deepseek — streaming CLI for delegation

A small CLI that lets Claude Code (or any orchestrator) delegate bounded chunks of work to DeepSeek, streaming. Replaces the older deepseek-agent-mcp pattern: instead of a black-box agentic loop running inside an MCP tool call (which times out, hides progress, and can't recover from partial failure), the orchestrator drives the loop via small visible CLI invocations.

Why CLI, not MCP

The previous version of this repo was an MCP server with deepseek_agent, deepseek_chat, web_fetch, web_search, web_fetch_raw tools. The agent tool didn't fit MCP's contract:

	MCP function call	Long agentic loop
Time	Bounded, atomic	Unbounded
Visibility	Final response only	Need progress
Recovery	Retry the whole thing	Resume mid-loop
Streaming	None	First-class

Squeezing a multi-turn agent loop into a function-shaped tool meant Claude waited 5+ minutes blind, transport timed out, partial work was lost. So we ripped it out. The CLI gives back streaming, visibility, recoverability — and Claude becomes the loop, which it was already doing one level up anyway.

The bounded tools (chat, fetch, search) became CLI subcommands too, for consistency.

Subcommands

deepseek chat       single-turn chat (streams)
deepseek fetch      URL → cleaned text → deepseek summary (streams)
deepseek fetch-raw  URL → cleaned text (no model)
deepseek search     Brave + deepseek synthesis (streams)

Every model-using subcommand accepts:

-m, --model <name>           per-subcommand default if omitted
--thinking <true|false>      omit → API default = on for v4 models
--reasoning-effort <l|m|h>   only meaningful when thinking is on
--max-tokens <n>             cap output
--show-thinking              stream reasoning_content to stderr
--system <text>              (chat only) system prompt
--temperature <n>            (chat only) sampling temp

Prompts can be a positional arg or piped via stdin.

Defaults

• chat → deepseek-v4-pro (when you explicitly ask, you want the better brain)
• fetch, search → deepseek-v4-flash (high volume, summarization is cheap work)
• All v4 models have thinking mode on by default — pass --thinking false to opt out.

Pricing (per 1M tokens, cache miss):

Model	Input	Output
v4-pro	$1.74 ($0.435 promo through 2026-05-31)	$3.48 ($0.87 promo)
v4-flash	$0.14	$0.28
Cache hits	1/10 of input price	—

Examples

# Single question
deepseek chat "What does this regex match: ^(?=.*[A-Z])(?=.*\d).{8,}$"

# Cheap & fast
deepseek chat "summarize this in one sentence" --thinking false --max-tokens 100 < article.txt

# URL summary
deepseek fetch "https://docs.example.com/v3" "list breaking changes" > breaking.md

# Web research
deepseek search "deepseek v4 thinking parameters" -n 8 > research.md

# See the model's reasoning
deepseek chat "what is 7*8" --show-thinking
# stdout: 56
# stderr: We are asked: "what is 7*8"...

# Override default model
deepseek chat "translate 'hello' to japanese" --model deepseek-v4-flash --thinking false

Install

git clone https://github.com/luongjames8/deepseek-mcp-server.git
cd deepseek-mcp-server
npm install
npm run build
npm link        # puts `deepseek` on PATH globally

Set DEEPSEEK_API_KEY in your environment (and optionally BRAVE_API_KEY for search):

echo 'export DEEPSEEK_API_KEY="sk-..."' >> ~/.bashrc
echo 'export BRAVE_API_KEY="..."' >> ~/.bashrc

Or drop a .env file in the project root, your cwd, or $HOME — the CLI auto-loads from any of those.

Use with Claude Code

Add a permission rule to ~/.claude/settings.local.json so Claude doesn't prompt on every call:

{
  "permissions": {
    "allow": ["Bash(deepseek:*)"]
  }
}

Then drop a skill at ~/.claude/skills/deepseek/SKILL.md (the one in this repo's claude-skill/ is a starting point) telling Claude when to delegate. Claude reads the skill on relevant prompts and invokes deepseek via Bash when appropriate.

Configuration

Per-subcommand defaults live in config.yaml at the repo root, project cwd, or ~/.config/deepseek-cli/config.yaml:

chat:
  default_model: "deepseek-v4-pro"

web_search:
  default_model: "deepseek-v4-flash"
  max_results: 10

web_fetch:
  default_model: "deepseek-v4-flash"
  timeout_seconds: 15
  max_content_chars: 50000

Every value is overridable per-call via flags.

Tests

npm test

25 vitest cases (~13s with API key, ~1s without):

• Unit: param translation, validation, config loading
• Smoke (real API): spawns the actual CLI binary against DeepSeek/Brave. Catches request-shape regressions, broken streaming, model-routing changes upstream.

Smoke tests skip cleanly if DEEPSEEK_API_KEY is unset.

Sharp edges

• Legacy aliases die 2026-07-24: deepseek-chat and deepseek-reasoner will stop resolving. Use deepseek-v4-pro or deepseek-v4-flash explicitly. Defaults are already correct.
• fetch min content size: 500 chars. Short pages (status pages, redirects) error out — use fetch-raw to bypass parsing.
• No interactive prompts: every call is one-shot. The CLI doesn't keep conversation history. If you want multi-turn, the orchestrator manages it.
• Strict tool calling (Beta): not exposed via flags. Set DEEPSEEK_BASE_URL=https://api.deepseek.com/beta to use the strict endpoint.

License

MIT