You're viewing an older version (read-only). Switch to Main to edit.
Start building an AI agent instantly.
Your idea
Research
Build
Image agentVideo agentAudio agentText agent
0:00
What will your agent do?
Describe it in one sentence. Ready to test, publish and earn.
Final Preview
Choose Template For Your Agent
No showcase media yet - test output will be used as the listing image.
Output
General & Style
Identity, listing copy, discovery keywords. The agent's marketplace card reads from here.
Template previews aren't available on mobile - continue on a desktop to see how your selected template looks.
Media
Hero gallery + card hover clip. Up to 5 of each media type; the first item is the desktop hero.
Plays on hover over the agent card.
Hero gallery on the agent page - images or 3-30s videos
Model Pipeline
Pipeline of model calls. The last step is always silent; the user sees only its final output.
No steps yet. Add at least one to continue.
Fields
The form the user fills in. Most fields land here automatically from your step schemas; you tweak the title, default, options.
No fields yet. Add a step to auto-populate, or add one manually.
Credits & Preview
The platform computes the provider cost floor and derives an allowed price range from it. Publish requires a passing preview.
Inspect JSON
Total per run-credits
Versions
Every published/saved version of this agent. Approved versions are live and cannot be deleted.
No versions yet. Publish this agent from the Setup tab to create its first version.
Version
Status
Updated
Approved
Actions
How the Agent Builder works
An agent is a small product on the marketplace: a few inputs from the user, a pipeline of AI steps, a result. You assemble one in five passes.
1 · Metadata
Name, description, category, template, tags. Version is managed for you (patch++ on every publish) and estimated duration is learned from real runs — neither is yours to set. The id can be edited up until moderation approves the agent; after that it's locked. Pick one that's not already taken — the wizard warns you inline if it is.
2 · Media
Hero gallery (showcase images / videos) + optional hover-clip on the marketplace card. All optional — skip this pass and the agent still publishes.
3 · Steps
The AI pipeline. + Add step first asks what kind of step:
Model step — pick a generation model (image / video / audio / text). The builder reads its input schema and pre-wires every required input as a form field on Pass 4. Optional inputs stay dormant under + Enable optional until you flip them on.
Content research — runs Claude instead of a generation model, optionally with web search / web fetch. Use it to look something up or write text/data before generating media. Choose the Claude model, toggle the tools you want (each adds a credit), and set the output to Text (markdown) or JSON (structured fields you define). No model schema to wire — just the soul.
Reference mode — some video models expose this toggle on a model step (it comes straight from the model's schema; you'll only see it when that model supports it). On: the uploaded image guides identity/style but the model invents its own scene from the prompt — keep a face, character, or place faithful, then put them somewhere new. Off (default): the model reproduces the same image directly, as the source/first frame. Pick by intent — "keep this person, new scene" turns it on; "animate this exact picture" leaves it off.
Default cap is five steps; the last step is the final output the user downloads.
Each step has a soul — a free-form textbox where you describe what the step does and wire form values in with ${id} placeholders. For a model step the runtime feeds the resolved soul into the model's prompt slot; for a research step it's the instruction to Claude. Type an unresolved ${id} and the builder warns you under the textarea. A later step can pull one field out of a JSON research output with a dotted placeholder — ${research.price}.
▶ Run on any step runs just that step against live models. If a required input is still empty, a small modal collects it (optional inputs you can skip). Cached outputs survive a refresh; in-flight Runs reattach on boot. Run every step clean and that counts as a passed preview — no full-pipeline run required. ▶ Run all on the toolbar wipes the cache and runs the whole pipeline fresh; its chip shows the summed credits.
Running cells show live timing (e.g. "⏳ Running… 5 s / ~30 s") — the estimate starts coarse and tightens after each successful full run. Debug Plays on earlier steps stay internal; only last-step Plays and full-pipeline runs land in the agent's public creation history.
4 · Fields
The form your end user fills in. Most rows land here automatically from Pass 3 (green glow + auto-added badge until you open them). + Add field for hand-added rows — pick catalog when the value goes into a model directly (an image, an aspect ratio, a duration), or custom when it only flavours a prompt (a style word, a free-form note). Click Edit to tweak title, default, options, visibility, cost multiplier; drag ⋮⋮ to reorder.
Fifteen control types, including three visual ones: Image picker (inline cards), Visual picker (modal grid with category chips — good for large libraries; both let you jot a per-step prompt note per option, while your soul references the chosen value via ${id}), and Interactive prompt (a sentence with clickable pills, e.g. "A [mood] cat in a [setting]", where each pill is its own slot with options). Leaving a field's title blank hides the label in the runner.
Required toggle marks a field the user must fill in before Generate. It's locked on when a step wires the field as a required model input (the AI can't run without it); otherwise it's an author-required nudge you can flip on for any field that's important even if the model is fine without it.
The Step toggle on a field makes it a gating step in the user's form — later fields stay locked until this one is answered.
5 · Credits & Preview
The builder computes the price floor from the models you picked. Set Credits per run at or above the floor; the rest is your markup. The same floor drives the Run all chip, the Pass 5 display, the publish gate, and the live agent card — one number everywhere. Run preview kicks the pipeline off as a background job — progress overlay, refresh-safe, 15-minute ceiling, Cancel button after 2 minutes (cancel does not refund — the provider call is in flight). A clean output unlocks Publish.
On Publish the server re-checks the floor (a tampered client can't ship a below-floor price) and persists your Credits per run on the agent doc. Once moderation flips the row to approved, the agent goes live on the marketplace without a redeploy. Previews and per-step Runs are billed — both reserve credits up front and refund on failure (cancel excepted).
My Agents & the "pick first, edit second" rule
Opening /agent-builder shows your agents list — every agent you've built as a card. In-progress agents live in your browser's local storage; published ones are loaded from the database (and re-seed local storage if you've cleared it). Agents land in the database only on Publish. Approved agents show a green ✓ Approved pill on their card; you can keep editing them, but you can't delete them from here.
Both Steps and Fields refuse to show an editor until you've nailed down what you're working with: adding a step opens the model picker first; adding a field asks catalog or custom? first.
First pick what kind of field this is - the rest of the form changes accordingly. Catalog fields reach the model; custom fields are prompt-only.
Catalog field
From the shared input taxonomy. The mapper translates it to whichever provider field the chosen model uses.
Custom field
A free-form knob you reference from a step's prompt via ${yourFieldId}. Never sent to a model directly.
New custom field
Fill in the inputs for this step
These are the form fields this step (and any upstream cached step it needs to run) reads from. Required fields must be filled in; optional ones you can leave blank. Then the step runs against the live models.