Developer API / docs
MCP server
One config block gives an AI agent every tool in the catalog. The server reads the live catalog at startup, so it never offers a tool the API cannot run.
Setup
You need an API key from the dashboard. Claude Desktop config:
claude_desktop_config.json
{
"mcpServers": {
"gigai-tools": {
"command": "npx",
"args": ["-y", "gigai-tools-mcp"],
"env": {
"GIGAI_API_KEY": "gk_live_...",
"GIGAI_MCP_OUT_DIR": "/path/where/results/go"
}
}
}
}claude code
claude mcp add gigai-tools -e GIGAI_API_KEY=gk_live_... -- npx -y gigai-tools-mcpWhat the agent gets
Every catalog engine becomes a tool - all of them, from merge_pdf and ocr_pdf to detect_pii and generate_qr. The list comes from the live catalog at startup, so new engines appear without updating the package. Tools take local file paths or URLs. Results are saved to GIGAI_MCP_OUT_DIR and the tool answer names the saved paths - file bytes never pass through the model's context.
Ask the agent things like "merge these three PDFs and give me the file" or "make a QR code for our launch page" and it picks the right tool with the right options.
agent_plan (v0.2.0)
One extra tool, agent_plan, turns a goal sentence into an ordered pipeline of this server's own tools with ready-to-use options - "shrink this photo to under 150 KB as WebP" comes back as resize → convert → compress steps the agent can execute directly. It sends only file metadata (basename, type, size), never contents, and is rate-limited to 10 plans per day per account. When the free-tier planning capacity behind it is unavailable it says so honestly and the agent falls back to picking tools from the list itself.
Source
The server is about 100 lines on top of the Node SDK: github.com/cbsshekhawat18-lab/gigai-tools-mcp. Files sent through it follow the same rules as the rest of the API: contents never logged, deleted after retention, deletion verified.