OttoPort

One API for every model. This package is three things over the same gateway: a CLI, an MCP server, and a Claude Code plugin — chat, image, video, speech, and music, on one key and one prepaid balance.

Get a key at https://ottoport.ai/api-keys (they look like op-…), then pick whichever surface fits.

CLI

npm install -g ottoport     # or run any command below through `npx ottoport …`
export OTTOPORT_API_KEY=op-...

ottoport models [--modality chat|image|video|tts|music] [--json]

ottoport chat  "<prompt>" [--model claude-sonnet-5] [--system "..."] [--no-stream]
                          [--temperature 0.7] [--max-tokens 512]
ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k]
                          [--quality high] [--aspect-ratio 16:9] [--n 1] [--seed 7]
                          [--image <url>] [--ref <url>[,<url>…]] [--out file.png]
ottoport image --model qwen-image-layered --image <url> [--layers 4]
ottoport video "<prompt>" [--model kling-3.0] [--duration 5] [--resolution 1080p]
                          [--aspect-ratio 16:9] [--image <url>] [--last-frame <url>]
                          [--ref <url>[,<url>…]] [--out clip.mp4]
                          [--webhook-url https://…] [--no-wait]
ottoport webhook status|retry <job-id>   # delivery attempts / redeliver now
ottoport speech "<text>"  [--model gpt-4o-mini-tts] [--voice alloy] [--out speech.mp3]
ottoport music "<prompt>" [--model suno-v5] [--duration 15] [--out song.mp3]

ottoport install <host>   # wire the MCP server into an agent
ottoport mcp              # run the MCP server on stdio

Chat streams by default. Image, video and audio commands print the result URL; --out downloads it. --url and --key override the environment for a single call.

Every command is a thin client over the public REST API, so anything the CLI does is available to your own code — the gateway is OpenAI-compatible, and any OpenAI SDK works by pointing baseURL at https://ottoport.ai/api/v1.

Claude Code plugin

Brings the MCP server, the ottoport skill, and six slash commands in one install:

/plugin marketplace add https://ottoport.ai/plugin/marketplace.json
/plugin install ottoport@ottoport

| Command | What it does | | --- | --- | | /ottoport:models | The catalog with live pricing, filterable by modality. | | /ottoport:chat | Send a prompt to any chat model (GPT, Claude, Gemini, Kimi, GLM …). | | /ottoport:image | Generate or edit an image. | | /ottoport:video | Submit a video job and wait for the URL. | | /ottoport:speech | Read text aloud, or generate music. | | /ottoport:setup | Diagnose key / gateway / MCP wiring. |

Other agents

Claude Code is the only host with a plugin format. Everywhere else, one command registers the MCP server by writing that agent's own config:

npx ottoport install codex     # or hermes, cursor, windsurf, claude-desktop, claude

Codex also gets the OttoPort skill copied into ~/.codex/skills/. Your key is forwarded from the environment rather than written into any config file.

MCP tools

ottoport_list_models, ottoport_chat, ottoport_generate_image, ottoport_generate_video, ottoport_video_webhook, ottoport_generate_speech, ottoport_generate_music. The server speaks stdio JSON-RPC and is launched by the host as a subprocess, so it needs Node 20+ on the PATH of whatever starts that host.

Modes and resolution

Video takes four modes — text-to-video, image-to-video (image_url), first-and-last-frame (image_url + last_frame_url), and reference-to-video (reference_image_urls). Images take one reference (image_url) or several (reference_image_urls). Both take a resolution: 480p…4k for video, 1k/2k/4k for images, and a tier above a model's base rate costs more.

The GPT Image family reads a quality on top of that — low, medium (the default every catalog rate is measured at), high, and on gpt-image-2.5 and gpt-image-2.5-flare also xhigh and max. It is how much compute the model spends, and the ladder is steep: max bills sixteen times medium.

Two image models take a picture apart instead: qwen-image-layered and seedream-5.0-layers split image_url into RGBA layers and return one URL per layer, bottom first (the prompt is an optional caption; num_layers, 2–10, is honoured by Qwen only). Each layer returned bills separately.

Which modes and tiers a model accepts differs per model, so ottoport_list_models — and ottoport models in the CLI — prints each one's menu beside its price. For one model in full, including every parameter it reads, pass model to that tool or run ottoport models <model-id>. Asking for something a model does not offer is refused with its actual list, before a generation is spent.

Webhooks

Pass webhook_url (MCP) or --webhook-url (CLI) with a video to have OttoPort POST the finished job to that HTTPS endpoint as video.completed or video.failed. Deliveries are signed: x-ottoport-signature: t=<unix>,v1=<hex HMAC-SHA256(secret, "<t>.<raw body>")>. Failed deliveries are retried after 1, 5, 30 and 120 minutes, 5 attempts in all (redirects are not followed); ottoport webhook retry <job-id> or ottoport_video_webhook sends one again at any time. Verification examples are at https://ottoport.ai/docs#webhooks.

Configuration

• OTTOPORT_API_KEY — your op-… key. Export it from your shell profile: the MCP server reads the environment its host was launched in, so a key exported inside a running session arrives too late.
• The CLI always targets https://ottoport.ai. Use the explicit --url flag only for local development or a self-hosted gateway.
• OTTOPORT_BASE_URL configures the MCP server only; it does not silently redirect CLI traffic.
• OTTOPORT_WEBHOOK_SECRET — the secret video webhooks are signed with, used when a webhook URL is given without an explicit secret. Keep it in the environment rather than on the command line.

Notes

• Image, video, and audio calls return URLs, and provider URLs expire. Download anything worth keeping.
• Video generation takes a minute or more and the call blocks until the job is terminal, unless a webhook is given or --no-wait is passed. A retry is a second billable generation, not a resumption.
• resolution and quality are price multipliers, not formatting hints. Leave them unset to bill at the model's base tier.
• Requests are billed against the prepaid balance on your OttoPort account.

Docs: https://ottoport.ai/docs · Support: [email protected]

MIT © LITBOX LLC