Deep Copilot
Chat with DeepSeek in VS Code to inspect, search, and edit project code.
- Source repo
- deep-copilot/DeepCopilot
- Stars
- ★ 79
- Last updated
- 5mo ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 60/100 · Some gaps
At a glance
- How it runs
- Works with
- Platform-specific
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A VS Code user wants DeepSeek to inspect workspace code, explain files, or search for symbols and text.
- Not a fit if
- Teams required to use OpenAI or Anthropic APIs
- Website
- marketplace.visualstudio.com/items ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-05-31latest release v0.43.1 (2026-05-31)
- Source review
- 60/100 · Some gaps
What does this agent do, and when should you use it?
Deep Copilot is a bilingual conversational coding extension that runs in the VS Code extension host and calls DeepSeek V4 Pro, Flash, or Reasoner through the DeepSeek API, which uses an OpenAI-compatible protocol. Its sidebar chat streams replies, reasoning content, and tool progress over SSE. The agent can read and edit workspace files, search code, and run shell commands, with writes and commands governed by configurable approval modes. It also supports Tavily web search and stdio connections to MCP tool servers. The extension has no runtime npm dependencies and needs no separate backend, but it requires VS Code 1.95 or later and a DeepSeek API key.
Open the Deep Copilot panel in VS Code and configure a DeepSeek API key, then send a request or attach context such as files, selections, diagnostics, or Git changes. Its agent loop can call tools including read_file, grep_search, write_file, apply_patch, and run_shell. Writes and shell commands follow the configured approval mode; after edits, LSP diagnostics can be appended for the model to review. Replies and tool results stream over SSE. Users can view Plan & Todos, manage sessions, revert file edits from the current turn, and manage context with /context, /compact, and /fork. Add a Tavily key to use web_search, or configure MCP servers through deepseekAgent.mcp.servers.
- A VS Code user wants DeepSeek to inspect workspace code, explain files, or search for symbols and text.
- A developer wants an agent to edit files or run shell commands under an approval policy, then inspect LSP diagnostics.
- A developer wants to investigate code in read-only Plan mode before deciding whether to make changes.
- A developer switching between tasks wants sessions to keep running and replay progress when reopened.
- A VS Code user needs to connect external MCP tools or use Tavily web search.
How do you install or deploy this agent?
In VS Code, search for Deep Copilot in the Extensions view and install it, then open its sidebar panel to configure a DeepSeek API key. You can also install a prebuilt VSIX from GitHub Releases:
code --install-extension deep-copilot-0.41.6.vsixBuilding from source requires Node.js 18 or later, npm 9 or later, and VS Code 1.95 or later:
git clone https://github.com/ZhouChaunge/DeepCopilot.git
cd DeepCopilot
npm install
npm run build
npm run package
code --install-extension deep-copilot-0.41.6.vsixHow do you use this agent?
After installation, click the 🐋 icon in the activity bar to open the panel. Click the 🔑 button at the bottom right, enter your DeepSeek API key, and save; this key is required for chat and agent tools. Send a request to start a conversation. To enable web search, enter an optional Tavily API key in the same settings dialog. Configure the tool approval mode in VS Code settings under deepseekAgent.approvalMode. MCP stdio servers can be configured like this:
{
"deepseekAgent.mcp.servers": [
{ "name": "my-db", "command": "npx", "args": ["my-db-mcp-server"] }
]
}What are this agent's strengths and limitations?
- Includes file reading, targeted replacement, apply_patch, ripgrep search, and shell tools with configurable approval policies.
- Provides read-only Plan mode, post-edit LSP diagnostics, and current-turn file rollback for investigation and review.
- Supports stdio MCP tool servers, optional Tavily search, and switching between running sessions.
- Has no runtime npm dependencies and requires no separately deployed backend.
- Model calls depend on the DeepSeek API and a network connection; API usage may incur service charges.
- It is delivered only as a VS Code extension and requires VS Code 1.95 or later.
- The agent can run shell commands and read or write workspace files; effective permissions depend on configuration.
- The README documents DeepSeek as the model provider and does not describe adapters for other providers; Tavily search requires an additional API key.
How does this agent compare with similar options?
The README presents Deep Copilot as a Copilot-style VS Code coding assistant and mentions a GitHub Copilot-style pending-edits panel and input experience. It does not provide feature or performance comparisons with other products.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Deep Copilot This agent | 60 · Some gaps | IDE extensionFree + model costs | ★ 79 | 5mo ago | JavaScript | — |
| wcgw Local Coding Terminal | 47 · Major gaps | MCP serverFree + model costs | ★ 679 | 2mo ago | Python | Claude.ai |
| Kode CLI | 60 · Blocked | CLIFree + model costs | ★ 5.2k | 18d ago | TypeScript | OpenAI API · Claude API |
| CoreCoder | 64 · Some gaps | CLIFree + model costs | ★ 1.9k | 1d ago | Python | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README and package manifest describe file, shell, MCP, and web tools, so the permission surface is broad and least privilege scores 1. Manual approval is the default, with read-only mode, deny lists, and per-turn rollback; however, auto-edit/autopilot can authorize actions, and the README explicitly says dangerous commands are silently allowed in autopilot, so confirmation scores 2. The README describes DeepSeek/Tavily data paths and log locations, but debug logging is enabled by default and records reasoning, tool, and API events; terminal history is also retained in memory by default, leaving sensitive-data handling thin. No runtime npm dependencies, a lockfile, and SHA-pinned CI Actions support a dependency-security score of 2. Shell, file edits, and MCP have broad external effects, while rollback covers only current-turn file changes, so external effects score 1 and rollback 2. Repository, copyright, and package publisher fields provide attribution clues, but the publisher is unverified per the prompt and maintenance/identity ownership remain unclear, so source attribution scores 1.
The README and package configuration disagree on defaults such as maxIterations, compactBudgetTokens, and debug logging; claims of no backend also coexist with legacy server settings and commands, so self-consistency scores 1. Node/VS Code prerequisites, the lockfile, and the no-runtime-dependency claim are documented, supporting dependency availability at 2. The code logs or falls back to heuristics for provider/token-counter failures, but the packet does not establish comprehensive user-facing failure messages, so that criterion scores 1.
Bilingual documentation covers chat, code editing, search, planning, read-only investigation, completions, and approval modes; audience and scenarios score 3. Read-only mode, tool descriptions, and settings establish some boundaries, but autopilot risks, unlimited iterations by default, and broad MCP tools leave them partial, scoring 2. VS Code commands, slash commands, context references, and skill invocation have concrete entry points, scoring 2 for trigger precision. VS Code and Node build requirements and key setup are clear, but silent shell defaults and platform-dependent terminal capabilities limit environment fit, scoring 2.
The README has a table of contents, tool table, architecture diagram, and bilingual sections, earning 3 for information architecture. Marketplace, VSIX, and source installation steps are given, earning 3 for install notes. Deep Copilot, deepseekAgent, and DeepSeek V4 naming coexists with some server/backend remnants, so naming stability scores 1. Configuration, keybindings, tools, and troubleshooting are covered, but examples/FAQ support is incomplete, scoring 2. Risks around dangerous auto-approval, terminal history, clipboard use, and iteration cost are disclosed, though limitations are not comprehensive, scoring 2. The MIT license is complete, scoring 3. A changelog and version are visible, but the presented change history is limited, scoring 2. Copyright and repository publisher fields offer clues without clearly assigning maintenance responsibility or support commitments, scoring 1.
The agent can read and edit files, run commands, search, present plans, and roll back; output usability scores 2 because results still depend on the model and tool outcomes. Multi-turn tool use, reviewable edits, diagnostics, and session features offer clear coding value, scoring 3 for marginal value. A required DeepSeek API key, optional Tavily key, unlimited iterations by default, and background-task budgets leave costs variable; telemetry helps but does not remove that exposure, so cost-benefit scores 1.
The README and settings give concrete descriptions of many features, defaults, and tools, with some supporting implementation files, so claim traceability scores 2. The packet is a sample and provides no execution evidence; README and manifest settings also differ, limiting cross-source corroboration to 1. Product claims are generally stated as features with little separation between established facts and inference, so fact/inference separation scores 1.
- Debug logging is enabled by default and records reasoning, tool, and API events; in autopilot mode, dangerous shell commands listed in the README are silently allowed.
- Several defaults differ between the README and package.json; this assessment uses only the supplied static material and includes no execution or independent verification.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.43.1 Deep Copilot v0.43.1
Release notes excerpt
Deep Copilot v0.43.1 > 中文版在下方 / English notes below. 🇨🇳 中文 主题:工具调用前穿插逐步文字解说 + 后台唤醒调度 + DeepSeek 前缀缓存调优 > 本版本主要打包 PR #180(v0.43.0)的全部能力,并以 0.43.1 重新构建 VSIX。 ✨ 新特性 1:每个工具调用前都有「人话解说」 此前多步工具回合常常表现为「一堆工具先跑,最后才蹦出文字」。日志里每个 ITER_END 的 assistant_chars:0 说明模型把「我现在要做 X」的解说全塞进了 DeepSeek 的隐藏 reasoning_content 通道,可见正文 content 全程为空。 修复(src/prompts/system.js): - Tone 段:把原来禁止前导解说的 No preamble, no "I'll now…" 改成「禁空话(Great question/Sure),但要求动作解说」——每次工具调用前用一句可见正文说明「要做什么 + 为什么」。 …View on GitHub ↗ -
v0.41.6 v0.41.6 — Archive pure-export + watchdog stays on duty
Release notes excerpt
Deep Copilot v0.41.6 > 中文版在下方 / English notes below. 🇨🇳 中文 主题:Archive 改为纯导出 · 值守任务不再提前结束会话 · 打包卫生 🐞 修复 1:Archive 语义改为「纯导出」(Issue #169) 旧版「存档」实际上是「软隐藏 + 导出」——点一下生成 Markdown,同时把会话从侧边栏抽走,连点两次第二次还会把它「unarchive」回来,体验和菜单文案完全不符。 新行为(PR #170): - archive(id) 只做一件事:调用 exportSessionToMarkdown 把整个会话渲染为 Markdown,写到 .deep-copilot/archives/yyyyMMdd-HHmmss-<title>.md。 - 不再 把 archived 置为 true,不再 把当前激活的 session 切走,不再 广播 sessionLoaded 空消息。 - 出错或用户在保存对话框里取消,session 状态都保持原样——只 toast 一下。 - 同一会话连点两次「存档」 → 生成两个 md 文件,会话本体不动。 一次性迁移 _migrateArchivedFlagIfNeeded: …View on GitHub ↗ -
v0.41.0 v0.41.0 - Context Window Overhaul
Release notes excerpt
Deep Copilot v0.41.0 > 中文版在下方 / English notes below. 🇨🇳 中文 主题:Context Window 优化套件 · 三个新斜杠命令 · 底部 Context Ring · 会话切换闪屏修复 ✨ 新功能 1:上下文窗口管理重构(Issue #142) 长对话越来越容易吃满 token 窗口,0.41.0 把"上下文管理"做成了一套完整的机制: - 结构感知截断(structure-aware truncation) - 保留最新工具调用的完整内容 - 较旧的轮次自动折叠为"摘要骨架"(保留语义,丢弃大段冗余) - 同文件读取去重 - 多次读取同一路径时,只保留最后一次完整内容 - 较早的读取被折叠为占位符:<file path=... read-collapsed='true'/> - 滚动摘要(rolling summary) - 超阈值时按需调用模型,把更早的对话压缩为结构化摘要节点 - 会话历史持久化层(session-store.js)已适配,重新加载会话能正确回放摘要节点 - MCP per-server opt-out:可以为某个 MCP server 显式关闭上下文压缩参与 …View on GitHub ↗
FAQ
Does using Deep Copilot cost money?
Can the agent edit files or run commands without confirmation?
Can it run without connecting to an external service?
Can I connect other tools or switch model providers?
Is Deep Copilot open source?
Is Deep Copilot safe to use?
Is Deep Copilot still maintained?
Who makes Deep Copilot?
What is Deep Copilot's official website?
Thanks — we use this to improve the page.