remaining pieces are an API process that also owns the agent loop, and a console bound to your identity provider. Anthropic's self-hosted overview, checked 16 November 2026, names that set for its stack, and other platforms rhyme.
Skip the database and the process holds agents and sessions in memory, so a restart drops them. A hosted product ships upgrades and watches for odd traffic. On your stack you take both jobs: you apply every release, you lock down the server and the memory stores, and you watch for usage that does not fit the job. Check the vendor's differences page before you treat the self-hosted option as the same product with a different bill. Some features are currently available only on the hosted service.
Split trust with separate deployments, not with named environments on one stack. Give the untrusted workload different workers, a different database and a different server API key. One key in one deployment can call every environment that deployment holds. Sharing one stack across trusted and untrusted code is how a low-trust session inherits a high-trust network.
A shell tool is the worker process's shell and it has that process's permissions. Isolation comes from the container or VM around the worker, not from the tool. If you start the worker yourself, run it on a container or VM built to be deleted, and store nothing there you would miss. That matches how we treat Skionis: no general shell in the action set, because a plan step must not become an arbitrary API call.
When is self-hosting worth it versus a managed agent platform?
We start from the ops load, not from the marketing page. Take the full runtime when agents, sessions and history have to live on disks you control, or when inference has to hit a named provider (such as Bedrock, Google Cloud or Microsoft Foundry) or your own gateway. Take it when the tools have to sit next to systems that will not accept a vendor worker. If the review only cares that code execution stays on your hosts, stop at self-hosted sandboxes and leave the loop with the vendor.
Pick a managed agent platform when you will not staff PostgreSQL, workers, upgrades and a console IdP for every trust boundary. Budget people for release installs, database backups, worker capacity, IdP mapping, secret-manager wiring and a separate deployment per trust level before you treat the self-hosted option as the lower cash bill.
If your security review only needs tool code to stay inside the VPC, pay for sandboxes and leave the loop hosted. If the review needs the session store and the vault off the vendor's disk, take the full runtime.
November 2026 made that buy-versus-build choice louder across vendors, from Mirantis's Kubernetes offering and IBM and Microsoft's Imagine Enterprise through Anthropic's self-hosted private beta. Read each company's own release in Further reading, and do not treat this page as a shortlist.
How do you secure a self-hosted agent runtime?
Start with the tool set: on Skionis the agent reaches AWS only through purpose-built tools, and there is no general shell. On Solarpunk the agent can act only through systems someone registered, and credentials are injected into the tool adapter at execution time so the model never holds them. Carry those habits into a self-hosted runtime: enumerate the actions, keep secrets outside the sandbox, and refuse a path that turns a plan into an arbitrary call.
Then add the controls the neighbours already cover, and do not rebuild them here. AI agent sandboxing is the isolation, egress allowlist and stop rules once the agent can run its own code or open a browser. AI agent permissions binds each action to a person, a target and a check in code.
An AI gateway sits on model traffic when more than one agent or team shares keys, budgets and routing. An AI audit trail keeps the evidence version, the rule and the review. An AI agent inventory is the live list of every agent with an owner and a last-acted time, and an AI agent kill switch is the flag that refuses the next model or tool call without a redeploy.
Give every session its own sandbox, and do not place the environment key where the session's code can read it. A console role only binds the signed-in person's own requests. It does not bind code that finds a credential already sitting in the sandbox. For MCP and the API calls the product supports, attach the secret on a proxy outside that sandbox, the same way we keep the Solarpunk vault on the user's machine.
How Bigcircle already places the vault, the boundary and the trace
We do not ship a product named self-hosted agent platform. The agents we build already keep the runtime decisions this page argues for: where credentials live, which tools exist, and what a run leaves behind.
On Solarpunk, credentials sit in an encrypted vault on the user's machine and are injected at the call. We stopped giving the model the registry, and each step queries it for the few contracts it needs. A trace per run goes from the goal through each call to the final application state, so claimed completion can be compared with what actually changed in the target system.
Plans completed with no person rose from about 55% to 80% once local retrieval and planner memory grounded the agent, in our own measurement during the build.
On Skionis, an engineer approves an architecture diagram before a change applies, and one record links the request to every resulting resource. Template errors before deploy fell from around 10% to 4% across the validation loop's introduction, again by our own measurement, and delivery took roughly 4 months to a working product.
On the insurance voice agent, permission checks and required disclosures sit in the tool contract, so the model's output is a request to call a tool and never an action by itself. A failing request escalates and does not degrade into something close. On a current engagement for a telecom marketing platform, we scope each agent to the entitlements of the person it acts for, and prompts stay a user-experience layer that no policy check depends on.
When a second team or a second trust boundary arrives, those same fields move into a runtime you operate: a vault outside the sandbox, a registered tool set, a session store you control, a gateway in front of model calls if several agents share keys, and a kill switch on the request path. The model placement stays a separate eval against self-hosted models.
Common questions
What is a self-hosted AI agent?#
An agent whose loop, tool workers, sandboxes, secrets, session history and identity console run on infrastructure you operate. The model can still be a hosted API, a named provider, or a model you host yourself. Self-hosting the agent runtime answers where the control plane sits, not where the weights sit.
How is a self-hosted AI agent different from a self-hosted LLM?#
A self-hosted LLM is about where the model runs and which GPUs answer tokens. A self-hosted AI agent is about where the loop, workers, sandboxes, secrets and sessions run. You can self-host the agent and still call a hosted model, or self-host the model behind a vendor-hosted agent. The self-hosted models service covers the model side.
Can AI agents run fully on premise?#
Yes, if you host both the runtime and the model on infrastructure you operate, and you keep tool egress inside that perimeter. A self-hosted runtime with an external model still sends a copy of the work to date: prompt text, tool arguments and tool returns, on every model call. Fully on premise means both decisions, not only the agent server.
What is a self-hosted agent platform?#
Software you install that runs the agent control plane on your hosts: the API, loop, session store, workers and usually policy, audit and identity. Mirantis's Kubernetes product and Anthropic's self-hosted private beta are two shapes of that decision. Pick by storage, inference routing and how much ops load you will carry, not by a ranked list.
What are private AI agents?#
In this guide, private means the runtime you operate: sessions, secrets, sandboxes and identity stay on your infrastructure. It does not automatically mean private inference. If the model is external, prompts and tool results still leave on each call. Treat "private AI agents" as a runtime claim until the model placement is named.
Which open-source options exist for a self-hosted personal agent?#
n8n Labs released Nod on 10 November 2026 as a research preview: a standalone personal agent you install on your computer or a server you host, with the model provider you choose, under Apache 2.0. Nod's own docs say it belongs to one person, keeps files and saved logins on the host, and sends the model provider what each task needs. That is Nod's account of itself, not a ranking against enterprise platforms.
Do self-hosted sandboxes keep tool data off the vendor?#
No. With self-hosted sandboxes only, tool inputs and outputs still flow back to the vendor-hosted loop so the model can decide the next step. You keep code execution and files on your hosts. You still send the observations the loop needs unless you also self-host the service.
Further reading
- Anthropic, Self-hosted Claude Managed Agents, checked 16 November 2026. Three deployment shapes, components you run, bash isolation, what leaves on each model call.
- Anthropic, Data and access in self-hosted Claude Managed Agents, checked 16 November 2026. Egress list, conversation contents on each call, ZDR, one deployment per trust boundary.
- Anthropic, Operate Claude Managed Agents inside your security perimeter, and Claude Platform release notes for 12 November 2026. Private beta announcement; harness, sandboxes, credentials and session history on your infrastructure.
- Anthropic, Differences from the hosted service, checked 16 November 2026. Features only your server has, and features only the hosted service has.
- n8n Labs, Meet Nod, 10 November 2026, and Host your own server, checked 16 November 2026. Personal agent on your machine or server; model of your choice.
- Mirantis, Mirantis Announces Agent Platform, 10 November 2026. Self-hosted governed agents on Kubernetes.
- IBM, IBM and Microsoft Announce Imagine Enterprise, 10 November 2026. Managed agent platform with self-management options.
- Self-hosted models, AI agent sandboxing, AI agent permissions, AI gateway, AI audit trail, AI agent inventory, AI agent kill switch and computer use.
- The Solarpunk build, the Skionis build and the insurance voice agent build.