Agent Access01 / 06
Scoped agent access
Identify agents and applications through revocable credentials associated with owners, teams, environments, and access policies.
Scoped access contextFlagship platformSelf-hosted · customer-managed
Govern every model request.
Agent Access Manager is the self-hosted control layer between your AI applications and model providers. Replace shared provider keys with scoped virtual access, enforce runtime policy, and keep a searchable record of governed activity.
RuntimePolicy Enforcement
ProtectedProvider Credentials
SearchableActivity Records
const client = new OpenAI({
baseURL: "https://gateway.acme.com/v1",
apiKey: process.env.VIRTUAL_KEY,
});
const reply = await client.chat.completions.create({
model: "fast", // governed model alias
messages: [{ role: "user", content: prompt }],
});OpenAI-compatibleDocker & KubernetesOffline installation
The control plane
One governed layer between agents and enterprise AI infrastructure.
Applications authenticate with scoped, revocable virtual keys while provider credentials stay encrypted inside your environment. No reusable vendor secret needs to live in agent code.
Every model request can be evaluated against identity, team, model, budget, rate, and guardrail policy before it is forwarded to a configured provider.
Security and platform teams get a consistent operational record across model requests, guardrail decisions, usage, configured cost, and outcomes handled through the platform.
Agents & Applications
Coding agents and assistants
Internal AI applications
SDKs and orchestration frameworks
Teams, projects, and environments
Agent Access Manager
Authenticate with scoped virtual keys
Authorize model requests by scope
Route to configured providers
Enforce request and response guardrails
Apply budgets and rate limits
Record decisions, usage, and outcomes
Model Providers
Cloud-hosted LLMs
OpenAI-compatible endpoints
Anthropic deployments
Gemini and Vertex AI
Self-hosted model servers
Product capabilities
Control access, policy, routing, and evidence in one platform.
Runtime governanceModels · providers · agents
Policy follows every governed request.
Establish who is calling, which model alias they request, and what must be enforced before a provider receives the request.
Agent Access Manager brings identity context, credential protection, routing, guardrails, cost control, and security records into the same model-request path.
Scoped virtual access
Issue revocable, expiring virtual keys for applications, agents, teams, and projects without distributing provider credentials.
Credential isolation
Keep provider credentials encrypted inside the gateway and inject them only when forwarding governed model requests.
Provider and model routing
Use model aliases, health-aware routing, and failover across configured cloud and self-hosted providers.
Runtime guardrails
Inspect requests and responses for PII, secrets, patterns, and content with allow, flag, redact, or block actions.
Budgets and rate controls
Apply spend budgets and RPM or TPM limits across organizations, teams, projects, applications, and keys.
Tool-call argument screening
Flag or block configured findings in model-generated tool-call arguments before the application executes a tool.
Security analytics and containment
Search model-call and guardrail events, run detections, and apply supported quarantine or throttling actions.
Searchable security records
Investigate governed model activity with the applicable identity, provider, model, usage, configured cost, and outcome.
Least privilegeFail closedCustomer-owned data
How it works
From standing provider credentials to governed access in three steps.
Keep your current SDKs and provider accounts. Change the endpoint and credential, then manage access and policy centrally.
Register customer-owned provider credentials and define the model aliases approved for application use.
Give every application, agent, team, or project a revocable virtual key with model, budget, rate, and guardrail policy context.
Applications call one gateway endpoint. Agent Access Manager authorizes and routes requests, enforces policy, and records governed activity.
DeploymentYour cloud or data center
PackagingDocker Compose or Kubernetes
Data boundaryPrompts and records stay with you
Provider accountsCustomer-owned credentials and spend
Integrations
Bring providers and agent frameworks behind one governed endpoint.
Connect supported cloud and self-hosted models, then point compatible SDKs, frameworks, coding agents, and internal applications at Agent Access Manager.
Model providers
Route through customer-managed provider connections.
Centralize provider configuration and model aliases while applications use scoped virtual keys instead of reusable vendor credentials.
OpenAIAnthropicAzure OpenAIAWS BedrockGoogle GeminiVertex AICohereMistral AIGroqOllamaMeta LlamaIBM watsonxNVIDIADatabricksDeepSeek
OpenAI
Anthropic
Azure OpenAI
AWS Bedrock
Google Gemini
Vertex AI
Cohere
Mistral AI
Groq
Ollama
Meta Llama
IBM watsonx
NVIDIA
Databricks
DeepSeekAdditional providers can connect through supported OpenAI-compatible endpoints.
SDKs and frameworks
Keep the development tools your teams already use.
In most integrations, applications change the base URL and replace the provider key with an Agent Access Manager virtual key.
LangChainLangGraphLlamaIndexDSPyVercel AI SDKLiteLLMSemantic KernelOpenAI SDKCrewAIAutoGen
LangChain
LangGraph
LlamaIndex
DSPy
Vercel AI SDK
LiteLLM
Semantic Kernel
OpenAI SDK
CrewAINative guides are available for OpenAI, Anthropic, Gemini, Vertex AI, LangChain, and supported self-hosted models.
Frequently Asked Questions
Questions teams ask before adopting.
How Agent Access Manager governs application and agent access to models in customer-managed environments.
What is Agent Access Manager?
Agent Access Manager is a self-hosted AI gateway and governance layer that sits between your applications and your LLM vendors. Applications get a single OpenAI-compatible endpoint and revocable virtual keys; the platform routes each call to the configured provider, enforces budgets, rate limits, and guardrails at runtime, and records governed activity in a searchable audit trail. It deploys in your own environment with Docker Compose or Kubernetes.
How do AI agents access models without seeing our provider API keys?
Provider credentials are registered once by administrators and kept AES-256-GCM encrypted inside the gateway; they are never returned to clients. Each application, team, or agent instead receives a scoped virtual key that is revocable, can expire, and carries its own budgets, rate limits, and guardrail policy. Revoking a virtual key cuts off that consumer without rotating the underlying provider credential.
Which LLM providers and SDKs does the gateway support?
The gateway exposes OpenAI-compatible chat, embeddings, responses, messages, and models APIs, with provider adapters for OpenAI-compatible services, Anthropic, Gemini, and Vertex AI, plus self-hosted backends such as Ollama, vLLM, and LiteLLM. Existing applications built on the OpenAI SDK, Anthropic SDK, LangChain, or coding agents like Claude Code and Codex connect by changing the base URL and key — no application rewrite.
Can Agent Access Manager run fully on-premises or air-gapped?
Yes. Agent Access Manager is customer-managed software deployed in your own environment — single-host Docker Compose or multi-replica Kubernetes with Helm — and supports offline installation. Prompts, responses, credentials, and audit data stay inside your infrastructure, and provider usage runs through your own vendor accounts.
How does Agent Access Manager handle model-generated tool calls?
The current gateway preserves or translates supported tool and function-call fields and provides tool-call argument screening through configured response guardrails using flag or block actions. It does not currently provide an MCP server registry or per-tool execution permissions. Tool execution must be authorized and audited by the agent runtime, tool service, or a separate tool gateway.
How does it help with SOC 2 or ISO 27001 audit evidence?
Every governed request is recorded with the applicable identity, policy decision, provider, model, token usage, cost, and outcome. Audit trails are searchable, and scheduled usage and security reports export in PDF, CSV, JSON, and HTML. Built-in SIEM search, Sigma detections, behavioral analytics, and SOAR containment support security operations on the same event stream.
See your first governed model call in under an hour.
Talk to the engineers who build Agent Access Manager. We will map your providers, agents, deployment environment, and model-governance requirements to a practical control architecture.
We typically respond within one business day.