Native to the agent
you already use
Connect Tough Tongue AI to the agent you already use. Create scenarios, pull reports, and fix agents in plain English.
Works inside
Two minutes to connect
One install and one browser sign-in, and your agent can operate the full Tough Tongue AI platform.
Install once
Add the ToughTongue AI app in ChatGPT, or install the plugin in Claude or your coding agent. The plugin brings the MCP server and the skills.
Sign in
Your agent opens a browser window. Sign in to Tough Tongue AI and approve access. No API key to create or paste.
Ask in plain English
"Create a practice scenario for pricing objections." Your agent does the rest.
The official ToughTongue AI app. Click Connect and approve access. Be signed in at app.toughtongueai.com first so the consent screen appears.
Full setup for Copilot, Windsurf, Gemini CLI, and other clients
Connecting from a server or CI job? Use a Personal Access Token instead.
Skills, not just access
Tools give your agent access. Skills give it judgment, learned from thousands of scenarios built on our platform. All three are open source.
Scenario Creator
From a brief, a URL, or a call transcript to a live practice scenario. Knows the difference between a cold call, a sales roleplay, and a coaching session, and applies the right rules to each.
"Build a roleplay from this lost-deal transcript. Make the buyer as tough as the real one." Scenario Refiner
A debugger for agent behavior. Pulls low-scoring sessions, finds the exact turn where things went wrong, and makes the smallest fix.
"The onboarding agent hangs up too early. Find out why and fix it." Session Analyst
Team performance answers with receipts. Aggregates scores across sessions and names weaknesses as coachable behaviors.
"Last 50 sessions on the discovery scenario: top 5 improvement areas." 35 tools across the platform
Everything your admin can do in the dashboard, your agent can do on request.
One node in your agent's toolchain
Tough Tongue composes with everything else your agent can reach. Pull transcripts from your call recorder, context from your docs, meetings from your calendar, then create the scenario, run the practice, and send the report. No integrations to build. The agent is the integration.
Open source, top to bottom
The skills, the plugin manifests, and the evaluation suite are MIT-licensed on GitHub. Read exactly what your agent has been taught. The MCP server wraps the same public API our dashboard uses, with an AI-readable reference at llms-full.txt.
Frequently asked questions
No. In ChatGPT or Claude, setup is a one-click install and a sign-in. In a coding agent, it is one command. After that, everything is plain English to your AI agent: describe what you want and it handles the rest.
ChatGPT, Claude (web and desktop), Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Windsurf, and Gemini CLI. Each one signs in with OAuth: you approve access in your browser, and there is no API key to create. In ChatGPT, install the official ToughTongue AI app from the ChatGPT app directory. In Claude, go to Settings > Plugins > Add marketplace and paste https://github.com/tough-tongue/toughtongue-skills. The plugin bundles the MCP server and the skills. Clients without remote MCP support can connect through mcp-remote, which also signs in with OAuth. For servers, CI, and other headless setups, send a Personal Access Token as a bearer header instead.
You sign in with OAuth and approve access on a Tough Tongue AI consent page. That grants a token on your account, which you can view or revoke at any time in Developer settings (app.toughtongueai.com/developer). If you use a Personal Access Token for a headless setup, keep it in the TTAI_PAT environment variable, never in shared config. Tool calls are rate-limited at 30 per minute per token.
The skills and plugin are free and open source, and the hosted MCP server is free to connect. They operate your existing Tough Tongue AI account on your current plan. New accounts start with free practice minutes.
Updates are surgical by design: the refiner skill fetches evidence from real transcripts first, changes the minimum, and verifies. You can also ask your agent to propose changes for review before applying them.
Some agents trim tool discovery to fit their context budget, or keep an old cached list. Start a fresh thread. If it persists, remove and re-add the MCP server. The main server endpoint always serves all 35 tools. The directory endpoint (/api/public/mcp/directory) is a separate 33-tool version that leaves out the two tools that place outbound phone calls.