`execute_chain` turns generate → complete → optimise into a single request, with each step seeing what the last one produced. That is a real saving when a task is three predictable passes rather than a conversation. Worth knowing: the API key is billed by DeepSeek, not by this server, so the built-in caching is the only cost control in the box.
A DeepSeek API wrapper for MCP clients. Three tools cover writing new code, completing existing code, and rewriting code against a target such as performance or readability; a fourth runs several of them in sequence, passing each step's output into the next.
- `generate_code` takes a `prompt`, a `language` and a `temperature`
- `complete_code` continues existing `code` given a `prompt` — for example adding input validation and error handling
- `optimize_code` rewrites code against a `target`, with `performance` and `security` shown as values
- `execute_chain` runs a list of `steps`, each naming a `toolName` and its params, with context carried between them
- Request caching with a TTL, plus caching of chain results, so repeated work does not re-bill the API
Node.js >= 18.0.0 and a DeepSeek API key. Published as deepseek-mcp version 1.0.0; the documented path is clone, `npm install`, `npm run build`, then point your client's `command` at node and its `args` at `build/index.js`. The key goes in the server entry's `env` as `DEEPSEEK_API_KEY`.
One command plus a key — npx -y deepseek-mcp, then supply credentials