Taichu MCP
Use GWC's full multimodal stack — image, video, voice, 3D, multi-agent orchestration — from inside Claude Code, Codex CLI, Cursor, or any MCP-aware editor. One token, five tools, your AI workforce in your terminal.
What it does
Taichu MCP is a Model Context Protocol server that exposes GWC capabilities as native tools to any MCP-compatible host. The host (Claude Code, Codex, Cursor) calls Taichu the same way it would call any other tool. Taichu routes the request to the right model, bills your GWC balance, and returns the result.
Five tools ship by default:
taichu_chat{ prompt, model?, system? }Text completion. Routes to your default Taichu model (DeepSeek V4 Pro by default).
taichu_image{ prompt, aspect_ratio?, model? }Image generation via Seedream 5.0 (default), imagen-4, qwen-image, or z-image-turbo.
taichu_video{ prompt, duration?, model? }Short video generation via Seedance 1.5 Pro, veo-3, or wan-t2v.
taichu_task{ goal, context? }Submit a high-level goal. MAOE decomposes it across specialist agents and returns a runId.
taichu_balance{ }Show GWC USD balance + monthly allowance remaining.
Setup in 90 seconds
1. Get a GWC API key. Sign in at gwc-corp.com/app/settings/api-keys and click Generate Key. Copy the gwc_live_… token shown once.
2. Run the setup wizard. If you have GWC Studio installed, the taichu CLI is already on your PATH. Otherwise install it via npm i -g @gwc/taichu.
# One-time setup. Opens an interactive wizard. $ taichu setup # Or step-by-step: $ taichu login # paste your gwc_live_ token $ taichu config --set defaultModel=deepseek-v4-pro $ taichu mcp # starts the MCP server (stdio) $ taichu whoami # confirm sign-in $ taichu logout # remove local credentials
3. Restart your AI CLI. Claude Code, Codex, and Cursor scan their MCP config at startup. New tools appear under the taichu_* prefix.
Manual configuration
If you'd rather not run taichu setup, drop the config in yourself.
Claude Code
// ~/.claude.json
{
"mcpServers": {
"taichu": {
"command": "taichu",
"args": ["mcp"],
"env": {
"GWC_API_KEY": "gwc_live_..."
}
}
}
}Codex CLI
// ~/.codex/config.json
{
"mcpServers": {
"taichu": {
"command": "taichu",
"args": ["mcp"],
"env": {
"GWC_API_KEY": "gwc_live_..."
}
}
}
}Cursor
Cursor stores MCP config in its settings DB, not a file. Open Settings → MCP → Add Custom Server, then paste:
// Cursor → Settings → MCP → "Add Custom Server"
{
"taichu": {
"command": "taichu",
"args": ["mcp"],
"env": {
"GWC_API_KEY": "gwc_live_..."
}
}
}What a real call looks like
From Claude Code:
> use taichu to make a hero image for my landing page and a 5s loading animation, and translate the headline "Move money like data" into Spanish, French, and Japanese. [Claude Code calls taichu_image, taichu_video, taichu_chat ×3 in parallel] ✓ Image: https://gwc-corp.com/m/abc123.png ✓ Video: https://gwc-corp.com/m/def456.mp4 ✓ es: "Mueve el dinero como los datos" ✓ fr: "Bouge l'argent comme les données" ✓ ja: "データのようにお金を動かす" Cost: $0.034 (deducted from your monthly $15 allowance)
Picking the main AI model
Taichu's text model is the one used by taichu_chat and any agentic loop that doesn't specify a model. You can set it three ways, in resolution order:
- Per call — pass
{ model: "gemini-3.1-pro-preview" }in the tool arguments. Highest priority. - Per host — set
TAICHU_DEFAULT_TEXT_MODELin your MCP configenvblock. - Per account — set the default at gwc-corp.com/app/settings. Syncs across web, mobile, Studio, and MCP.
Billing & cost transparency
Every MCP tool call is charged the same way as a call from the GWC web or mobile app — your monthly allowance first, then USD balance. Text responses include a [model: …, tokens: in→out, cost: N¢] suffix so the host can see what each call cost.
If you'd rather route through your own provider keys, add them via BYOK in GWC Studio's settings (desktop app) and Taichu will pass-through at provider cost (no markup).
Security
- API keys are stored locally in
~/.gwc/credentials.jsonwith file mode 0600. - Tokens are validated against
api.gwc-corp.comon every call — no offline acceptance. - The MCP server uses stdio only — no inbound network port opens on your machine.
- Each token can be revoked from the keys page; revocation is enforced within seconds.
- Rotate tokens by issuing a new one and revoking the old.
Troubleshooting
Tools don't appear in Claude Code
Claude Code reads MCP config on startup only. Restart it after changing ~/.claude.json. Verify with:
$ taichu mcp # should print "Taichu MCP server starting (api=...)" to stderr
# if you see "No GWC auth token", run taichu login first"HTTP 401: Invalid token"
The token was revoked or never valid. Re-mint at /app/settings/api-keys, then run taichu login.
"HTTP 402: Insufficient balance"
You ran out of allowance and have no USD balance. Top up at /app/billing or wait for the next monthly refresh.
Slow first call
Cold-start on the GWC API can be ~1s for the first request after idle. Subsequent calls reuse the warm instance.
Related
- GWC SDK — programmatic access from Node/Python/Go
- REST API — raw HTTP endpoints
- MCP overview — what MCP is, and how GWC exposes other servers
- Studio Guide — the desktop IDE that ships
taichu