thinkany-ai/dscode

What it solves

DSCode is a local-first coding agent designed to provide a secure, cost-effective, and provider-agnostic way to automate software development tasks. It addresses the need for a coding assistant that maintains local control over sessions and security while allowing users to switch between various AI model providers without losing their project context.

How it works

DSCode operates as a runtime that manages local sessions stored as JSONL files. It uses a provider-aware routing system to connect to models like DeepSeek (the default), OpenAI, Anthropic, and others. To ensure security, it executes commands in an OS sandbox (using Seatbelt on macOS or Docker on Linux/Windows) and implements a granular permission system (plan, ask, auto, full) to control what the agent can access.

Who it’s for

It is built for developers who want a powerful coding agent that is MIT-licensed, locally controlled, and optimized for economical models like DeepSeek V4 Flash, while retaining the flexibility to use other high-end providers.

Highlights

• Multi-Provider Support: Native integration with DeepSeek, OpenAI, Anthropic, OpenRouter, and more.
• Local-First Architecture: Sessions are stored locally as tree-shaped JSONL, ensuring transparency and inspectability.
• OS Sandboxing: Commands run in isolated environments with network access blocked by default to prevent unauthorized actions.
• Parallel Execution: Supports running up to four parallel tasks with specialized roles (explorer, implementer, reviewer, tester).
• Cost Awareness: Built-in tracking of context, cache hits, and estimated costs, specifically optimized for DeepSeek's prefix cache.
• Vision Integration: Supports image inputs for vision-capable models via TUI or file attachments.
• Extensibility: Compatible with MCP (Model Context Protocol), Agent Skills, and project-specific instructions via AGENTS.md and CLAUDE.md.