Agents - IMP Docs

The agent type is the central building block of IMP. Each agent carries its own model binding, system instructions, tool and skill lists, and access rules. See Agents overview for the product-level documentation.

type agent {
  name: String
  image: String
  defaultagent: Boolean
  category: String
  feedback: Boolean
  suggestions_enabled: Boolean
  description: String
  welcomemessage: String
  instructions: String
  memory: String
  model: String
  active: Boolean
  tools: JSON
  skills: JSON
  animation_idle: String
  animation_responding: String
  sandbox_enabled: Boolean
  max_tool_steps: Float
  guest_access: Boolean
  guest_auth_mode: String
  guest_cover_image: String
  last_processed_at: Date
  embeddings_updated_at: Date
  rights_mode: String
  created_by: Float!
  createdAt: Date
  updatedAt: Date
  providerName: String
  modelName: String
  streaming: Boolean
  capabilities: AgentCapabilities
  maxContextLength: Int
  authenticationInformation: String
  systemInstructions: String
  workflows: AgentWorkflows
  slug: String
  guest_has_password: Boolean
  budget: JSON
  id: ID!
  RBAC: RBACData
}

Field notes

Field	Notes
`rights_mode`	`"public"` — visible to all users; `"private"` — only accessible to users, roles, or teams listed in `RBAC`.
`created_by`	Numeric ID of the user who created this agent.
`model`	Internal model ID string referencing a `model` record.
`providerName`	Computed: the display name of the model’s provider (e.g. `"Anthropic"`). Read-only; populated from the underlying `model` record.
`modelName`	Computed: the human-readable model name (e.g. `"claude-sonnet-4-5"`). Read-only.
`systemInstructions`	Computed: final resolved system prompt after merging `instructions` and `memory`. Read-only.
`capabilities`	Computed: the content types this agent accepts, derived from the underlying model’s capabilities.
`sandbox_enabled`	When `true`, the agent opts into the file-system sandbox for tool execution. Sandboxed runs operate in an isolated environment; default is `false`.
`guest_access`	When `true`, the agent is reachable via a guest link. The link may still require a password or a login depending on `guest_auth_mode`.
`guest_auth_mode`	Controls how guest visitors authenticate: `"public"` (no gate — open access), `"password"` (shared password required), or `"regular"` (login required — authenticated or external users only).
`guest_cover_image`	URL of the cover image shown on the guest landing page for this agent.
`guest_has_password`	Computed: `true` when a guest password has been set on this agent. The password itself (`guest_password`) is input-only — it is hashed server-side and never returned in queries or filters.
`budget`	JSON budget envelope attached to this agent; shape matches the LiteLLM budget object.
`last_processed_at`	Timestamp of the last knowledge-embedding pass over this record.
`embeddings_updated_at`	Timestamp when the vector embeddings for this record were last regenerated.

Example query

query {
  agentsPagination(
    limit: 10
    filters: [{ active: { eq: true } }]
    sort: { field: "name", direction: ASC }
  ) {
    items {
      id
      name
      description
      modelName
      providerName
      capabilities {
        text
        images
        files
      }
      rights_mode
    }
    pageInfo {
      currentPage
      pageCount
      hasNextPage
    }
  }
}