A new terminal-native coding agent called cwcode is positioning itself as a lean, local-first alternative to bloated AI IDE extensions. Built in Go (~12k lines), it drives any OpenAI-compatible endpoint—including DeepSeek V4 Pro, Qwen3.6-27B, Kimi, and Azure—as a tool-using coding assistant through a Bubbletea TUI interface.
Cost That'll Make You Actually Use It
Running cwcode costs roughly $0.40 per hour on DeepSeek V4 Pro. That's 5% of what you'd burn through Claude for the same workload. The team ships with pre-configured profiles for DeepSeek (Pro and Flash), Azure OpenAI, Kimi for Coding, and a local vLLM/llama.cpp setup for Qwen3.6-27B running on your home server. Switching between them mid-session is a single slash command away.
Hash-Anchored Edits: No More Token Waste
The most interesting technical detail is hash-anchored editing. The read_file tool annotates every line with a 3-character content hash like "42:a3f| return x." When you want to edit, the edit_lines tool takes (line, hash, new_text) and rejects the entire batch if any hash drifted since the last read. This means the model never has to reproduce content character-perfect just to land an edit—it came from Can Akay's February 2026 post, ported to Go in about 200 lines. The team reports output tokens per session dropped 30–40% on V4 Pro as a result.
Sticky Prefix Cache Hits 85%+ After Turn Three
The system prompt is byte-stable across turns because tool definitions serialize deterministically and reasoning content gets stripped from outbound requests by default on every provider. DeepSeek's prompt-cache hit path runs ~120× cheaper than the miss path, and their /cache slash command shows session-cumulative hit ratios routinely exceeding 85% after the third turn.
Checkpoints, Rewinds, and Storm-Breaker Error Handling
Before any file-mutating tool executes, the harness snapshots pre-state of every path that tool will touch. Snapshots are SHA-256-keyed blobs stored in ~/.cwcode/sessions/
Plan Mode vs Code Mode Toggle
A single Shift+Tab toggles between read-only planning (the LLM only sees non-mutating tools) and full execution. The model doesn't see the flag—it just operates with a different (smaller) tool registry and an altered system prompt. The human holds final control unless you explicitly opt into YOLO mode for autonomous operation.
Autonomous Goal Loop for Overnight Runs
/goal
No SaaS Lock-In
Config lives in JSON files. Sessions are JSON. Checkpoints are content-addressed blobs. Memory store is a SQLite file. Everything sits under ~/.cwcode/. There's no account, no telemetry, no remote control plane. If the network's down and your model endpoint is local, the agent keeps working.
Key Takeaways
- cwcode runs for ~$0.40/hour (5% of Claude's cost) via DeepSeek V4 Pro integration
- Hash-anchored edits cut output tokens 30–40% by eliminating character-perfect reproduction requirements
- Sticky prefix cache achieves 85%+ hit ratios after turn three, dropping costs further on cached turns
- Pre-built binaries available now (v1.11) for macOS arm64/amd64 and Windows amd64; source goes public once API surface stabilizes
The Bottom Line
cwcode isn't trying to be everything—it's a focused terminal tool that respects your wallet and your data. If you're already paying for DeepSeek access, this gets you a persistent coding agent with proper error recovery and session rewind for almost nothing extra.