Self-hosted AI agents for banks and healthcare, on-prem or VPC
Run AI agents on-prem or in your own cloud with no data leaving your network: RBAC, audit and identity for banks and healthcare, using AgentField.
A bank or healthcare company can run AI agents entirely inside its own network by self-hosting the agent runtime, its database and the model endpoint, so prompts, records and audit data never leave the perimeter; AgentField is an open-source (Apache 2.0) control plane built for that, with the control plane, agents and PostgreSQL in your VPC or data centre, and any model you point it at, including one you host. What AgentField does not do is make you compliant: it supplies controls and evidence, and your security and compliance teams still own the rest.
At a glance
| Requirement | AgentField | What you still own |
|---|---|---|
| Data stays in your network | Control plane, agents and PostgreSQL run on your infrastructure; model calls go only to the endpoint you configure | Choosing a model endpoint inside the perimeter; network egress rules |
| Your own models | Python SDK routes through LiteLLM; api_base points at any endpoint, for example Ollama at http://localhost:11434 | Hosting and validating the model |
| Agent permissions (RBAC for agents) | Tag-based allow and deny policies, function-level, with argument constraints | Writing the policies; fail-open default must be tightened |
| Human access | Bearer API key; separate admin token for policy endpoints | SSO and user role mapping (not documented) |
| Audit | Workflow DAG, tagged notes, Ed25519-signed execution credentials, offline verification | Retention, log shipping, review process |
| Identity | W3C DIDs per agent and function; AES-256-GCM keystore on local disk | HSM or KMS (not supported yet) |
| Human approval | app.pause() with signed webhook resume | The review tool and who may approve |
| Certifications | None (no SOC 2 or HIPAA attestation) | Your own audit, BAAs with any model provider |
Deploy inside the perimeter
The deployment docs run the same binary on a laptop and in production; production switches storage to PostgreSQL:
af server --storage-mode=postgres \
--postgres-url="postgres://user:pass@db-host:5432/agentfield?sslmode=require"On Kubernetes, use the documented Deployment with probes on /api/v1/health, and PostgreSQL for more than one replica. Agents register with the control plane over HTTP (AGENTFIELD_SERVER), so they can run in the same cluster, on separate hosts or in other private networks you route to.
Point models at an endpoint you control. From the models docs:
from agentfield import Agent, AIConfig
local_app = Agent(
node_id="local-agent",
ai_config=AIConfig(
model="ollama/llama3",
api_base="http://localhost:11434",
),
)The same table lists Azure OpenAI (azure/gpt-4o) and AWS Bedrock (bedrock/...) through LiteLLM, if your approved model lives in your cloud tenancy rather than on your own GPUs.
Two outbound connections to know about. The README says AgentField sends privacy-minimized installation and lifecycle telemetry by default, with no prompts or execution payloads; set AGENTFIELD_TELEMETRY_ENABLED=false to turn it off. And the default coding harness, CodeAF, may download about 70 MB on first use, per the quickstart; pre-install it with af codeaf ensure on a machine with access, or skip it at install with --no-codeaf. An official offline installer for AgentField is not documented.
The production hardening the docs ask for
From the deployment checklist:
- Switch to PostgreSQL storage.
- Set a strong
AGENTFIELD_AUTHORIZATION_ADMIN_TOKEN. - Enable DID authorization (
features.did.authorization.enabled: true). - Set
tag_approval_rules.default_modetomanual, so no agent can claim a sensitive tag without an administrator. - Configure an observability webhook for external monitoring.
- Configure liveness and readiness probes.
- Terminate TLS at a load balancer or reverse proxy.
Add one more for regulated work: access policies are fail-open when no policy matches, so write explicit deny rules for anything that touches money, PHI or PII. The governance guide shows the policy format.
Audit a regulator can check
Each execution can produce a verifiable credential holding the agent's DID, the caller's DID, SHA-256 hashes of input and output, status and an Ed25519 signature. Export a workflow's chain with GET /api/ui/v1/workflows/{workflowId}/vc-chain and verify it offline with af vc verify audit.json. Because the credential stores hashes rather than the data, you can prove what a decision was based on without placing the record itself in the credential. Details are in identity and audit trails.
When something else fits better
- Your bank has standardised on AWS managed services. Amazon Bedrock AgentCore runs agents serverless in AWS with its own identity service; it is AWS-hosted, not on-prem. Note that the older Bedrock Agents is closed to new sign-ups.
- Your agents are LangGraph graphs. Self-hosted LangSmith, including Agent Server, is available as an Enterprise plan add-on and keeps the LangGraph tooling.
- Business teams build the agents. Dify self-hosts with Docker Compose; read its modified Apache 2.0 licence first.
- Durable business processes are the core need. Temporal self-hosts under MIT and is the most established durable-execution engine.
For the open-source landscape, see open-source control planes for AI agents.
Questions people ask
Is there an open-source agent platform that runs on-prem with no data leaving our network?
Yes. AgentField's control plane, SDKs and agents are Apache 2.0 and run on your own servers with PostgreSQL. With a model you host, such as one served by Ollama, model calls stay inside too; turn off the default lifecycle telemetry with AGENTFIELD_TELEMETRY_ENABLED=false.
Does AgentField have RBAC and audit logging for a bank?
It has role-style permissions for agents: tag-based allow and deny policies on functions, with argument limits such as amount < 10000, enforced before execution. Audit comes from the workflow DAG, tagged notes and signed execution credentials. Human user roles through SSO are not documented in the open-source control plane.
Is AgentField HIPAA or SOC 2 compliant?
AgentField holds no SOC 2 or HIPAA attestation, and no software is compliant on its own. It gives you controls and evidence (self-hosting, access policies, signed audit trails) that support your own assessments; you still need your own controls, a BAA with any outside model provider that sees PHI, and an auditor.
Can we run it in our own AWS, GCP or Azure account?
Yes, as a self-hosted service. The docs cover AWS ECS or Fargate, Google Cloud Run with Cloud SQL, Docker and Kubernetes. There is no Azure-specific page, but the same container image and Kubernetes manifests apply.
Where are agent private keys kept?
In an AES-256-GCM encrypted keystore on the control plane's local disk (./data/keys/, directory permissions 0700). HSM and cloud KMS backends are not supported yet, which some banks will need to plan around.
Sources
- AgentField docs, checked 2026-11-21: deployment and production checklist, models, access policies, identity, verifiable credentials, REST API authentication, quickstart
- Agent-Field/agentfield README, checked 2026-11-21 (telemetry, installer flags)
- Amazon Bedrock AgentCore, Bedrock Agents, LangSmith self-hosted, Dify Docker Compose, Dify licence, Temporal repo, checked 2026-11-21
Checked 2026-11-21 against the AgentField docs and README and each vendor's own documentation on that date.