Pollo Generate

pollo-generate

pollo-cli already owns auth, token refresh, and request shaping. Your job starts one step later: pick the right model, learn its input fields, build a valid input, submit, and hand back the result. Everything below follows that order — model, then input, then submit, then result — because that's also the order the server validates in, so skipping a step just turns into a server error one step later.

Setup (once per session)

1. If pollo isn't on $PATH, install the released binary — don't build from source:

   npm i -g @pollo-ai/cli
   

   Without Node (macOS / Linux only), fall back to the standalone installer:

   curl -fsSL https://raw.githubusercontent.com/polloaiofficial/cli/main/install.sh | sh
   

2. Run pollo account status. An availableCredits number means you're already logged in to prod (pollo.ai, profile default). A missing/expired token error means:

   pollo auth login   # opens the browser, logs into pollo.ai -> profile "default"
   

   The CLI owns its own credential storage and refreshes the short-lived token before each request, so you won't need to log in again mid-session. Never read, print, or copy the stored credentials — every command below authenticates on its own.
3. Stay on the default profile unless the user names another one. Profiles let one machine hold several logins; pollo profile list shows them (* marks current), pollo profile use <name> switches, and pollo --profile <name> <command> overrides for a single command.

Step 1 — Find the model, then read what it actually accepts

Never guess a brand/alias or a field name — they change as models ship, so a guess just surfaces as a server validation error one step later instead of saving you a lookup now:

pollo model list                        # every standard model, capabilities aggregated
pollo model list --type text2video      # only models with that capability
pollo model get <brand>/<alias> --fields   # this model's input shape

Each model is listed once, with a capabilities[] array. A capability is an input mode — text2video | image2video | ref2video | text2image | image2image — and each one carries its own path (the output endpoint, /video or /image) plus its own fields/constraints. So a model that takes both text and images has several capabilities sharing one path, and a dual-output model like kling-v3-omni has capabilities across both paths.

--type on model list/get filters by capability (the values above). Don't confuse it with --type on generate create/cost, which is the output endpoint (video|image) — see Step 3.

model get --fields (same output as --json) returns one entry per endpoint, each already flattened to:

• fields[] — {name, type, description, values?, min?, max?}. values is the enum choices; min/max bound a numeric field.
• constraints[] — {id, message, expr}, a boolean rule over the eventual input (e.g. a field that's only valid in a certain mode). If your planned input would fail one, message tells you what to change before you submit instead of finding out from a 400.

The input mode is a capability, but it is not a separate endpoint — at request time it's decided by which fields you put in input: prompt only → text-driven, image → image-driven, refs → reference-driven. So "does this model support references?" = "does it have a ref2video capability whose fields include refs?" — check that, don't guess. If a model exposes both a /video and an /image path, pick the one matching the user's intent and build input only from that capability's fields.

If the user hasn't named a specific model, read references/model-routing.md first — it maps what the user actually said to a concrete brand/alias (video branches for on-screen text / single image / multi-reference / default, quality tiers, and an image routing table), with the fallback chain to use when a generation disappoints. Brands are not guessable (HappyHorse and Wan are alibaba/, Sora is openai/, Grok is xai/, Seedream is bytedance/), so route from that file rather than from memory, and verify with model list before you commit.

Once you know which video model you're using, write the prompt field's wording following that vendor's own formula in references/prompt-engineering.md — Kling, Veo, Seedance, Sora, Vidu, Wan, and Pixverse each respond best to a different prompt structure, and a Seedance-style prompt on Veo (or vice versa) tends to underperform.

Step 2 — Prepare inputs

Image/video fields (commonly image, images, or refs) accept either a local file path or an HTTPS URL — you don't upload anything yourself. When you pass a local path, the CLI uploads it to Pollo storage as part of the generation and substitutes the resulting URL before submitting; pass a hosted HTTPS URL and it's used as-is. (base64 isn't accepted.) There is no separate upload command — uploads only happen as a step inside generate create/cost.

Step 3 — Submit and read the result

# optional: estimate before an expensive run, or whenever the user is cost-sensitive
pollo generate cost <brand>/<alias> [--type video|image] --name value ...
# -> cost / singleCost / discountCost / discountSingleCost (+ their *Usd counterparts)

pollo generate create <brand>/<alias> [--type video|image] \
  --name value ... --json    # -> { "taskId": "...", "status": "waiting" }

pollo generate wait <taskId> [--timeout 10m] [--interval 3s] --json
# -> { "taskId", "credit", "generations": [{ "status": "succeed|failed",
#      "url", "mediaType", "failMsg" }] }

Notes on the flags above:

• You usually don't need --type: the CLI resolves the endpoint from /v1/models on its own. Pass --type video|image only when the model has both output types (the CLI errors and tells you so), or to skip the extra lookup when you already know.
• Every other --name value becomes input.<name>. Values auto-type: true/false → boolean, numeric strings → number, everything else → string. Repeat a flag to build an array (e.g. multiple --images <url>); refs is always sent as an array even from a single --refs <url>.
• --webhookUrl is lifted to the request envelope's top level, not into input.
• There is no --wait flag on create — an unknown flag would be sent as an input field instead of erroring, so always go through create then wait as two separate calls, and use --json on create so you can parse taskId reliably.
• The deliverable lives in the generations array: each item carries the output url, mediaType, and per-item status/failMsg. Parse it from the --json output — without --json, wait/get print a human summary of only the first generation, so multi-output results get truncated. Hand the user the URL(s); everything before that (taskId, polling) is just plumbing you don't need to narrate.

Credits & recharge

Credits are spent per generation. Gauge affordability up front by comparing pollo account status (availableCredits) with pollo generate cost ….

When the user is short on credits — either you can already see availableCredits is below the cost, or a generate create fails with insufficient credits — recharge at <url> — treat it as terminal: do not switch models, retry, or loop. Tell the user how many more credits they need and point them to the recharge page <host>/pricing, where <host> is the environment you're logged into (pollo profile current shows it, e.g. https://pollo.ai/pricing). Always give the actual link, not just "please recharge".

Worked examples

pollo account status
pollo model list --type text2image
pollo model get google/nano-banana --fields

# Text-to-video
pollo generate create bytedance/seedance-2-0-fast \
  --prompt "a red sports car drifting on a mountain road" \
  --aspectRatio 16:9 --duration 5 --json

# Text-to-image (endpoint auto-resolved to /image for an image-only model)
pollo generate create pollo-ai/pollo-image-v2 \
  --prompt "a cozy cabin in a snowy forest, warm light" --aspectRatio 16:9 --json

# Image-to-image (image field = local path, uploaded automatically; or an HTTPS URL)
pollo generate create google/nano-banana \
  --image ./photo.png --prompt "make it nighttime" --aspectRatio 1:1 --json

# Reference-to-video: refs on the video endpoint (repeat --refs for multiple).
# Only refs-capable models accept it — confirm `refs` shows up in model get --fields.
pollo generate create vidu/viduq3 \
  --refs ./character.png --prompt "the character walks into frame" --json

# Dual-output model (both /video and /image): pick with --type
pollo generate create kling-ai/kling-v3-omni --type image --prompt "..." --json

pollo generate cost pollo-ai/pollo-image-v2 --prompt "..." --aspectRatio 16:9
pollo generate wait <taskId>
pollo generate get <taskId>        # one-shot status check, no polling

When it fails

• Insufficient credits (insufficient credits — recharge at …, or you can see availableCredits is below the generate cost) — terminal, not a model problem. Do NOT switch models or retry; give the user the recharge link <host>/pricing and how many more credits are needed. See "Credits & recharge".
• 403 "This model is not enabled for API access." — that model isn't exposed on this account/environment, even though model get and generate cost may still succeed for it. Go back to Step 1, pick a different model, retry.
• Any other validation error from the server — re-run pollo model get <brand>/<alias> --fields and check the offending field/value against the fields/constraints you got back; something in input doesn't match what the model actually declared.
• Expired/invalid token — pollo auth login again. This is only needed once the long-lived refresh token itself has expired; the short-lived one refreshes automatically.
• Generation fails uploading a local image (/file/sign 403) — the upload step is gated to some accounts. Pass an existing HTTPS image URL for the image field instead of a local path.
• generate list returns the server's raw history shape (always JSON) — treat it as opaque history to skim, not a typed result to parse fields out of.

Everything else

pollo --help / pollo <command> --help cover the full flag list: global flags (--profile, --json, …) and profile management (pollo profile list|use|current|remove). The request body create/cost build is { "input": { ...your --name value... }, "webhookUrl"? }, posted to /v1/generation/{brand}/{model}/{video|image} (/estimate for cost) and validated against that endpoint's fields/constraints — the same data model get --fields shows you up front.

The same capabilities are also exposed over MCP (pollo mcp, a local stdio server) for non-terminal agent clients like Cursor or Claude Desktop — same login and /v1 endpoints, tools pollo_list_models (pass format: 'fields' to read a model's input fields) / pollo_estimate_generation_cost / pollo_generate_image / pollo_generate_video / pollo_get_generation_status. Image inputs there also take a local path (uploaded automatically) or an HTTPS URL — there's no separate upload tool. As this skill drives the CLI directly you won't call those here; see the README "Use as an MCP server" for client setup.