Tough Tongue AI Skills

Agent skills and MCP server for Tough Tongue AI, the platform for handling tough conversations. Some, the AI takes: voice agents that answer and place calls, qualify leads, run demos, screen candidates, and book meetings. Others, you nail: hyper-realistic roleplay that gets you ready for negotiations, interviews, and coaching conversations.

Install once, then work with Tough Tongue AI from Claude Code, Codex, Cursor, or any agent that supports the Agent Skills format:

• "Pull the last 3 lost deals from our call notes and create a practice scenario for the pricing objection."
• "For the Enterprise Discovery Call scenario, pull the last 50 sessions. What are the top 5 improvement areas?"
• "The onboarding agent keeps ending calls too early. Pull the low-scoring sessions, find out why, and fix the scenario."

These compose with the rest of your MCP ecosystem: Gong, Notion, calendar, slides, email. See What you can do for the full journeys.

Table of contents

• What's included
• How this repo is structured
• Prerequisites
• Which setup fits you?
• Set up
  • One command, every agent
  • Claude Code
  • Codex
  • Cursor
  • Copilot / Windsurf / Gemini CLI / other agents
  • Skills only
  • MCP only
• Pin a version
• Verify your setup
• What you can do
• MCP Server
• Troubleshooting
• Repository structure
• Related
• License

What's included

This repo ships two layers that work together, plus plugins that bundle both:

Skills — workflow guidance your agent loads automatically when the conversation matches:

Skill	When it activates	What it does
ttai-agent	Any Tough Tongue AI / ttai question	Intelligence layer: scope, capability, entities, scenario quality, runtime, MCP guidance
scenario-maker	"Create a scenario", "create a voice agent", "fix the scenario", …	Create, edit, or refine scenarios via MCP.
session-analyst	"How is my team doing?", "top improvement areas", …	Turn session data into structured reports.
browser-demo-builder	"Record browser demo steps", …	Pre-recorded browser demo steps via MCP.

MCP server — live actions in your Tough Tongue AI account. 27 tools over the public API: scenarios, sessions, analytics, organizations, SIP, meeting bots, and collections. Full catalog in MCP.md.

Plugins (Claude Code, Codex, Cursor) — bundle the skills and the MCP server registration in one install.

How the layers relate:

• ttai-agent is the provider-neutral intelligence layer: it turns intent into scoped, capability-aware decisions using the entity handbook and features/mcp adapter guidance.
• Workflow skills add judgment for a job (create, refine, analyze, record a demo).
• With skills only, your agent can advise. With MCP only, it can call tools. With both, it picks the right workflow and completes it. The plugin gives you both in one step.

Need a specific git tag, commit, or a single skill? See INSTALL.md.

How this repo is structured

flowchart TB
  subgraph install["You install one of these"]
    P["Plugin — skills + MCP registration"]
    Sonly["Skills only — guidance, no live tools"]
    Monly["MCP only — 27 ttai tools, no workflows"]
  end
  subgraph layers["Skill layers — load down, don't copy up"]
    L1["Workflows — scenario-maker · session-analyst · browser-demo-builder"]
    L0["ttai-agent — kb (entities, recipes) + features/mcp"]
    L1 --> L0
  end
  P --> layers
  P --> MCP["Hosted MCP https://api.toughtongueai.com/api/public/mcp"]
  Sonly --> layers
  Monly --> MCP

Workflow skills tell the agent when. They load ttai-agent for what exists and how to act. Situation recipes (cold call, sales roleplay, coaching, demo, cascade TTS) live under skills/ttai-agent/kb/scenario-recipes/ — not copied into every workflow skill.

The intelligence layer has no terminal or UI dependency. Coding agents load it before calling MCP; web conversational plugins can pass verified account, workspace, and focused-resource context to the same layer, then execute the result through their server-side adapter.

Prerequisites

Every path needs a Tough Tongue AI account — sign up at app.toughtongueai.com.

That's it for interactive setups. Authentication is OAuth-first: the first time your agent calls a Tough Tongue AI tool, the client opens a browser consent page — approve once and the client stores the token (Claude Code and Codex use your system keychain). No environment variables, no app restarts.

Upgrading from a PAT-based install? After updating the plugin you'll see a one-time OAuth prompt (Claude Code: run /mcp and authenticate; Codex: codex mcp login ttai or approve the automatic prompt). Your existing PAT keeps working for manual/headless configs.

Browser OAuth can't run in headless environments. Create a Personal Access Token (PAT) at app.toughtongueai.com/developer and export it where your agent runs:

export TTAI_PAT="<your-token>"

Then register the server manually with a bearer header — per-client commands in MCP only and MCP.md.

Note: never commit the PAT or paste it into config files — reference it only through the TTAI_PAT environment variable.

Which setup fits you?

Setup	Best for	What you get
Plugin (recommended)	Claude Code, Codex, Cursor	Skills + MCP, auto-configured in one install
Skills + MCP, manual	Copilot, Windsurf, Gemini CLI, other agents	Same capability, assembled in two steps
Skills only	Any Agent Skills client, no live tools needed	Workflow guidance; the agent advises but can't act
MCP only	Developers who want raw API tools	27 tools; no workflow guidance

Not sure? Use the plugin — it's the least setup. Per-client instructions below. To freeze a git tag, commit, or a single skill, see INSTALL.md.

Set up

One command, every agent

The fastest path. The plugins CLI detects the coding agents on your machine — Claude Code, Cursor, Codex, Grok Build, Kimi Code, GitHub Copilot CLI, VS Code — and installs the skills plus the MCP server into all of them at once:

npx plugins add tough-tongue/toughtongue-skills

Restart your agent, then say "get me started with Tough Tongue AI". The ttai-agent intelligence layer verifies the connection and routes you to the right workflow. On the first tool call your client opens the browser OAuth consent — approve once and you're connected.

This repo is also a standard Agent Plugin (root plugin.json, skills/, mcp.json), so clients that load Agent Plugins natively can point straight at it. Prefer your client's own plugin system? Per-client instructions below.

Claude Code

The plugin bundles the skills and registers the Tough Tongue AI MCP server (.mcp.json) in one install: no separate claude mcp add step needed. On the first tool call, Claude Code opens a browser OAuth consent — approve once and you're connected (re-authenticate anytime with /mcp).

Inside Claude Code:

/plugin marketplace add tough-tongue/toughtongue-skills
/plugin install toughtongue@toughtongue-skills

Or from the terminal:

claude plugin marketplace add tough-tongue/toughtongue-skills
claude plugin install toughtongue@toughtongue-skills

Then say "get me started with Tough Tongue AI" or call ttai:list_organizations. ttai-agent verifies the connection, takes a lightweight inventory, and selects the relevant workflow.

Skills are namespaced after install: invoke them as /toughtongue:scenario-maker, /toughtongue:session-analyst, or /toughtongue:browser-demo-builder; or just describe the task and Claude picks the right skill automatically.

Upgrade — refresh the marketplace catalog, then move the installed pin to the latest version:

claude plugin marketplace update toughtongue-skills
claude plugin update toughtongue@toughtongue-skills

Then run /reload-plugins in your session to apply.

Local development / testing — load the plugin without installing:

claude --plugin-dir /path/to/toughtongue-skills
# after edits, run /reload-plugins inside the session

Codex

The plugin bundles the skills and registers the Tough Tongue AI MCP server (.mcp.json) in one install:

codex plugin marketplace add tough-tongue/toughtongue-skills
codex plugin add toughtongue@toughtongue

Then restart Codex and start a new thread. Codex detects the server's OAuth support and prompts you to log in (or run codex mcp login ttai). Say "get me started with Tough Tongue AI" to verify the setup and start your first workflow.

Upgrade — both steps are needed; the first refreshes the marketplace snapshot, the second re-pins the installed plugin to it:

codex plugin marketplace upgrade toughtongue
codex plugin add toughtongue@toughtongue

Local development / testing — register the checkout as a local marketplace:

codex plugin marketplace add /path/to/toughtongue-skills
codex plugin add toughtongue@toughtongue

Cursor

In Cursor, go to Settings > Plugins > Team Marketplaces > Add Marketplace > Import from Repo, point it at https://github.com/tough-tongue/toughtongue-skills, then install toughtongue.

Or assemble it manually in two steps:

1. Install the skills:

   npx skills add tough-tongue/toughtongue-skills
   

2. Add the MCP server: Install in Cursor — one click adds the server; Cursor prompts for OAuth consent on first use. Or add it manually via "Cursor Settings" > "MCP" (config JSON in MCP.md).

Re-import the marketplace from Settings > Plugins > Team Marketplaces. If you installed the skills via the CLI, refresh them with npx skills update.

Copilot / Windsurf / Gemini CLI / other agents

No plugin for these yet — install the skills and the MCP server as two steps:

1. Install the skills. Works with any Agent Skills-compatible client:

npx skills add tough-tongue/toughtongue-skills

The CLI prompts you to pick which skills to install and which agents to configure. To install everything non-interactively:

npx skills add tough-tongue/toughtongue-skills --all

Update existing skills later with:

npx skills update

2. Add the MCP server. Point your client at the hosted server — the general shape is:

{
  "mcpServers": {
    "ttai": {
      "url": "https://api.toughtongueai.com/api/public/mcp"
    }
  }
}

Clients with OAuth support prompt for consent on first use. For clients (or headless setups) that need a bearer token instead, add "headers": { "Authorization": "Bearer ${TTAI_PAT}" } with the PAT from Prerequisites. Exact config per client (Copilot/VS Code, Windsurf, Gemini CLI) is in MCP.md.

Stdio-only clients (Claude Desktop, Zed, older VS Code): bridge to the hosted server with mcp-remote — see MCP.md.

Skills only

Want the workflow guidance without connecting your account?

npx skills add tough-tongue/toughtongue-skills

The agent can advise on scenario design, evaluation rubrics, and coaching patterns — but it can't create or modify anything in Tough Tongue AI. Add the MCP server later (step 2 above) when you want action.

MCP only

If you only want the tools (no skills) — OAuth flow starts on first use:

# Codex
codex mcp add ttai --url https://api.toughtongueai.com/api/public/mcp
codex mcp login ttai

# Claude Code
claude mcp add --transport http ttai https://api.toughtongueai.com/api/public/mcp
# then run /mcp in a session to authenticate

Headless / CI (PAT bearer auth instead of OAuth):

# Codex
codex mcp add ttai --url https://api.toughtongueai.com/api/public/mcp \
  --bearer-token-env-var TTAI_PAT

# Claude Code
claude mcp add --transport http ttai https://api.toughtongueai.com/api/public/mcp \
  --header "Authorization: Bearer ${TTAI_PAT}"

More clients in MCP.md.

Verify your setup

Ask your agent:

Call the ttai MCP tool list_organizations and show me the result.
Then list my scenarios.

You should see two things:

• The agent reaches your Tough Tongue AI account and returns real data — the MCP connection works.
• The agent reasons in Tough Tongue AI terms — scenarios, sessions, rubrics, report cards — the skills are loaded.

On a skills-only setup, verify with "What can I do with Tough Tongue AI?" — the agent should describe the workflows but won't be able to call tools.

What you can do

Coding agents are where work happens now. These journeys show Tough Tongue AI composing with the other tools already connected to your agent: copy any prompt to start.

1. Practice this sales call

A sales manager spots an AE struggling with pricing objections in real calls.

Pull the last 3 calls from Gong where we lost on pricing. Create a Tough Tongue AI scenario to practice handling that objection, using our positioning doc from Notion. Give me the shareable practice link.

Gong MCP (search_calls / list_calls → get_call_transcript) + Notion MCP → scenario-maker builds a sales roleplay from the real objections → create_scenario → shareable practice link.

2. How is my team doing?

A VP of Sales wants the team's skill gaps, not raw transcripts.

For the "Enterprise Discovery Call" scenario, pull the last 50 sessions. What are the top 5 improvement areas? Build me a 3-slide deck.

session-analyst → list_sessions (scores, strengths, weaknesses per session) → aggregates report-card topics and weakness themes → slides MCP for the deck.

3. Refine from real conversations

A CS manager notices low scores on the onboarding scenario.

Pull the 5 lowest-scoring sessions for our onboarding scenario, figure out what went wrong, and fix the scenario.

list_sessions (sorted by score) → get_sessions_batch (full transcripts + evaluations) → scenario-maker diagnoses the root cause → surgical update_scenario: live for the next session.

4. Prep me for this meeting

A sales rep has a discovery call in 30 minutes.

I have a call with Sarah Chen from Acme Corp in 30 minutes. Create a quick practice scenario so I can rehearse.

Calendar MCP + web search for attendee/company context → scenario-maker → create_scenario → start practicing in minutes.

5. Automated post-call coaching

An engineering team wires coaching into their call pipeline.

Every time a call ends in Gong, analyze the transcript and email the rep a coaching report.

Webhook script → create_session (ingest transcript against a coaching scenario) + post_process_session → poll get_sessions_batch until analysis completes → email MCP sends the report. Full recipe in session-analyst's report templates.

MCP Server

The plugin registers Tough Tongue AI's hosted MCP server at https://api.toughtongueai.com/api/public/mcp (Streamable HTTP, OAuth 2.1 with dynamic client registration; PAT bearer auth for headless setups). There is nothing to install and no local process to run. It exposes 27 tools over the public API: scenarios (create, update, generate, access tokens), sessions (list with evaluations, single and batch fetch, ingest, post-process), analytics and organizations, SIP phone calls, meeting bots, and collections.

See MCP.md for the full tool catalog, per-client setup (Claude Code, Codex, Cursor, Copilot, Windsurf, Gemini CLI), and troubleshooting.

The same OAuth flow powers the web connectors. In claude.ai, add a custom connector (Settings > Connectors) pointing at the server URL with the Client ID and Secret left empty. In ChatGPT (Business/Enterprise/Edu), enable developer mode and create a custom MCP app with OAuth. Full steps for both are in MCP.md.

Troubleshooting

• Fewer than 27 ttai tools listed: your agent trimmed or cached tool discovery. Start a fresh thread; if it persists, remove and re-add the MCP server.
• 401 / authentication errors: the OAuth login hasn't completed for this client. Claude Code: run /mcp and authenticate in the browser. Codex: codex mcp login ttai. Cursor: Cursor Settings > MCP > log in on the ttai server. If you're on a manual PAT config instead, check TTAI_PAT is visible to the agent process (re-export, launchctl setenv on macOS, fully restart the app).
• Skills installed but the agent can't do anything: skills are guidance only — the agent also needs the MCP server for live actions. See Which setup fits you? and add the MCP server for your client.
• Using claude.ai or ChatGPT web?: same OAuth flow, via custom connectors. See the web connector sections in MCP.md.
• Scenario edits not taking effect in a running call: scenario changes apply to new sessions only; sessions compile their prompt at start.

If /plugin install fails or claude plugin list shows stale entries, do a clean reinstall — run these in order in any Claude Code session:

1. /plugin marketplace remove toughtongue-skills
2. /plugin marketplace add tough-tongue/toughtongue-skills
3. /plugin marketplace update toughtongue-skills
4. /plugin install toughtongue@toughtongue-skills

Step 3 forces Claude Code to re-read the marketplace manifest. After step 4, claude plugin list should show one toughtongue@toughtongue-skills entry.

Repository structure

On-disk layout. How to pin a plugin, skill, or git ref: INSTALL.md.

toughtongue-skills/
├── plugin.json            # Agent Plugins 1.0.0 manifest (agent-plugins.org)
├── mcp.json               # Agent Plugins MCP config (streamable-http; no credentials)
├── .plugin/               # Marketplace entry for the `npx plugins` CLI
├── .claude-plugin/        # Claude Code plugin + marketplace manifests
├── .codex-plugin/         # Codex plugin manifest
├── .cursor-plugin/        # Cursor plugin manifest
├── .agents/plugins/       # Codex plugin-marketplace entry
├── .mcp.json              # Claude-native MCP server registration (OAuth; no credentials)
├── MCP.md                 # MCP server docs: setup per client, tool catalog
├── skill-evals/           # Evaluation scenarios per skill
└── skills/                # Each: SKILL.md (+ kb/ / features/ or references/) + agents/openai.yaml
    ├── ttai-agent/            # Grand map: kb (entities, recipes) + features/mcp
    ├── scenario-maker/        # Create / edit / refine (loads ttai-agent)
    ├── session-analyst/       # Report templates
    └── browser-demo-builder/  # Deterministic browser demo steps: format, selectors, example

Related

• Sign up · Developer portal / PAT · Platform docs · llms-full.txt (AI-readable API reference)
• Privacy policy · Terms of service — the canonical legal URLs for Tough Tongue AI, required when submitting the MCP server to the Claude Connectors Directory or a plugin marketplace. Note the paths: /privacy-policy/ and /terms/, not /privacy or /tos.
• voice-ai-quickstart: starter templates for building apps on Tough Tongue AI (Next.js, Flask, co-navigation demo, scenario-as-code CLI) and the toughtongue-ai integration skill for developers embedding the platform.

License

MIT