# deepseek-mcp

A local MCP server that makes DeepSeek available to your primary coding agent
for code review and second opinions.

The server reads the repo itself. Your agent passes a git ref or a list of
paths — never the code — so a review costs about twenty tokens on the way in
instead of several thousand.

## Tools

| Tool | Use it for | Model |
|---|---|---|
| `review_changes` | Review a diff. Uncommitted work by default, or a branch via `base`. | review |
| `review_files` | Review whole files — architecture, an unfamiliar module, a suspected bug. | review |
| `consult` | Ask a specific question. Supports multi-turn via `history`. | consult |

All three are read-only and take paths, not content.

## Install

```bash
npm i -g @berrydev-ai/deepseek-mcp
```

That puts a `deepseek-mcp` binary on your PATH; no checkout needed. Or skip
the install and let your MCP client fetch it on demand with
`npx -y @berrydev-ai/deepseek-mcp`, at the cost of a slower cold start.

One caveat if you install from source instead: avoid
`npm i -g git+https://github.com/berrydev-ai/deepseek-mcp.git`. On npm 10.x
it appears to succeed but symlinks the global package at a bare clone in the
cache with no working tree, leaving `bin` dangling. Use the registry, or a
tarball. `dist/` is committed so that installing from git or a tarball still
works — a global install has no devDependencies, so nothing can compile at
install time.

## Configure

There is no dotenv loader. The server reads `process.env` and nothing else,
so `.env` is a reference for what to set rather than a file that gets read.
Configuration belongs in your MCP client's `env` block.

### Claude Code

```bash
claude mcp add deepseek -s user -e DEEPSEEK_API_KEY=sk-... -- deepseek-mcp
```

`-s user` registers it across all your projects. Without it the default is
local scope, which is limited to the directory you ran the command in.

### Codex

```bash
codex mcp add deepseek --env DEEPSEEK_API_KEY=sk-... -- deepseek-mcp
```

Note `--env` rather than Claude's `-e`, and that the command to launch goes
after `--` in both. Codex writes to `~/.codex/config.toml` globally, so there
is no scope flag to think about. The equivalent by hand:

```toml
[mcp_servers.deepseek]
command = "deepseek-mcp"

[mcp_servers.deepseek.env]
DEEPSEEK_API_KEY = "sk-..."
```

Verify with `codex mcp get deepseek`, and `codex mcp remove deepseek` to undo.

### Any other MCP client

```json
{
  "mcpServers": {
    "deepseek": {
      "command": "deepseek-mcp",
      "env": {
        "DEEPSEEK_API_KEY": "sk-..."
      }
    }
  }
}
```

Note that a missing key does not stop the server from starting, so the client
will report it as connected either way — `assertConfigured` only fires on the
first tool call. If every call comes back with "DEEPSEEK_API_KEY is not set",
the `env` block is the place to look.

The server operates on its working directory, which is normally your project.
Set `DEEPSEEK_REPO_ROOT` to override.

## Local development

```bash
git clone https://github.com/berrydev-ai/deepseek-mcp.git
cd deepseek-mcp && npm install && npm run build
```

Run `npm run build` and commit `dist/` alongside any source change, otherwise
installs keep serving the previous build.

## Direct or through AI Gateway

One env var, no code change. `https://api.deepseek.com` is the default and has
no dependencies. Pointing `DEEPSEEK_BASE_URL` at a Cloudflare AI Gateway
endpoint instead gets you request logs, spend caps, caching and the ability to
swap models without touching this repo — worth it the moment you want to
compare two models on the same review.

## Tuning the prompts

`prompts/review.md` and `prompts/consult.md` are the entire behaviour of the
server. They are plain markdown; edit and restart. Both are written to keep
output tight, because whatever DeepSeek returns lands in your primary agent's
context window and competes with everything else there.

The review prompt enforces severity grouping and `file:line` anchors, and
forbids restating what the code does. If reviews come back vague, the fix is
almost always in that file rather than in the TypeScript.

## Getting it actually invoked

The most common failure mode is not a bug — it is that the primary agent never
calls the tool. Coding agents are reluctant to delegate. The tool descriptions
are written to draw a clear line (substantial changes yes, one-line edits no),
but you may still need to ask for it explicitly at first, or add a line to your
`CLAUDE.md` telling the agent to get a DeepSeek review before opening a PR.

Delegation only pays for itself on work that involves reading and analysing
real amounts of code. On small tasks the per-call overhead dominates.

## Guards

- Paths are resolved and rejected if they escape the repo root.
- Git refs are validated so an argument cannot be smuggled in as an option.
- Input is truncated at a character budget, and the response says when that happened.
- Model calls abort at 55s, below the ~60s tool-call timeout most clients enforce,
  so you get a clear message rather than a hang.
- Every response carries a footer with file count, model and token usage.

## Notes

Built on `@modelcontextprotocol/server` v2 via `serveStdio`, which serves both
the 2025 and 2026-07-28 protocol revisions from one factory — clients on either
revision work without configuration.

Model defaults are `deepseek-v4-flash` for reviews and `deepseek-v4-pro` for
consults. Check them against whichever endpoint you point at — the names move
over time.

Watch out for the older `deepseek-chat` and `deepseek-reasoner` names. They
still return 200, but the response comes back as `deepseek-v4-flash`, so a
stale value costs you the reasoning model without ever raising an error. The
model in the response footer is read from the provider's reply rather than
from your config, so it will tell you what actually ran.

## License

MIT — see [LICENSE](LICENSE).