Agent Skills: polli — Pollinations CLI

polli — Pollinations CLI

Thin wrapper around gen.pollinations.ai. Generates images, text, audio, video; transcribes speech; manages API keys and usage.

When to use this skill

• User asks to generate an image / text / audio / video via pollinations
• User mentions polli, pollinations, pollen, pollinations.ai
• User wants to transcribe speech or run TTS
• User asks about their pollen balance, usage, or API keys
• User wants to browse or filter available models

Quick reference

| Intent | Command | |---|---| | Log in once | polli auth login | | Generate image | polli gen image "<prompt>" --output out.png | | Generate text | polli gen text "<prompt>" | | Text with stdin as context | echo "<ctx>" \| polli gen text "<question>" | | Describe an image (vision) | polli gen text "what is this?" --image <url> | | One-shot TTS | polli gen audio "<text>" --output speech.mp3 | | Speak out loud | polli gen audio "<text>" --play (uses afplay on macOS; ffplay/mpv/mpg123 on Linux) | | Generate video | polli gen video "<prompt>" --output out.mp4 | | Transcribe audio | polli gen transcribe path/to.mp3 | | Upload a local file | polli upload path/to.png (prints public URL) | | List all models | polli models | | Filter models by type | polli models --type image | | Model health + latency | polli models --stats (default 60m, --window <min>) | | Check balance | polli usage | | Machine-readable output | append --json to any command |

Setup

One-time: polli auth login (device-flow). Verify with polli auth status. Override the stored key for a single command with --key <key>.

Recipes

Generate an image to a file

polli gen image "a fox reading a book, studio ghibli style" --output fox.png

Defaults: zimage, 1024x1024. Pick a different model with --model flux (see polli models --type image). For edits / img2img, pass one or more --image <url> flags — must be public http(s) URLs, local paths are rejected client-side. Only models that list "image" in input_modalities actually consume the flag — flux and zimage are text-only and will silently ignore --image. Find i2i-capable models with polli models --type image --json | jq -r '.[] | select(.input_modalities | contains(["image"])) | .name' (common choices: nanobanana, kontext, p-image-edit). To use a local file, upload it first with polli upload (see next recipe).

Upload a local file to get a public URL

URL=$(polli upload cat.png)
polli gen image "make the cat purple" --image "$URL" --output purple.png

polli upload <file> posts to media.pollinations.ai (10MB max, 14-day TTL, content-addressed so duplicates dedupe). Human mode: URL on stdout, id/size/contentType/duplicate on stderr. --json: full upload response on stdout. The returned URL is public (no auth to fetch) and works anywhere --image is accepted — gen image, gen video, etc.

Generate text

polli gen text "summarize the three laws of robotics"

Save to file: --output summary.txt. Use --system "<msg>" to set system prompt. For reasoning models, pass --reasoning low|medium|high to control reasoning effort. Only send --reasoning to models where reasoning: true in polli models --type text --json — the flag is not validated client-side, and non-reasoning models may silently accept it (openai), ignore it, or return a 400 (mistral).

Describe an image with a vision model

URL=$(polli upload selfie.jpg)
polli gen text "turn this person into a cartoon pet in one playful sentence" --image "$URL"

gen text --image <url...> attaches one or more public https URLs as an OpenAI-style multimodal message — repeatable for multi-image prompts. Local paths aren't supported; run them through polli upload first (see the upload recipe above). Only text models with "image" in input_modalities actually read the image — filter with polli models --type text --json | jq -r '.[] | select(.input_modalities | index("image")) | .name'. Non-vision models silently ignore the attachment. Good defaults: openai, gemini, claude.

Pipe stdin as context into text generation

cat README.md | polli gen text "what does this project do?"

stdin becomes context; the positional argument is the question.

Interactive chat session

polli gen chat --model openai --system "you are a terse assistant"

Slash commands inside the session: /exit, /clear, /save <path>.

Text-to-speech

polli gen audio "hello world" --voice nova --output hello.mp3
echo "long script" | polli gen audio --voice nova --output out.mp3

Default voice is sage. To discover the full live voice list, use the model registry: polli models --type audio --json | jq -r '.[].voices[]?' — each audio model entry includes its voices[] array. Format defaults to mp3; --format opus|aac|flac|wav to change. Accepts stdin (same as gen text). Add --play to save and then play the audio back (handy for narration/demos). Playback starts after the file is fully written, and the command blocks until playback finishes — if you want fire-and-forget, wrap in a subshell: ( polli gen audio "..." --play & ). Player on macOS: afplay; on Linux it tries ffplay, then mpv, then mpg123 in that order.

Generate music (elevenmusic)

polli gen audio "lofi hip-hop beat" --model elevenmusic --duration 30 --instrumental --output track.mp3

Generate video

polli gen video "a spacecraft landing on mars" --model wan-fast --duration 5 --output mars.mp4

Cheapest path: --model wan-fast at ~$0.01/sec, fixed 5-second output (any --duration value is ignored — you always pay for and receive 5 sec). For image-to-video, pass --image <url> with a public HTTPS URL (local file paths and 404/rate-limited hosts will fail with a server error).

Flag support varies per model and is not enforced client-side. --duration, --aspect-ratio, --audio, --negative, and --enhance are forwarded to the server but may be silently ignored — verified on wan-fast where duration is locked to 5s, --aspect-ratio 9:16 still returns 16:9, and --audio produces no audio track. Always inspect the output (file, ffprobe) before trusting a flag worked. Check polli models --type video --json for per-model capabilities.

Video is not tracked by --stats. polli models --type video --stats returns empty — the stats pipe only records text/image/audio events. To compare video models, fall back to polli models --type video --json and look at price/description fields.

Transcribe audio to text

polli gen transcribe recording.mp3 --language en

Models: whisper (default), scribe. Accepts common audio formats (mp3, wav, m4a, flac, ogg); non-audio input (e.g. a .txt file) returns a clear 400 invalid_request_error: extension "txt" not supported — no need to pre-validate with file. Default output is the plain transcript on stdout as a single line (pipe-friendly). Use --json for structured output: whisper returns word-level timestamps, segments, and duration; scribe returns only {text: "..."} — use whisper if you need timing data. --language <ISO-639-1> (e.g. en, fr) is an optional hint that can improve accuracy for non-English or accented speech — whisper honors it and echoes the value in the JSON response; scribe silently ignores it (no error, no effect).

Discover models

polli models --type text              # text models only
polli models --type image --verbose   # with context length / pricing
polli models --stats                  # health + avg latency + err% (60m default)
polli models --stats --window 5       # last 5 minutes only

Use --stats before choosing a model. Caveat: the err% column counts 5xx only — a model can show 0.0% while having massive 4xx rates (auth, validation, etc.). For the full picture use --stats --json and read errors_4xx, errors_5xx, latency_p95_ms.

Pricing fields are per-token, not per-request. completionImageTokens: 0.000008 means each output image-token costs that much — a single 1024x1024 image from gptimage lands at ~$0.008, not $0.000008. Flat-priced image models (flux, zimage) expose completionImageTokens as the whole-image price because they emit exactly one "token" per image. When in doubt, make one call and read the true cost from polli usage --history --limit 5 --json.

Check usage and balance

polli usage              # current pollen balance
polli usage --history    # recent individual requests
polli usage --daily      # daily cost summary

History is eventually consistent — a request you just made may not appear for 30–60s. When matching costs to freshly-generated media, use --limit 50 and filter by timestamp, and retry if the expected entry is missing. polli usage --json returns {"pollen": <number>} — the current balance only; use --history --json or --daily --json for cost breakdowns.

Manage API keys

polli keys list                                                    # list all keys on the account
polli keys info                                                    # details about the CURRENTLY AUTHENTICATED key only (takes no id)
polli keys create --name "my-bot" --type secret --budget 1000 --permissions balance usage   # scoped key
polli keys revoke <id>                                             # id comes from `keys list --json`

--permissions <perms...> scopes what the new key can do on the account (e.g. balance usage lets it call polli --key <new> usage). Without --permissions, new scoped keys can generate media but cannot read account state — polli --key <new> usage will 403. "keys" is auto-stripped from the list so a scoped key can never mint further keys. To inspect a specific key other than the current one, use polli keys list --json | jq '.[] | select(.id == "<id>")'. keys info is intentionally scoped to the caller's own key.

Read API docs

polli docs                          # full llm.txt reference
polli docs /v1/chat/completions     # filter to one endpoint
polli docs --open                   # open in browser

Output contract

• Default (human mode): varies by command — most emit key: value pairs or tab-separated tables with a header row. Exceptions: gen text prints the full response to stdout; gen transcribe prints the transcript as plain text; gen chat runs an interactive REPL. Status/progress messages go to stderr, so pipes stay clean.
• --json: every command emits machine-parseable JSON to stdout; all human messages go to stderr. Always prefer --json when piping into jq or parsing — it's the only shape with a stable contract.
• Exit codes: 0 on success, non-zero on auth failure, rate limit, network error, or invalid args. Error messages go to stderr.

Agent operating rules

1. Run polli auth status first if you don't know whether the user is logged in. Fail fast with a clear "run polli auth login" message if not.
2. Prefer --json whenever you'll parse the output. Never grep human-formatted tables.
3. Don't hardcode model IDs. Fetch the live list with polli models --type <type>. Model availability changes.
4. Before picking a model for production use, check polli models --stats. Rule of thumb for "healthy": err% ≤ 5, avg latency in a reasonable range for the modality (standard text <5s, image <10s, video <60s), and requests high enough to be statistically meaningful (ignore rows with <10 requests — noise). Filter by capability first, then optimize by health — e.g. for a reasoning task, narrow to models where reasoning: true (via polli models --type text --json), then cross-reference against --stats output. The healthiest model overall may not support the capability you need. Reasoning models are inherently slower — expect 5–50s, not <5s; when picking among them, prioritize low err% and request count over raw latency, and compare latency only within the reasoning-capable subset.
5. Always pass --output <path> for gen image, gen audio, gen video — otherwise the file lands in the current directory with a default name.
6. For stdin-as-context on gen text, pipe the context and pass the question as the positional argument: cat file | polli gen text "question about the file".
7. For exact flag lists, run polli <cmd> --help or polli gen <cmd> --help. This skill's recipes cover the common path; the CLI's own help is always the source of truth.
8. Use polli docs [endpoint] over guessing API shapes. It prints the canonical llm.txt reference from the live API.

Common pitfalls

• Forgetting --output on binary generators (image/audio/video) — the file goes to a default path, which may not be what the user wants.
• Using polli gen text --json expecting OpenAI chat-completions shape — the CLI's --json wraps its own structure. Use polli docs /v1/chat/completions to see the raw API shape if you need it.
• Running commands without auth — polli auth status tells you the tier and balance in one call.
• gen text streams to a TTY, buffers when piped. The default now auto-detects — a human at the terminal sees tokens tick in, a pipe/redirect gets the full response once. Force either mode with --stream or --no-stream. For scripts and chains like polli gen text … | polli gen audio …, you don't need to do anything; buffering happens automatically.
• Translating a polli workflow into a browser app. gen.pollinations.ai requires a bearer token, so a plain client-side fetch with no auth returns 401. The only anonymous escape hatch is the legacy https://text.pollinations.ai/openai endpoint, which accepts exactly openai and openai-fast — the healthy-model advice from --stats does not carry over. For anything beyond those two text models in the browser, mint a scoped key with polli keys create and proxy via your own backend.