Connector and plugin for AI apps

Use Tokentide from AI apps and agents such as Claude, ChatGPT and Codex, or from any other app that supports MCP (Model Context Protocol). The Tokentide connector is a remote MCP server that your app signs in to with OAuth. With it, your app can explain how Tokentide’s prices work, read current prices, estimate and quote costs, browse the model catalog, show your balance and usage, and run models within the limits you choose. The Tokentide plugin bundles the connector with skills, starting with one that turns a product into short-form videos. One plugin package serves every app that supports plugins.

Install#

Install the plugin to get the connector and the skills, or add the connector on its own. For step-by-step guides, see Tokentide in Claude and Tokentide in ChatGPT and Codex.

Apps that can’t sign in with OAuth can send a Tokentide API key instead, as Authorization: Bearer tt_live_… (in Codex, set bearer_token_env_var for the server). The key’s limits apply just as they do for a connection.

Connector URLs#

The connector serves MCP over Streamable HTTP, without sessions, to apps on the current MCP version and on the 2025 versions. Tools that only read public data work before you sign in. When your app first calls a tool that uses your account, the connector answers with a sign-in challenge and the app asks you to connect. Each tool also declares whether it needs your account (securitySchemes in its metadata), which ChatGPT uses to show its sign-in prompt.

Signing in#

1. Your app opens Tokentide

   Your app registers itself with Tokentide’s authorization server (OAuth 2.1 with PKCE, dynamic client registration and the iss response parameter) and opens Tokentide’s consent page. Apps on your computer, such as Claude Code and Codex CLI, receive the answer on localhost or 127.0.0.1. Sign in to your Tokentide account if you aren’t already.

2. You choose the limits

   Name the connection, then pick a daily spending cap (default $5) and a privacy mode. Tokentide creates an API key just for this connection, with those settings and the default Standard lane.

3. Your app receives sealed tokens

   Your app gets an access token that lasts one hour and a refresh token that lasts 90 days and changes on every use. Both are sealed (encrypted): the app never sees the API key inside them. To disconnect the app at once, revoke the connection’s key under Dashboard > API keys.

The connector can do exactly what the key allows, and nothing more. The key’s daily cap, privacy mode and rate limits apply, and so do your account’s spending limits. Every paid call is billed to your balance and listed in your usage.

Tools#

Every paid tool reports what its call cost: the exact charge, the list cost, the multiplier and the lane. Paid tools are marked as actions that are not read-only and can’t be undone, so Claude, ChatGPT and Codex ask before each call. Most apps also let you choose, tool by tool, whether to ask first.

The connector also offers two resources: tokentide://pricing-guide (the pricing guide as Markdown, from GET /v1/market/rules) and tokentide://openapi.json (the OpenAPI document). It also offers three prompts: pricing-overview, cheapest-time-to-run and spending-report.

What it leaves out#

Video and music generation, text-to-speech, transcription, uploads and the live price stream are not available as tools, for three reasons: chat apps can’t play audio or video results; keeping finished media to offer a download link would break zero data retention (we never store outputs); and Anthropic’s directory doesn’t accept connectors that generate audio or video. For those, use the API, or the video skill in the Tokentide plugin, which runs the generation in your app’s code environment with your own API key. The Responses API is left out because chat_completion covers the same models.

The Tokentide plugin#

The plugin lives in the public repository mds-main/tokentide-plugins, which is also a plugin marketplace. One package serves every app, and each app reads its own files:

The plugin is proprietary and works only with the Tokentide service. More skills will follow.

The video skill#

The viral product video skill researches a product, writes a strategy to make it go viral, then plans, generates, edits, checks and packages vertical videos for TikTok, Instagram Reels and X with Tokentide’s image, video, voice and music models and FFmpeg. It runs Python scripts, so generation works where your app can run code with network access: Claude Code, Codex (in the terminal or the ChatGPT desktop app) and the Claude apps with code execution turned on. It needs:

• FFmpeg and Python 3.8 or later;
• network access to api.tokentide.app. In the Claude apps, allow the domain in Settings > Capabilities. Codex’s sandbox blocks the network by default, so approve the skill’s requests when Codex asks;
• an API key in TOKENTIDE_API_KEY, or in a key file. Create it with privacy mode all, or with private_text to keep text models private. It needs one of these modes because background removal, most premium image-edit and video models, and every music model are anonymized. The skill never asks you to paste the key into the chat.

Without a key or network access, the skill still does the research, the strategy and the video specs, and hands them over.

App directories#

Because the video skill generates media with AI models, the plugin isn’t listed in Anthropic’s directory, which doesn’t accept such software; the connector on its own can be listed there. Tokentide isn’t in OpenAI’s plugin directory yet either. Until it is, install the plugin from Tokentide’s marketplace, as shown above.

Staying in step with the API#

The connector runs inside the API and calls it in process, so it always matches the deployed version. A contract test maps every API operation to a tool, a resource or a documented exclusion, and fails the build when an endpoint changes but the connector does not. The plugin’s manifests, for every app, and its scripts are checked the same way.