Core Concepts

Security Model

An eve agent runs in two contexts: the trusted app runtime and the isolated sandbox. Every secret stays in the app runtime. Use this model to decide what an agent, and the model driving it, can reach.

Trust boundaries

App runtimeSandbox
process.env / secretsYesNo
Your Node.js codeYesNo
NetworkUnrestrictedControlled by policy
FilesystemApp's ownIsolated /workspace

The app runtime is the trusted side. Your tool implementations, model calls, connections, and session state run here, with process.env and full Node.js available. On Vercel, this is a Vercel Function.

The sandbox is the isolated side. It has its own /workspace filesystem but no process.env, no secrets, and no path back into the app runtime. On Vercel, each sandbox is a Vercel Sandbox microVM with hardware-level isolation.

Only shell commands execute in the sandbox. The built-in bash, read_file, write_file, and edit_file tools, and opt-in tools such as glob and grep, run in the app runtime and proxy into the sandbox.

The model sees only tool definitions and results. When the model calls a custom charge_card tool, its execute runs in the app runtime, reads process.env.STRIPE_KEY, calls Stripe, and returns { ok: true }. The model sees only { ok: true }, and the call never touches the sandbox.

The just-bash provider has a weaker boundary. Its shell is a JavaScript interpreter with a virtual filesystem inside the app runtime process, with no network isolation. Any experimental.customCommands you register run as host application code outside the virtual filesystem. Use a Vercel or Docker sandbox when the model runs commands you would not run in the app runtime.

See Agent loop and sandbox for how eve connects these contexts while keeping their state and lifetimes separate.

Data flow at a glance

eve sends data only where your agent configuration sends it:

  • Inbound channel data flows through the channel provider you configure, then into the eve app runtime.
  • Model inputs and outputs flow to the model selected in agent.ts, such as a Vercel AI Gateway model id or a provider model instance (EveLanguageModel).
  • Tool and connection calls flow to the external services, MCP servers, OpenAPI endpoints, and channels you configure.
  • Sandbox commands can reach network destinations allowed by the sandbox network policy.
  • Telemetry flows to destinations configured under agent/instrumentation/. By default, eve also records local traces during eve dev and exports to Vercel Agent Runs in preview and production.
  • Eval data flows to the reporters configured in eval settings.

eve stores the session and workflow state it needs to resume conversations, stream events, replay completed steps, and show run observability.

You decide whether your channels, model providers, connected services, and sandbox egress destinations suit your data and use case. The same applies to telemetry exporters, retention settings, and deletion controls.

Credential brokering

Credential brokering gives the model authenticated network access from inside the sandbox, such as a git clone of a private repo. Use it when no tool or connection can route the call. On the Vercel Sandbox backend, the sandbox's network firewall injects auth headers for matching domains. The secret stays in the app runtime, and the sandbox process sees only the response. See Vercel Sandbox Credential Brokering for the platform mechanism, and Sandbox for the eve policy API.

Connection credentials

Connection tokens (MCP and OpenAPI) come from either getToken() or an interactive OAuth flow. eve injects the resolved token into every outbound request and never saves it in session state.

Channel verification

A channel authenticates its own inbound traffic. The built-in platform channels follow two rules, and so must any channel you write:

  • Verify signatures in constant time. Platform channels (Slack, GitHub, Telegram, Twilio) verify the platform's HMAC signature over the raw request body. They compare in constant time, so response timing can't reveal a forged signature. Never use === on a signature or secret.
  • Don't trust body-supplied identity. Derive the caller from a verified signature or token. Never use a principalId (or similar) from the request body, because a body field is attacker-controlled.

A turn runs as one principal

A turn's principal (its authenticator, issuer, principal type and id) never changes while the turn runs; only its claims may update. Only that principal's messages steer the turn or answer its pending questions by text. In a shared thread, another person's message waits for the turn to end and then runs as their own turn, so their text never drives the turn's credentials and the turn never continues under their identity. Another person can still settle a pending request with a button or structured answer; they are recorded as its responder, and the turn keeps its caller. See What reaches ctx.session.auth.

Authored markdown is data

Skill and schedule frontmatter is data: it must parse to a plain YAML object. A fence that names any language other than YAML, such as ---js, throws instead of evaluating code.

Auth fails closed

Routes reject unauthenticated traffic by default. If no AuthFn in the walk accepts the request, it gets a 401; admitting anonymous callers takes an explicit none(). The scaffold's placeholderAuth() keeps a half-configured app closed in production until you replace it. See Auth & route protection for the full walk and verifiers.

Pre-production checklist

Before exposing an agent to real traffic:

  • Replace placeholderAuth() in agent/channels/eve.ts with a real AuthFn (vercelOidc(), httpBasic(), oidc(), or your own). Verify an unauthenticated production request gets 401.
  • Verify channel signatures. Each platform channel needs its signing secret set; custom channels must verify signatures in constant time and never trust body-supplied identity.
  • Keep secrets in process.env, never in compiled artifacts, never passed into the sandbox. Route privileged calls through tools or connections.
  • Limit connection tokens to the least privilege the agent needs; they reach hosts but never the model.
  • Set a sandbox network policy tighter than allow-all if the model shouldn't have open egress; use credential brokering for authenticated egress.
  • Don't surface untrusted text as markup. Model- or user-controlled strings rendered into a channel UI should be escaped for that surface.