dao-code
Open-source TypeScript terminal coding agent for DeepSeek-V4 — builds on DeepSeek's strong price-performance and ultra-cheap cache pricing, engineering byte-stable prefixes and cache-reusing forks so cross-session memory and a continuous self-correction layer add almost no token cost; 1M context, Skills/MCP/Hooks, Claude Code config compatible.
A terminal AI coding assistant that reads, writes, and fixes code using DeepSeek's V4 model. It gives you a Claude Code-like experience for 18-30x less cost by aggressively optimizing cache efficiency.
Dao Code is a terminal-based AI coding assistant that reads, writes, and fixes code for you using DeepSeek's V4 model. The core pitch is simple: it gives you a capable AI pair programmer in your terminal, but at a fraction of the cost of similar tools. Where most premium AI coding agents charge premium rates, this one leverages DeepSeek's cheaper pricing and aggressively engineers around its caching system to drive costs down even further — claiming real-world bug-fix tasks cost roughly 18-30x less than running the same work through Claude.
Under the hood, it works like other terminal coding agents: you give it a task, it streams its reasoning, calls tools (reading files, running commands, editing code), and loops until the job is done. Every write or shell command passes through an approval gate unless you flip on auto-approve mode. What makes it different is the obsession with cache efficiency. DeepSeek charges drastically less when the AI reuses the same "prefix" of context (like system instructions and tool definitions) — about 1/120th the price of fresh input. The agent keeps that prefix byte-stable across turns and runs its memory and self-reflection checks as "forks" that reuse the same cached prefix, so extra quality checks add almost nothing to the bill.
Beyond cost, it has a self-verifying memory system: at the end of a session it distills your preferences and project conventions, then on the next startup it checks those memories against your actual code and drops anything stale. It also has a reflection layer that kicks in when it gets stuck or starts drifting from the original goal — a skeptical "challenger" review that questions its own premises rather than blindly retrying the same approach.
The people who'd use this are developers who want a Claude Code-like experience but either find it too expensive or can't easily access it due to regional restrictions. It's especially aimed at users in mainland China, where DeepSeek is directly accessible without workarounds. If you're a solo developer or small team doing real bug-fixing work and your eyes water at API costs for top-tier models, this is built for you.
Where it fits
- Fix bugs in your codebase at a fraction of standard AI API costs.
- Get a terminal-based AI pair programmer without paying premium rates.
- Maintain project conventions across sessions with self-verifying memory.
- Access an AI coding agent in regions where other services are restricted.