Manifest reference

Manifests are declarative JSON documents (shown here as YAML for readability) that describe a moulds.ai agent: inputs, pipeline steps, output, and pricing. The platform runtime executes the manifest — you never ship server code for Tier-1 agents.

v2 note: Tier-2 agents use manifest_version: 2 and add runtime, package, ui.routes[], and integrations[]. See Tier-1 vs Tier-2 for the Tier-2 workflow.

Top-level fields

Shapes

• batch_analyzer — user uploads a file or URL; the agent ingests, runs AI, optionally runs Python, renders a report. Good for: receipt analyzers, contract extractors, data enrichment.
• knowledge_qa — user submits a text query; the agent answers from a knowledge collection (vector store) and renders the response. Good for: policy Q&A, GAAP lookup, internal wikis.

landing block

landing:
  tagline: Turn receipt PDFs into a clean expense ledger   # 1–140 chars
  description: |
    Upload a folder of receipts. We extract vendor, date, line items,
    tax, and total — then output CSV ready for your bookkeeper.         # 1–2000 chars
  hero_image_blob: mld://blobs/agent-assets/receipts-hero.png          # optional, must start with mld://

inputs[]

Three kinds, discriminated by kind. All share name (snake_case, ^[a-z][a-z0-9_]{0,40}$) and required (boolean, default false).

file_upload

- kind: file_upload
  name: receipts
  accept: [application/pdf, image/png, image/jpeg]   # required, min 1 MIME type
  required: true
  max_size_mb: 50                                      # default 25, cap 100

url

- kind: url
  name: source_page
  required: true

text

- kind: text
  name: query
  required: true
  multiline: true       # renders a textarea, default false
  max_chars: 4000       # default 2000, cap 10000

pipeline[]

Ordered steps; outputs of earlier steps are available as template variables in later steps. Four step types:

ingest — extract content from files

- step: ingest
  config:
    extract: [text, tables]   # min 1 element; each must be "text" or "tables"

Use {{ingest.documents.0.text}} (not {{ingest}}) in downstream templates.

ai_generate — LLM call with structured output

- step: ai_generate
  config:
    model: pro                      # "flash" (fast/cheap) or "pro" (accurate)
    prompt_template: |
      Extract every line item from the receipt below.
      Content: {{ingest.documents.0.text}}
    output_schema:
      type: object
      properties:
        vendor: { type: string }
        total_cents: { type: integer }
        items:
          type: array
          items:
            type: object
            properties:
              description: { type: string }
              amount_cents: { type: integer }
    knowledge_collections: []      # vector-store collection IDs; required for knowledge_qa shape

Step output keys default to the step type (ai_generate, ingest, etc.). Add name: to override if you have two steps of the same kind.

compute_python — sandboxed Python script

- step: compute_python
  config:
    script_blob: mld://blobs/agent-scripts/receipt-totals-v3.py   # must start with mld://
    packages: [pandas, numpy]                                       # PyPI packages
    timeout_sec: 90                                                  # default 120, cap 240

Upload the .py file via the builder console first, then reference the returned mld:// URI.

render_report — final output via Mustache template

- step: render_report
  config:
    template: |
      <h1>{{vendor}}</h1>
      <p>Total: {{total_cents}} cents</p>
      <ul>
      {{#items}}
        <li>{{description}} — {{amount_cents}}</li>
      {{/items}}
      </ul>

Max template size: 100 KB. Triple-stache {{{var}}} is banned — the renderer throws BAD_REQUEST if it sees {{{ anywhere. Use standard {{var}} (HTML-escaped).

review_ui (optional)

Pauses the pipeline after ai_generate and shows an editable form before continuing.

review_ui:
  enabled: true
  fields: [vendor, total_cents]   # output_schema properties to expose; [] = all fields

pricing

pricing:
  free_trial:
    quantity: 10       # number of free runs; 0 disables the trial
    period_days: 30    # trial window length in days
  plans:
    - id: starter               # snake_case, ^[a-z][a-z0-9_]{0,40}$
      price_cents: 1900         # 1900 = $19.00
      interval: month           # "month" or "year"
      feature: report           # what the plan grants
      quantity_per_period: 100
    - id: pro
      price_cents: 14900
      interval: month
      feature: report
      quantity_per_period: 1000

pricing.plans is required but may be [] (free-only agent). platform_fee_pct defaults to 15 (moulds.ai's revenue share from Stripe Connect).

Common gotchas

• Slug format: must match ^[a-z][a-z0-9-]{1,62}$. BadSlug!, My-Agent, 1agent, and single-char a all fail. Use kebab-case: receipt-analyzer, gaap-lookup.
• Identifier format: inputs[].name and plans[].id use ^[a-z][a-z0-9_]{0,40}$ (snake_case). Hyphens are not allowed — that's slug territory.
• {{ingest}} vs {{ingest.documents.0.text}}: {{ingest}} is the whole result object. Use the dotted path to get the raw extracted text.
• Step naming collision: if you include two ai_generate steps, add a name: field to each so their outputs don't overwrite each other.
• Blob URIs: both landing.hero_image_blob and compute_python.config.script_blob must start with mld://. HTTP URLs and bare file paths are rejected.
• Pipeline order matters: a render_report step that references {{vendor}} only works if a prior ai_generate step produces a vendor field.
• Triple-stache banned: {{{var}}} throws BAD_REQUEST at render time. Use {{var}}.
• Limits cheat sheet:
  • file_upload.max_size_mb: default 25, cap 100
  • text.max_chars: default 2000, cap 10000
  • compute_python.timeout_sec: default 120, cap 240
  • name: 1–80 chars
  • landing.tagline: 1–140 chars
  • landing.description: 1–2000 chars
  • category: 1–40 chars