Native Coding Agent Optimized for Local LLM and DeepSeek v4 with Vector Memory

cwcode

A terminal coding agent built around DeepSeek V4 Pro, Qwen3.6‑27B, Kimi, Azure, and anything else that speaks OpenAI’s chat API.

Written in Go. Lives in your terminal. Edits real code. Recovers from its own mistakes. Costs about $0.40 to leave running for an hour.

5%

of Claude’s token cost on DeepSeek V4 Pro

85%+

prefix-cache hit ratio after turn 3

~12k

lines of Go no external services

What it is

cwcode is a Bubbletea TUI that drives any OpenAI-compatible chat endpoint as a tool-using coding agent. It ships with profiles for DeepSeek (Pro and Flash), Azure OpenAI, Kimi for Coding, and a local vLLM / llama.cpp profile for Qwen3.6-27B on a home server. Switching profiles mid-session is one slash command.

It has bash, file edit, glob, grep, web fetch, headless-Chrome fetch (driven via CDP through your real browser), sub-agents, a persistent semantic-memory store, content-addressed checkpoints with rewind, a plan/code mode toggle, and an autonomous goal loop. The tool registry is six hundred lines and adding a new tool is a two-method Go interface.

It is not a SaaS. There is no account, no telemetry, no remote control plane. Your API key sits in ~/.cwcode/config.json. Your session history sits in ~/.cwcode/sessions/. If your network is down and the model endpoint is local, the agent keeps working.

Why it’s different

Hash-anchored edits

The read_file tool annotates every line with a 3-character content hash: 42:a3f| return x. The edit_lines tool takes (line, hash, new_text) and rejects the entire batch if any hash drifted. The model never has to reproduce content character-perfect to land an edit. Adopted from Can Akay’s February 2026 post and ported to Go in about 200 lines. Output tokens per session dropped 30–40% on V4 Pro.

Sticky prefix cache

The system prompt is byte-stable across turns. Tool definitions serialize in a deterministic order. Reasoning content is stripped from outbound requests on every provider by default. DeepSeek’s prompt-cache hit path is ~120× cheaper than the miss path, and our /cache slash command shows session-cumulative hit ratio that routinely exceeds 85% after the third turn.

Plan vs code mode

A single Shift+Tab toggle between read-only planning (the LLM only sees non-mutating tools) and full execution. The model doesn’t see the flag — it just sees a different (smaller) tool registry and a system-prompt addendum. The human holds final control unless you opt into YOLO mode.

Checkpoint & rewind

Before any file-mutating tool runs, the harness snapshots the pre-state of every path the tool declares it will touch. Snapshots are SHA-256-keyed blobs in ~/.cwcode/sessions//objects/, deduped automatically. /rewind N restores files, truncates conversation history, and pre-fills the input box with the original prompt.

Storm-breaker

When the same tool fails identically three times in a row, the harness doesn’t silently abort. It synthesizes a plain-language response (“I’m unable to continue: read_file failed three times because the path was empty. Please clarify…”), streams it like a normal reply, and appends it to history so follow-ups have context.

Autonomous goal loop

/goal appends a goal to goals.md. /goal on starts an autonomous loop that runs back-to-back turns until every checkbox is marked done or until a safety cap of 20 consecutive cycles. We use this for four-hour overnight runs on annotated tasks.

No SaaS lock-in

Config is JSON. Sessions are JSON. Checkpoints are content-addressed blobs. Memory store is a SQLite file. Everything lives under ~/.cwcode/. If the project disappeared tomorrow your sessions are still readable.

What it looks like

Captured during real work on our dose-prediction codebase: the agent proposing an edit_file change to a Go test, with a unified diff highlighted inline, the reasoning trace streaming below, and the current task list pinned to the bottom of the pane.

cwcode running a Go test edit; multi-tab tmux session, dose-prediction project, DeepSeek profile.

Install

Download a pre-built binary for your platform from the Google Drive release folder (current build: v1.11; macOS arm64 / amd64 and Windows amd64). Drop it somewhere on your PATH and make it executable:

curl -L -o ~/.local/bin/cwcode chmod +x ~/.local/bin/cwcode cwcode -version

You’ll need an OpenAI-compatible endpoint (DeepSeek API key, Azure deployment, local vLLM, or whatever else you have on hand).

Configure a profile in ~/.cwcode/config.json:

{ "active_profile": "deepseek-pro", "profiles": { "deepseek-pro": { "provider": "deepseek", "endpoint": "https://api.deepseek.com", "model": "deepseek-v4-pro", "api_key": "sk-...", "ctx_size": 262144 } } }

Run it.

cwcode # Bubbletea TUI cwcode -p "fix the bug" # one-shot, no session cwcode -continue # resume the most recent session cwcode -plain # stdout REPL (no TUI)

Built-in tools

namepurposeneeds approval

bashrun a shell command (streaming output)yes

bash_backgroundspawn a long-running processyes

read_fileread with per-line content hashesno

write_filecreate or overwrite a fileyes

edit_fileexact-string replace with whitespace recoveryyes

edit_filesatomic multi-file batch (exact-string)yes

edit_lineshash-anchored line replacementyes

globfind files by patternno

grepsearch files for a regexno

lslist directory contentsno

web_fetchfetch a URL and clean it upno

chrome_fetchdrive your real Chrome via CDP for bot-blocked pagesno

taskspawn a sub-agent with its own contextyes

rememberadd a fact to the persistent memory storeno

recallsemantic search over past sessionsno

todo_writeupdate the visible task listno

FAQ

Why Go?

Single static binary, fast startup, easy cross-compile. Three platform builds in 90 seconds. The TUI binary on macOS is 24 MB with debug symbols stripped.

Why a terminal app and not a VS Code extension?

Because we wanted the agent to be the primary interface, not a side panel. The TUI gives the model the whole pane to work in and gives us a small surface to debug. If you live in VS Code, you can run cwcode in the integrated terminal.

Does it work with Claude?

Not directly — cwcode speaks the OpenAI /v1/chat/completions shape. Claude has its own API. You can put Claude behind a translating proxy if you want, but we built this for the cost shift in the other direction.

What model do you use day to day?

DeepSeek V4 Pro for most coding work, Flash for quick questions and one-shot scripts, the local Qwen3.6‑27B profile when we want zero latency or are working offline.

Is the source available?

Pre-built binaries are on Google Drive. Source is currently private; we plan to open it once the API surface settles. If you want a peek before then, get in touch.

Who built this?

A small team that uses it daily for dose-prediction model training, financial research agents, and writing cwcode itself. The agent ships its own bugs and writes its own fixes.