fal-setup

name	fal-setup
description	Wires up fal.ai for image / video generation in cookiedclaw. Saves FAL_KEY to ./.cookiedclaw/keys.env and registers fal's hosted HTTP MCP server in Claude Code's user-scope config. Invoke when the user explicitly asks to set up / install / enable fal.ai in their cookiedclaw workspace.
allowed-tools	Bash(claude mcp *) Bash(mkdir *) Bash(chmod *) Bash(test *) Bash(pwd) Read Write Edit

fal.ai integration

Standalone setup for the fal.ai integration. The user runs this when they want image / video generation (text-to-image, image editing, lipsync, video, etc.) available to the agent.

This skill is independent from /cookiedclaw:setup — call it when the user asks for "set up fal", "add image generation", or similar. It assumes the cookiedclaw workspace already exists (./.cookiedclaw/ directory). If it doesn't, point the user at /cookiedclaw:setup first.

Step 1 — Check workspace + existing config

Run pwd and confirm we're in a cookiedclaw workspace. A workspace has ./.cookiedclaw/keys.env (or at least the ./.cookiedclaw/ directory). If not:

"This doesn't look like a cookiedclaw workspace yet. Run /cookiedclaw:setup first to bootstrap one, then come back here."

Then check for existing config:

• ./.cookiedclaw/keys.env — does it already have FAL_KEY?
• claude mcp list — is fal-ai already registered?

If both are present, ask the user with AskUserQuestion:

• Question: "fal.ai already looks set up. What now?"
• Options:
  • Re-do — replace the key with a new one
  • Cancel — leave it as is

If they pick cancel, exit cleanly.

Step 2 — Get the fal key

Tell the user:

"Open https://fal.ai/dashboard/keys, create a new key, and paste it here."

Wait for the key. Validate format — fal keys typically look like <id>:<secret> or start with fal_. If it doesn't match, ask them to double-check.

Save to ./.cookiedclaw/keys.env:

• mkdir -p ./.cookiedclaw first if needed
• If file already has a FAL_KEY= line, replace it; otherwise append
• chmod 600 ./.cookiedclaw/keys.env

Don't echo the key value back. Just say "saved".

Step 3 — Register the fal-ai MCP server

Use AskUserQuestion:

• Question: "Register the fal.ai MCP server in Claude Code now? (Stores your token in CC's user-scope config so every session can use it.)"
• Options:
  • Yes, register it
  • No, I'll do it manually later

On yes, run:

claude mcp add --transport http fal-ai \
  https://mcp.fal.ai/mcp \
  -s user \
  --header "Authorization: Bearer <value>"

-s user makes it available in every CC session, not just this workspace.

If fal-ai is already registered, run claude mcp remove fal-ai -s user first, then re-add with the new key.

Step 4 — Wrap up

Confirm: "fal.ai set up. After you restart Claude Code, both this Telegram session and any other CC session can call mcp__fal-ai__* tools (run_model, submit_job, recommend_model, etc.) for image / video generation."

Mention briefly:

• The agent can now respond to "generate an image of …" by calling run_model and embedding the result with [embed:<url>].
• For long jobs, use submit_job + check_job so the bot doesn't block on the wait.
• ~/.cookiedclaw/keys.env (the OLD location) is not read anymore — keys live in the workspace's ./.cookiedclaw/keys.env.

Don'ts

• Don't run claude mcp add without confirming the exact command first.
• Don't echo the API key back in confirmations.
• Don't try to add fal-ai to a project-scope (.mcp.json) — user-scope is correct so non-cookiedclaw sessions also benefit.