Eight providers behind one tier knob — images, speech, video, and talking avatars. One ledger. One budget cap. Hard-enforced before the call, sidecar-tracked after, with two independent $0/call local engines so the free tier isn't a single point of failure. Generate without the bill being a surprise.
Code written for Gemini runs unchanged against OpenAI or a local
Kokoro model — swap --provider and the call still
runs. No per-vendor quirks in your prompt code, no special-casing
voice IDs. Empty cells are honest: that slot just isn't shipped.
| provider | modality | small | mid | pro | batch |
|---|---|---|---|---|---|
| cloud · image | |||||
| Google Geminidefault · cheapest | image | flash-lite$0.034 | 3.1-flash$0.067 | 3-pro-image$0.134 | all · 50% |
| OpenAIgpt-image-2 · to 4K | image | low$0.006 | medium$0.053 | high$0.211 | yes · 50% |
| OpenRouterpassthrough | image | passthrough | passthrough | passthrough | — |
| cloud · tts | |||||
| Google Geminiflash · studio | tts | 2.5 flash tts | 3.1 flash tts | 2.5 pro tts | yes · 50% |
| OpenAItts-1 ··· hd | tts | tts-1 | gpt-4o-mini-tts | tts-1-hd | — |
| ElevenLabs+ word timestamps | tts | turbo | multilingual v2 | — | — |
| cloud · video & avatars — own ladder, see § 02 | |||||
| Replicate3 models · per sec | video | draft · low · normal · high · ultra$0.005 — $0.15 / sec | — | ||
| local · $0 · auto-detected | |||||
| Local serveropenai-compatible | i · t | --model <id>$0.0000 | --model <id>$0.0000 | --model <id>$0.0000 | — |
| Voicebox7 engines · cloning | tts | kokoro · piper$0.0000 | orpheus · qwen3$0.0000 | chatterbox · xtts$0.0000 | — |
| pocket-ttskyutai · 26 voices · cloning | tts | pocket-tts$0.0000 | — | — | — |
Motion prompts and lip-sync are hard to predict, so the bottom rung exists to be thrown away. At half a cent a second you can try twenty framings for the price of one finished take, then move up once the movement is right. Video and talking avatars use the same five words — one vocabulary, learned once.
| tier | generate_video | $/sec | generate_avatar | $/sec | ceiling |
|---|---|---|---|---|---|
| cheap rungs · prunaai/p-video · one engine, both tools | |||||
| draftpreview · throwaway | 720pdraft mode | $0.005 | 720pdraft mode | $0.005 | 20s |
| low | 720p | $0.02 | 720p | $0.02 | 20s |
| normaldefault | 1080p | $0.04 | 1080p | $0.04 | 20s |
| specialist rungs · purpose-built, no length cap | |||||
| high | grok 480pricher motion | $0.08 | fabric 480pbetter sync | $0.08 | 15s / audio |
| ultra | grok 720p | $0.14 | fabric 720p | $0.15 | 15s / audio |
Measured on a square input, normal renders 1408×1408 while ultra renders 960×960. The specialist rungs cost more for motion coherence and lip-sync fidelity, not frame size. If you want a bigger picture, normal is both cheaper and larger — the docs say so out loud, because the names imply the opposite.
The cheap rungs generate straight from the prompt — no still to prepare, nothing to pay for twice. The specialist rungs are image-to-video only, so asking for one without a frame gets a refusal that names the tiers that would have worked rather than a stack trace.
Hand p-video 35 seconds of audio and it returns a successful-looking 20-second clip — no error, nothing in the response to catch it by. Verified: 35.44s in, 20.02s out, status: succeeded. So the tool refuses before spending, and prices the uncapped alternatives from your actual audio length. It will not silently double your bill by escalating on your behalf.
Per-call, session, and per-project ledgers, all written through
state/store.ts with a lockfile so parallel calls
don't race. Hard caps fire pre-call, not after
the charge — the budget guard returns a structured
BUDGET_EXCEEDED with a suggested fix. Identical
repeats hit the cache and return at $0.
estimate_cost ranks every implemented (provider,
tier) combo as a dry-run before you spend. CSV / JSON receipts
export by month. Pricing lives in a versioned JSON with a 30-day
staleness warning surfaced by health_check.
Plenty of MCP servers wrap a vendor. genkettle wraps eight and adds the cross-cutting work that, without it, lands on the caller: budgets, ledger, sidecars, failover, chunking, batch, captions, cloning.
Per-call, session, and per-project ledgers. Hard daily / weekly / monthly caps enforced pre-call, not after the charge. Dry-run estimate_cost ranks every (provider, tier). A $0 cache returns identical repeats for free.
Every output writes a hidden .regenerate.json sidecar — prompt, provider, model, tier, params, lineage. regenerate re-runs the brief; iterate adds a tweak and threads parent → child.
Sentence-aware splitter, ffmpeg stitch, single deliverable. Triggers pre-emptively (text > slot max) and reactively when a provider rejects shorter input as too long — INPUT_TOO_LONG retries on the same provider so voice + cache stay stable.
Automatic on rate-limit, 5xx, or timeout. Each retry logs its cost delta. Voice cloning pins to the preferred provider so the reference isn't silently dropped when a non-cloning backend would have been chosen.
Same plugin, same skills, same sidecars, no API key, no network, no bill. Route to Kokoro-FastAPI, Speaches, Orpheus-FastAPI, Chatterbox, Voicebox, or pocket-tts. Two independent $0 engines matters: a single free provider is a single point of failure, and when one stops answering the other still ships. Auto-detected at startup.
50% off, ≤ 24h. batch_submit queues, batch_status polls and fires notifications/message on completion. Image batch on Gemini Flash + OpenAI; TTS batch on Gemini.
SRT / VTT from ElevenLabs word-level timestamps. Share-target presets (og, twitter, favicon, linkedin, instagram-square) plus a local ONNX background remover. And gallery reads every sidecar on disk into one self-contained page — thumbnail grid, prompt, model, params, and what each one cost.
Elicitation when ≥2 prompts queue (batch vs. sync). Sampling for the prompt rewriter. Resources for recent outputs in the asset panel. Structured errors with code, message, suggestedFix — never raw provider blobs.
Point genkettle at any OpenAI-compatible local server and generate images or speech without a key, network round-trip, or dollar spent. Sidecars, cache, regenerate, iterate, post-processing — everything keeps working.
Auto-detected at startup. If
LOCAL_BASE_URL/models responds, the local provider
joins the failover chain automatically. Set
LOCAL_ENABLED=false only to force-exclude a
reachable server. Voicebox and pocket-tts resolve the same way
via /health.
# 1. start a local TTS backend (Kokoro-FastAPI shown) docker run -p 8880:8880 \ ghcr.io/remsky/kokoro-fastapi-cpu:latest # 2. confirm reachable + which models loaded node mcp-server/dist/cli.js --check-local # 3. generate against the local server, $0/call node mcp-server/dist/cli.js --speech \ -p "hello world" \ --provider local --model kokoro # 4. zero-shot voice cloning (Chatterbox / XTTS backend) export LOCAL_BASE_URL=http://localhost:4123/v1 node mcp-server/dist/cli.js --speech \ -p "read this in my voice" \ --provider local --model chatterbox \ --reference-audio ~/voice-samples/me.wav
A side-by-side that doesn't editorialize. Either the row ships, or it doesn't.
| genkettle | raw OpenAI / Gemini | ElevenLabs sub | guinacio's claude-image-gen | |
|---|---|---|---|---|
| multi-provider · one interface | 8 providers | — | — | OpenAI only |
| pre-call budget enforcement | daily / weekly / monthly | — | plan limits | — |
| per-call + per-session ledger | JSON + CSV export | — | dashboard | — |
| batch mode (50% off) | image + TTS | manual | — | — |
| sidecar / regenerate | .regenerate.json | — | — | — |
| local $0/call escape hatch | pocket-tts + Voicebox + 4 more | — | — | — |
| voice cloning | 2 local engines + ElevenLabs IDs | n/a | subscription | — |
| video · text-to-video · talking avatars | $0.005–$0.15 / sec | — | — | — |
| long-text TTS auto-chunk | sentence-aware · reactive | manual | partial | n/a |
| provider failover | logged cost delta | — | — | — |
| license | MIT · semver · changelog | vendor terms | vendor terms | MIT |
Inside an existing Claude Code session. The plugin registers slash commands and MCP tools automatically on enable.
You'll get the slash commands (/gen-image, /gen-speech, /gen-cost, /gen-budget, /gen-batch-status, /gen-presets, /gen-health) and the MCP tool surface in one go.
/plugin marketplace add sherifButt/claude-image-tts-gen /plugin install claude-image-tts-gen@claude-image-tts-gen-marketplace
Grab a free Gemini key at aistudio.google.com/apikey. Append to ~/.zshrc for persistence — new terminals + sessions inherit it.
# required: at least one provider key OR a local server export GEMINI_API_KEY='your-key' export OPENAI_API_KEY='your-key' # optional: per-provider default voices export GEMINI_DEFAULT_VOICE=Charon export OPENAI_DEFAULT_VOICE=onyx export ELEVENLABS_DEFAULT_VOICE='<voice-id>' # optional: video + talking avatars export REPLICATE_API_TOKEN='your-token' # optional: local $0 engines (auto-detected at startup) export LOCAL_BASE_URL=http://localhost:8880/v1 export POCKET_TTS_BASE_URL=http://localhost:8000
MIT-licensed, semver'd, and shipping. Open an issue, send a PR, or just install it and ship a feature with $5/day worth of cloud spend you can actually account for.
