DeepSeekCode

English | 中文 | 日本語

DeepSeekCode is a DeepSeek-first terminal code agent for the local development loop: inspect a repository, edit files, run checks, review the diff, and keep working from the same terminal.

Public beta status: usable today for Linux/macOS dogfooding and repository work. v0.1.6 ships GitHub Release binaries, a verified GHCR image, TUI and service smoke gates, deepseek quickstart, a release-binary smoke verifier, a verified Homebrew tap, verified npm/npx install paths, and model-backed README launch media. Larger external repo evidence and broader hosted product proof are still product-hardening work.

Try It

npm install -g @deepseek-code/cli
deepseek quickstart
deepseek

No install:

npx @deepseek-code/cli quickstart

Homebrew:

brew tap willamhou/deepseekcode
brew install deepseek
deepseek quickstart

DeepSeekCode interactive REPL builds a playable 2048 game from an empty repo

Generated game, played locally

_{Both recordings come from the same real interactive deepseek chat run against an empty disposable web repo.}

Why It Exists

DeepSeekCode is meant to feel closer to Claude Code CLI or Codex CLI than to a plain chat wrapper. The default path is terminal-first and repo-aware:

• deepseek opens the full-screen coding-agent TUI in a real TTY.
• deepseek chat keeps the line-oriented REPL available.
• deepseek run executes one-shot coding tasks.
• Sessions, threads, events, tasks, usage, and automations are persisted under .dscode/runtime/.
• File tools, patching, diff review, rollback snapshots, todos, hooks, skills, subagents, diagnostics, MCP/ACP, and local runtime APIs share the same permission and recovery paths.
• Shell work supports foreground commands, background jobs, replay, bounded interactive attach, stdin, resize metadata, cancellation, and a local shell-supervisor bridge.

Quick Start

Install with npm:

npm install -g @deepseek-code/cli
deepseek version
deepseek quickstart

Or run without installing:

npx @deepseek-code/cli version
npx @deepseek-code/cli quickstart

Install with Homebrew (verified on macOS x64/arm64):

brew tap willamhou/deepseekcode
brew install deepseek
deepseek version
deepseek quickstart

Or install from source:

cargo install --git https://github.com/willamhou/DeepSeekCode.git --locked
deepseek version
deepseek quickstart
deepseek doctor --json

Or download a release archive:

deepseek update download-plan --version 0.1.6
curl -L -o deepseek-linux-x64.tar.gz \
  https://github.com/willamhou/DeepSeekCode/releases/download/v0.1.6/deepseek-linux-x64.tar.gz
curl -L -o deepseek-linux-x64.tar.gz.sha256 \
  https://github.com/willamhou/DeepSeekCode/releases/download/v0.1.6/deepseek-linux-x64.tar.gz.sha256
shasum -a 256 -c deepseek-linux-x64.tar.gz.sha256
tar -xzf deepseek-linux-x64.tar.gz
./deepseek version

Or run the published container:

docker run --rm ghcr.io/willamhou/deepseekcode:0.1.6 version

For a local checkout:

cargo install --path .
deepseek quickstart
deepseek config init
printf '%s\n' '<api-key>' | deepseek config auth DEEPSEEK_API_KEY --stdin
deepseek config preset auto
deepseek doctor --json

Run a coding task:

deepseek
deepseek chat
deepseek run --preset auto "explain the current repository structure"

Start the local runtime and connect the TUI:

deepseek serve --http --addr 127.0.0.1:13000
deepseek tui --runtime-url http://127.0.0.1:13000

Set DEEPSEEK_API_KEY for real model calls. Local .env files are ignored by git.

What Works

• Full-screen TUI with Plan / Agent / YOLO modes, approval modals, first-run setup guidance, live agent task timeline, command palette, provider/model picker, and MCP management.
• REPL with raw-mode line editing, history, session list/load completion, SIGINT cancellation, /save, /load, /sessions, and custom slash commands.
• OpenAI-compatible single and same-turn batch tool calls, with independent read-only chunks parallelized conservatively and writes/shell/approvals kept as serial barriers.
• Recoverable DeepSeek-style malformed tool-call arguments are repaired through a bounded pipeline and recorded as observable runtime repair events.
• Defensive loop-hardening for weak-model failure modes: redundant re-reads of the same target are short-circuited even when size caps differ, a stuck-directive forces an edit decision after sustained inspection without progress, and apply_patch falls back to a unique-match whitespace-tolerant splice when the model reproduces a find block that is close but not byte-exact. See docs/demo/deepseek-code-calc-bugfix.gif for a recording where the stuck-directive fires and the agent then lands a small-anchor apply_patch.
• Runtime evidence commands: deepseek stats, deepseek events replay, and deepseek events diff summarize cost/cache/tool/failure traces without reading raw .dscode/runtime JSON.
• DeepSeek model presets and estimated-cost budgets sync into runtime session/thread records, so TUI and daemon task sessions can enforce or clear budgets across process restarts.
• Guided first-run checks through deepseek quickstart and deepseek quickstart --json.
• Local HTTP/SSE runtime, ACP stdio adapter, MCP client/server surfaces, and side-effect tooling behind explicit trust/approval controls.
• RLM helpers for recursive and long-input analysis, model-session context, live queue status, event replay, cancellation, recovery, and drain controls.
• CI-smoked Linux/macOS/Windows entrypoints plus release assets for Linux x64, Linux arm64, macOS x64, macOS arm64, and Windows x64.
• Verified model-backed README demos and online multi-file external fixture evidence for the current release-readiness path.

Current Limits

For the Linux/macOS local CLI milestone, the core interaction loop is already in place. The remaining gaps are mainly evidence depth and product hardening:

• optional larger external repo fixtures beyond the disposable Python/Rust/Node samples.
• scope of autonomy is model-bound: the agent is reliable on guided, scoped tasks — greenfield generation (e.g. the 2048 demo), single-file or single-line bugfixes (e.g. the calc bugfix demo), and focused features driven by clear prompts. Open-ended multi-file autonomous repair is currently a weaker surface. The defensive loop-hardening above keeps these failures observable and recoverable, but doesn't close the model-side gap.

Windows long-tail service proof, hosted IDE evidence, and installed service proof are broader product-hardening work. They are not blockers for the Linux/macOS local code-agent CLI milestone.

Evidence

Useful local checks:

cargo fmt --check
cargo test --lib -- --test-threads=1
node scripts/check-secrets.js
deepseek quickstart --json
deepseek dogfood repair-cache-evidence --json
deepseek stats --json
deepseek events replay <thread-id> --limit 50
deepseek events diff <before-thread-id> <after-thread-id> --json
deepseek update publish-status --json
deepseek update release-smoke --version 0.1.6 --json
deepseek tui --entrypoint-smoke --smoke-bin "$(command -v deepseek)"

For release and dogfood evidence, see:

• Release checklist
• Evidence summary
• Dogfood evidence
• Current status
• Additional demos: scripted 2048 capture, TUI recording, TUI first-run UX evidence, model-backed edit/test loop, and guided calc bugfix (stuck-directive + small-anchor apply_patch)

Documentation

• Install
• Public beta guide
• Current status and roadmap
• Launch kit
• Release checklist
• Evidence summary
• Dogfood evidence
• Demo assets
• Architecture
• DeepSeek-native loop design
• Runtime contract
• TUI workbench
• REPL mode
• Agent tasks
• Skills and profiles
• PR / CI integration
• Roadmap
• Changelog

Acknowledgements

DeepSeekCode is independently implemented, but several compatibility surfaces and terminal-agent workflow ideas were informed by Hmbown/CodeWhale, formerly DeepSeek-TUI. DeepSeekCode does not vendor or copy CodeWhale source code; the compatibility work is tracked as interface and workflow parity.

Repository Notes

This repository is public for transparency and collaboration. Public visibility does not imply a separate open-source grant beyond the terms in LICENSE.

Do not commit local credentials, API keys, runtime state, or private .env files. The tracked examples use placeholders only.