Deep Code CLI
A terminal AI coding assistant optimized for deepseek-v4, with deep thinking, reasoning effort control, Agent Skills, and MCP support.
Evidence shows a permission control mechanism (permissions field with always allow/ask/deny), but no default policy or least-privilege principle is detailed, so least_privilege scores 1. User confirmation exists (AskUserQuestion tool, permission prompts), with tests covering related logic, scoring 2. Data flow transparency is insufficient: README mentions config and notification scripts but does not explain how data is sent to external APIs or stored locally, scoring 1. Sensitive data handling: API_KEY is stored in settings.json without encryption or protection details, scoring 1. Dependency security: dependencies are recent but no security audit or vulnerability scan evidence, scoring 1. External effects: supports shell commands, file writes, network access, but permission control can limit, scoring 1. Rollback: /undo command exists but implementation details or reliability not described, scoring 2. Source attribution: README and LICENSE clearly attribute to lessweb, but publisher unverified, scoring 1.
Self-consistency: README and package.json names/descriptions are consistent, but version inconsistency (README mentions @vegamo/deepcode-cli, package.json is @vegamo/deepcode), scoring 2. Dependency availability: dependencies are common but no lockfile or integrity checks, scoring 1. Failure messages: tests cover error handling but no user-visible error message examples, scoring 1.
Audience and scenarios: targets developers, provides CLI and VSCode extension, clear scenarios, scoring 2. Capability boundaries: README lists supported models and features but no explicit limitations, scoring 2. Trigger precision: slash commands and shortcuts are clearly defined, scoring 2. Environment fit: supports multiple platforms (CI matrix) but no system requirements, scoring 2.
Information architecture: clear documentation structure with README and docs directory, scoring 2. Install notes: provides npm install command, scoring 2. Naming stability: package name inconsistent between README and package.json, scoring 1. Examples and FAQ: multiple FAQs and config examples, scoring 2. Known limitations: not explicitly listed, scoring 1. License: MIT license file present, scoring 2. Versioning/changelog: version number exists but no CHANGELOG, scoring 1. Maintenance responsibility: CI and contribution guide present, but publisher unverified, scoring 1.
Output usability: interactive CLI with various commands, output format not detailed, scoring 2. Marginal value: optimized for DeepSeek, unique features, scoring 2. Cost-benefit: claims cost reduction but no concrete data, scoring 1.
Claim traceability: README references architecture article and benchmark, but no detailed data, scoring 1. Cross-source corroboration: references external projects but no independent verification, scoring 1. Fact-inference separation: README distinguishes facts and opinions but not clearly, scoring 1.
- API_KEY is stored in plaintext in settings.json; consider using environment variables or a secret manager.
- Package name inconsistency between README and package.json may cause installation confusion.
- No CHANGELOG provided, version changes are not transparent.
- Default permission policy is not explicit; review default configuration.
What does this agent do, and when should you use it?
Deep Code CLI is a terminal-based AI coding assistant finely tuned for DeepSeek models. Installed globally via npm, it runs in any project directory and leverages DeepSeek's thinking mode, reasoning effort control, and context caching for cost efficiency. It offers a rich set of slash commands for managing conversations, models, and skills. The tool natively scans Agent Skills from project and user directories, integrates with MCP for external services, and shares its configuration with a companion VSCode extension. Permissions, notifications, and web search can be customized via a single settings file. Licensed under MIT, it focuses specifically on maximizing DeepSeek's performance in coding workflows.
Deep Code CLI provides an interactive AI coding assistant in the terminal. It reads configuration from ~/.deepcode/settings.json (including MODEL, BASE_URL, API_KEY, thinkingEnabled, reasoningEffort), and calls the DeepSeek API (default deepseek-v4-pro or deepseek-v4-flash) for conversational coding. It supports context caching and thinking mode, with /model to switch models and reasoning effort. Slash commands like /new, /resume, /fork, /continue manage conversations, while /undo restores code and chat state. Agent Skills are scanned from multiple paths (.deepcode/skills and .agents/skills) and listed via /skills. It integrates MCP servers (configured via mcpServers) and shows their status with /mcp. It supports pasting images from clipboard with Ctrl+V, and can execute a notify script after tasks complete. Granular permission policies control shell commands, file writes, and network access via the permissions field.
- Developers using DeepSeek models who want a coding assistant specifically tuned for their chosen model to improve code quality and efficiency.
- Users who need multi-turn conversation-based coding in the terminal, especially for complex tasks requiring deep thinking mode.
- Developers who want to extend the assistant's capabilities using Agent Skills with custom scripts or tools.
- Users who need to connect to external services (e.g., GitHub, browsers, databases) via MCP within their coding workflow.
- Developers using VSCode who want a consistent configuration and seamless switching between the terminal CLI and the editor extension.
- Security-conscious users who want to control AI actions (shell commands, file edits) through fine-grained permission settings.
What are this agent's strengths and limitations?
- Specifically optimized for DeepSeek models, with context caching to reduce costs and native thinking mode support; benchmarks suggest better outcomes than generic tooling combined with DeepSeek.
- Supports Agent Skills and MCP integration, enabling a highly extensible coding workflow.
- Fine-grained permission control and notification scripts enhance safety and automation.
- Shares configuration with a VSCode extension, allowing seamless switching between terminal and editor.
- Heavily tied to the DeepSeek model ecosystem; support for other models is limited and may not fully leverage its optimizations.
- Depends on the DeepSeek API as the primary backend, requiring network access and an API key, creating external service dependency.
- As a relatively new project, documentation may be less comprehensive, and advanced configuration (MCP, permissions) requires reading extra docs.
- Terminal-only interface with no GUI, which may have a learning curve for users unfamiliar with command-line tools.
How do you install or deploy this agent?
Install globally via npm: npm install -g @vegamo/deepcode-cli. Node.js is required.
How do you use this agent?
Run deepcode in any project directory to start the interactive terminal UI. First, create a configuration file at ~/.deepcode/settings.json with env.MODEL (e.g., deepseek-v4-pro), env.BASE_URL (e.g., https://api.deepseek.com), env.API_KEY, and optional thinkingEnabled and reasoningEffort. Then use slash commands (e.g., /new, /model, /skills) and keyboard shortcuts (Enter to send, Shift+Enter for newline, Ctrl+V to paste images, Esc to interrupt).
How does this agent compare with similar options?
Compared to general-purpose coding tools (e.g., Claude Code), Deep Code may deliver better performance and efficiency specifically with DeepSeek models thanks to its tailored architecture, though it may have a smaller feature ecosystem and community.