How Agent Vault works

The pieces of an Agent Vault setup and how a session gives an agent access to APIs and MCP servers.

Agent Vault lets an AI agent call APIs and MCP servers without having access to the real credentials. Instead, the agent sends its requests through an Agent Vault proxy, which attaches the credentials on the way out.

​

Architecture

To find the architecture that best fits what you’re building, plan your setup.

Every Agent Vault setup has an orchestrator that creates a session and starts the agent. Then, the agent sends every request through the proxy:

Piece	What it does	Credentials it stores
Orchestrator	Creates a session and starts the agent with the session token	Your Infisical login, or a machine identity that can create sessions
Agent	Does the task, in an environment such as a container, a virtual machine (VM), a Kubernetes pod, or a managed sandbox	The session token
Proxy	Fetches the real credentials from Infisical and attaches them to the agent’s requests, from its own container or machine	The proxy token
Infisical	Stores your services and access bundles in an Agent Vault project	The real credentials

​

Orchestrators

An orchestrator is the program (or person) that creates a session and starts an agent with the session token. Agent Vault works with any orchestrator, such as:

• Your own service: a backend service uses the Infisical API to create a session and starts an agent for each task. It can call the API directly or through the Python SDK or Node.js SDK. For example: a Slack bot that starts an agent in a new sandbox for each message, an internal coding agent that picks up tickets, or a product feature that starts an agent when a customer asks for something
• A scheduled job: a cron job or CI pipeline uses the Infisical CLI to create a session and starts the agent. For example: an agent that labels new issues every hour
• You: a person who creates a session in the Infisical dashboard and starts an agent with the session token. For example: a coding agent on a dev box, or an assistant on a home server

The orchestrator decides which access bundle each session gets, and the agent can’t change it. For example, your backend might create most sessions with a read-only access bundle. However, if an on-call engineer makes the request, the backend could create the session with an access bundle that includes your deployment API instead.

Keep the credentials that the orchestrator uses out of the agent’s environment. A machine identity can create a session for any access bundle it has access to, so if an agent read the machine identity’s credentials, it could give itself more access.

​

Session duration

How long each session lasts is up to you. A session lasts until it expires or the orchestrator revokes it, so it can last:

• Until the agent finishes a task, with the orchestrator revoking the session when the task ends
• For a fixed amount of time that you set when you create the session
• With no expiry, until someone revokes it

An agent keeps its access until its session expires or you revoke it. If a session token leaks, anyone with the token has the same access until then. Revoke the session when the agent no longer needs it, or if you think the token leaked.

​

Agent environment

The agent can run anywhere: a Kubernetes pod, a Docker container, a managed sandbox such as Modal or Daytona, or a VM, such as an Amazon EC2 instance. Wherever it runs, we recommend that:

• The proxy runs separately from the agent, such as in its own container, so the agent can’t read the proxy token
• The agent’s network only reaches the proxy, so the agent can’t skip the proxy and connect to an API directly

If you run the agent on your own computer, we recommend isolating it, such as in a container or a VM. Otherwise, the agent runs as your user account, with access to your files and your Infisical CLI login.

​

Proxy placement

You can run one proxy for several agents, or give each agent its own proxy, depending on where your agents run.

​

Common proxy

In this setup, one proxy runs on its own machine (or as a Kubernetes Deployment), and every agent sends its requests to that proxy. Agents send their session token to the proxy over plain HTTP, so use this setup only if your agents and the proxy share a private network, such as your own servers, a Kubernetes cluster, or self-hosted CI runners.

​

One proxy per agent

When the orchestrator creates a sandbox or container for a task, it also starts a proxy in a separate container, on a private network that only the agent and its proxy share. When the task ends, the orchestrator deletes both. Only the agent can reach its proxy, so nothing outside the task can send requests through it.

On a managed sandbox platform like Modal or Daytona, each sandboxed agent needs its own proxy.

If each sandboxed agent shared the same proxy, that proxy would need to be exposed to the internet. Since session tokens are sent to the proxy over plain HTTP, this would create a security vulnerability: anyone who gained access to the session token could then use the proxy to make authenticated requests from anywhere.

Each platform has its own way to run the proxy:

• Modal: a sidecar in the agent’s sandbox (see the Modal guide)
• Daytona: a linked sandbox (see the Daytona guide)
• Docker: a second container in each run’s Docker Compose project (see the Docker guide)

You create the proxy in Infisical once and start every copy with the same proxy token.

​

Access model

Agent Vault uses four primitives to create secure sessions for AI agents:

• Services: The APIs, MCP servers, and anything else the agent talks to, with any credentials to add
• Access bundles: Groups of services, granted as a unit
• Sessions: Time-bound access grants that an agent uses to authenticate to the proxy for the services in an access bundle
• Proxies: The forward proxy that intercepts an agent’s outbound traffic and attaches real credentials to authorized requests

Here’s how the four primitives connect:

1. The access bundle groups the services (external APIs or MCP servers) you want your agent to reach.

2. The orchestrator creates a session for the bundle. Infisical returns a session token scoped to that bundle until the session expires or is revoked.

3. The orchestrator starts the agent with the session token, and the agent uses the token to authenticate to the Agent Vault proxy on every outbound request.

4. Every outbound request goes through the proxy, with the session token attached as the proxy credential.

5. The proxy reads the matching service’s credential from the access bundle, attaches the credential, and forwards the request to the upstream host.

​

Concepts

​

Services

A service represents one external system that an agent can reach, such as GitHub’s REST API or Linear’s MCP server. Each service holds:

• Target: Where the system is, such as the host api.github.com or the server URL https://mcp.linear.app/mcp
• Credential: The real token, username and password, or OAuth authorization that the proxy attaches to the agent’s requests, if the service uses one
• Rules (HTTP APIs): The HTTP methods and paths the service allows
• Custom headers: Extra headers the proxy adds to each request
• Substitutions (HTTP APIs): Placeholders the proxy replaces with real values

Agent Vault supports two types of service: HTTP APIs and MCP servers.

You don’t give services to an agent directly. You add services to an access bundle, and a session gives the agent access to the services in its access bundle.

Services

Anything an AI agent talks to during a session, such as an API or an MCP server.

​

Access bundles

An access bundle groups a set of services under one name. A code-review bundle might hold services for GitHub, Slack, and Anthropic. An on-call bundle might hold services for PagerDuty, Datadog, and Slack.

You grant a bundle to a user, group, or machine identity. Anyone with that grant can create sessions that reference it.

Access bundles

Lists of services that AI agents can access during a session.

​

Sessions

A session is a time-bound access grant that’s scoped to one access bundle. Infisical returns a session token that identifies the session; the agent uses that token to authenticate to the Agent Vault proxy on every outbound request.

When the orchestrator creates a session, it picks one access bundle and sets when the session expires. The agent never sees the real credentials inside the bundle. All it has is the session token, which the proxy only accepts while the session is live.

Even if an attacker stole the session token, they could only reach the hosts, methods, and paths that the access bundle’s services allow, and only until the session expires or you revoke it.

Sessions

Time-bound grants that let AI agents access services without holding real credentials.

​

Proxies

A proxy runs outside the agent’s environment. It intercepts every outbound request, authorizes the request against the session’s access bundle, attaches the matching service’s credential, and forwards the request to the upstream host.

The proxy attaches the credential as a header, or replaces a placeholder that the agent sent in the headers, path, or query string. If the request breaks the matching service’s method or path rules, the proxy refuses it instead.

Proxies

Forward proxies that intercept an AI agent’s requests and attach real credentials on the way out.

​

Example flows

​

GitHub API

Here’s what happens when an agent calls api.github.com through an Agent Vault session:

1. The agent’s HTTPS client routes the request through the proxy (because the HTTPS_PROXY environment variable points at it), including the session token as the proxy credential.
2. The proxy answers the agent with a certificate for api.github.com that the proxy’s certificate authority signed. The agent accepts the certificate because infisical agent-vault run added that certificate authority to the agent’s trusted certificates, so the proxy can read the request.
3. The proxy looks up the session with Infisical, checking if it’s still live and which access bundle it carries.
4. The proxy looks inside the session’s access bundle and finds a service whose host pattern covers api.github.com. If the service limits methods or paths, the proxy refuses a request that doesn’t match them. The proxy reads the real GitHub token from that service.
5. The proxy opens a new HTTPS connection to api.github.com, attaches the real token in the authorization header, and forwards the request.
6. GitHub responds. The proxy pipes the response back to the agent.

If the agent asks for a host that isn’t covered by any service in the access bundle (like api.internal.example.com), the proxy either forwards the request without credentials or refuses it outright, depending on its traffic policy.

​

Linear MCP server

Here’s what happens when an agent calls a tool on Linear’s MCP server through an Agent Vault session. You add the server to the agent by its real URL, https://mcp.linear.app/mcp, the same way you would without Agent Vault, but without a token and without signing in from the agent:

1. The agent’s MCP client sends the tool call to mcp.linear.app through the proxy, with the session token as the proxy credential.
2. The proxy finds the MCP service in the session’s access bundle that covers mcp.linear.app, and reads the credential that Infisical stores for it, such as an OAuth access token or a Bearer token.
3. The proxy records the tool’s name for the session log, attaches the credential, and forwards the request.
4. Linear responds, and the proxy pipes the response back to the agent.

If the service uses OAuth, Infisical refreshes the access token before it expires. If Linear rejects the token anyway, the proxy gets a fresh one from Infisical and retries the request once. Either way, the agent never signs in to Linear itself.

​

Why no code changes are needed

Agent Vault uses the HTTPS_PROXY environment variable, which curl, git, gh, Python’s requests, and most SDKs already read. Any HTTP client that reads it sends its requests through the proxy without code changes.

infisical agent-vault run sets HTTPS_PROXY for the agent, trusts the proxy’s certificate authority, and passes the session token as the proxy credential.

The same applies to MCP servers. Your agent’s MCP configuration keeps each server’s real URL and has no token, and the proxy authenticates the agent’s requests to the server.

Some sandbox providers, such as E2B, can send a sandbox’s HTTP and HTTPS traffic to the proxy over SOCKS5 instead, so the agent doesn’t need HTTPS_PROXY.

Node.js’s built-in fetch (undici-backed) doesn’t read HTTPS_PROXY by default. If your agent uses it directly, run Node with --use-env-proxy or set NODE_USE_ENV_PROXY=1 (Node 24 or later). Third-party clients like axios, node-fetch, and got read the environment variable without extra flags.