Manifest

The manifest is the declarative configuration that defines your agent's identity, model, knowledge sources, tools, channels, and constraints. Every setting you configure through the UI is reflected in the manifest, and vice versa.

Overview

You can view and copy your agent's manifest from the Manifest tab. The manifest is a YAML document that represents the complete shape of your agent in a single, portable format.

The manifest is useful for:

• Version control — Track agent configuration changes over time
• Reproducibility — Recreate an agent from its manifest (coming soon)
• Auditing — Review the full agent setup in one place
• Automation — Create or update agents programmatically

Schema Reference

The current schema version is 2.0.0. All top-level sections are optional — omitted sections mean "not configured".

identity

The agent's name, description, system prompt, avatar, and LLM configuration.

knowledge

Training data and information sources the agent can reference when answering questions.

tools

Capabilities the agent can use during conversations.

channels

Deployment endpoints where users interact with the agent. See Channels for detailed configuration.

constraints

Guardrails and limits for the agent.

pipelines

A pipeline is a lightweight workflow — a series of tool calls stitched together where the output of one step acts as input to the next. Pipelines are not yet configurable through the UI.

Example

A minimal manifest for a support agent with a product dataset:

schema_version: 2.0.0identity:  name: SnowBot  description: Snowboarding Supply bot. Product catalog, FAQs and tech specs.  prompt: |    You are a technical support agent. Help users resolve issues efficiently.     When responding:    - Ask ONE clarifying question if the problem is unclear    - Provide numbered steps for solutions    - Use plain language, avoid jargon unless the user is technical    - If you can't solve it, clearly state what information you need     Stay patient and solution-focused. Never blame the user.  model:    provider: anthropic    credentials: anthropic:production    model: claude-opus-4-6    temperature: 0.2 knowledge:  indexing: index-sources  datasets:    - title: Snowboarding Supplies      file: agents/277878251123712/datasets/277878476329984/snowboarding-supplies.csv      template: |        # {{Product_ID}}         **Title**: {{Title}}        **Description**: {{Description}}        **SKU**: {{SKU}}        **Price**: {{Price}}        **Inventory**: {{Inventory}}        **Vendor**: {{Vendor}} channels:  widget:    display_name: SnowBot    initial_messages:      - "Hey 👋 How can we help?"    suggested_messages:      - Burton Boards      - Jackets    style:      theme: auto      primary_color: "#F5055A"      bubble_color: "#7B55F6"      position: bottom-right    intake:      allowed_domains:        - localhost pipelines:  - name: index-sources    description: Default indexing pipeline    steps:      - step: chunk        provider: unstructured        config:          strategy: fast      - step: embed        provider: openai        credentials: openai:default        config:          model: text-embedding-3-small

Notes

• All top-level sections are optional. Omitted sections mean "not configured".
• credentials fields use the format provider:tag (e.g. openai:default, slack:production).
• The knowledge section uses replace-not-merge on PATCH — sending knowledge replaces the entire section.
• Manifest v1 (integer schema_version or absent) is auto-migrated to v2 on load.
• The manifest contains some sections not yet exposed through the UI, including pipelines and stash.
• All tool invocations are saved against the message that triggered them. You can inspect which tools the LLM is choosing in the Playground by enabling Show Tools.
• No credentials are ever exposed in the manifest. Credentials are stored encrypted at the workspace level and made available to agents when services are configured. The manifest only references credentials by tag (e.g. openai:default), which opens possible futures for agent portability between different environments and workspaces — providing the credential tags match.