ovis
Ovis (formerly herdr web UI): free, open-source browser and phone client for herdr. Chat and live terminal for Claude Code, Codex and every agent pane, remote PCs over SSH, web push alerts
- 2.2k
- TypeScript
- MIT
- 17 min ago
Local-only AI coding agent powered by Gemma 4 via candle. Runs entirely offline — packaged as a ZIP for air-gapped environments.
Repo devswha/zipcode, read from GitHub 1 min ago. devswha has not joined Take My Repo.
devswha/zipcode is a public repo on GitHub by devswha. Local-only AI coding agent powered by Gemma 4 via candle. Runs entirely offline — packaged as a ZIP for air-gapped environments. It declares no license on GitHub. Public does not mean open source: without a license, nobody is granted the right to reuse it. It is written mainly in Rust, has 5 stars and last changed on 2026-05-17. Nobody has reviewed it on Take My Repo yet, which says nothing about its quality. Read from GitHub on 2026-11-20.
In the maintainer's words. Shown from GitHub; Take My Repo does not keep a copy.
Local-only AI coding agent. No API keys. No cloud dependency. Just code.
Quick Start • Air-Gapped Deploy • Tools • Architecture • Config
zipcode is a Rust-based coding agent that runs entirely on your machine using local LLM inference. It can run Gemma-family GGUF models through candle, native llama-cpp, or a llama-server fallback for newer Gemma 4 models that outpace the current Rust bindings. Responses stream token-by-token via SSE for instant feedback.
The intended entrypoint is plain zipcode: if the machine is ready, you land in the REPL; if it is fresh or broken, zipcode points you at setup or repair steps instead of dropping you into backend jargon.
Ship it on a USB stick. Run it in a SCIF. It just works.
$ zipcode
zipcode 0.1.0 (abc1234+dirty, build 1770000000) — local AI coding agent
Type /help for commands, Ctrl+D to exit
> read main.rs and add error handling to the database connection
Reading src/main.rs...
> edit_file {"file_path": "src/main.rs", ...}
< Successfully edited src/main.rs
I've wrapped the database connection in a proper error handler with
retry logic and connection pooling...
| Problem | zipcode Solution |
|---|---|
| API keys expire, leak, or get rate-limited | No API. Model runs locally. |
| Corporate networks block LLM endpoints | No network needed after setup unless you explicitly ask zipcode to fetch a GitHub repo URL. |
| Sensitive code can't leave the machine | Everything stays on disk. |
| Cloud LLMs add latency | GPU inference on your hardware. |
Setup requires npm, pip, Docker... |
Single static binary. |
| Setup | Response Time | Tokens/sec |
|---|---|---|
| CPU only (12 threads) | ~150s | ~0.8 |
| GPU (RTX 2070 SUPER) | ~7.6s | ~15 |
SSE streaming delivers tokens to your terminal as they're generated — no waiting for the full response.
git clone https://github.com/devswha/zipcode.git
cd zipcode
./install.shThe clone installer builds zipcode, installs a zipcode launcher into ~/.local/bin,
copies any model artifacts you point it at, and when nothing is ready it walks you
through a setup interview: download in the terminal now, show the download links,
paste the local paths after you fetch the files, or skip for now.
Common flags:
./install.sh --model /path/to/model.gguf --tokenizer /path/to/tokenizer.json ./install.sh --llama-server /path/to/llama-server ./install.sh --skip-build --binary ./target/release/zipcode ./scripts/download_model.sh --yes --dir ~/.zipcode/models ./scripts/download_model.sh --yes --preset 31b --dir ~/.zipcode/models ./scripts/build_llama_server.sh ~/.zipcode
The online download path may require Hugging Face login / access approval for the Gemma GGUF repository.
# Automated (requires internet + hf or huggingface-cli) ./scripts/download_model.sh # Or manual: download a Gemma GGUF + matching tokenizer.json # Place both files in ~/.zipcode/models/
# Start here. Ready machines drop into the REPL. zipcode # Scripted first-run / repair path zipcode setup --skip-smoke # Detailed readiness + repair report zipcode doctor
zipcode is the default entrypoint. Use setup when you want zipcode to discover the model,
write or refresh the global config, and optionally run a smoke prompt. Use doctor when you
want the repair summary without entering the agent.
Every local build embeds a build label: package version, Git commit, dirty marker for tracked local changes, and a Unix build timestamp. Use it to confirm which binary you are testing:
zipcode --version zipcode doctor
The fullscreen TUI header shows the Git label as well, so stale binaries are easier to spot during manual QA.
# Interactive entrypoint zipcode # One-shot zipcode prompt "explain this codebase" # Guided config + smoke test zipcode setup # Guided config only zipcode setup --skip-smoke # Readiness / repair report zipcode doctor # Custom model zipcode --model ./my-model.gguf # Gemma 4 via the llama-server helper ZIPCODE_LLAMA_SERVER_BIN=/path/to/llama-server \ zipcode --backend llama-server --model ~/.zipcode/models/gemma-4-e2b-it-Q8_0.gguf prompt "hello" # GPU acceleration (CUDA) ZIPCODE_GPU_LAYERS=99 ZIPCODE_FLASH_ATTENTION=1 zipcode # Build llama-server with CUDA (auto-detects GPU) ./scripts/build_llama_server.sh # If CUDA 11.x + newer GCC fails, prefer gcc-10/g++-10 for GPU builds CC=/usr/bin/gcc-10 CXX=/usr/bin/g++-10 \ ZIPCODE_LLAMA_SERVER_CMAKE_ARGS="-DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-10" \ ./scripts/build_llama_server.sh # Read-only mode (safe exploration) zipcode --permission-mode read-only
zipcode was designed from the ground up for isolated networks. Package everything into a ZIP,
move it on a USB drive, and let bare zipcode handle first-run versus recovery on the target
machine.
# Smallest bundle ./scripts/package.sh # Or bundle llama-server so Gemma 4 stays fully offline on first run / repair ./scripts/package.sh --with-llama-server=/path/to/llama-server # Creates: dist/zipcode-v0.1.0-linux-x86_64-cuda.zip
zipcode-v0.1.0-linux-x86_64-cuda.zip
|- zipcode # single binary (~30 MB)
|- install.sh # extracted-bundle installer
|- install_llama_server.sh # repair helper for adding llama-server later
|- build_llama_server.sh # build llama-server from source when needed
|- download_model.sh # model download helper
|- README.md
|- llama-server # optional, when bundled during packaging
|- scripts/lib/install_common.sh
'- models/
|- PLACE_MODEL_HERE.txt
'- PLACE_TOKENIZER_HERE.txt
Internet Machine Air-Gapped Machine
================ ==================
1. Build or download zipcode.zip
2. Download model.gguf + tokenizer.json
3. Optional: bundle llama-server for Gemma 4
4. Copy to USB
---- USB ---->
5. Unzip
6. ./install.sh
7. cp model.gguf tokenizer.json ~/.zipcode/models/
8. zipcode
9. If repair is needed:
zipcode doctor
zipcode setup --skip-smoke
zipcode comes with 11 built-in tools that the model can invoke autonomously:
| Tool | Description | Permission |
|---|---|---|
| Bash | Execute shell commands via bash -c |
Needs approval in workspace-write |
| ReadFile | Read file contents with line numbers, offset/limit | Always allowed |
| WriteFile | Create or overwrite files, auto-creates parent dirs | Blocked in read-only |
| EditFile | Targeted string replacement in existing files | Blocked in read-only |
| FetchRepo | Clone a public GitHub repo into .zipcode-remote/ for analysis |
Blocked in read-only |
| GlobSearch | Find files by glob pattern (**/*.rs, src/*.py) |
Always allowed |
| GrepSearch | Search file contents with regex, returns file:line: |
Always allowed |
| TodoWrite | Structured todo list persistence (.zipcode-todos.json) |
Blocked in read-only |
| REPL | Execute Python/Node.js code snippets | Blocked in read-only |
| Agent | Spawn sub-agent for delegated tasks | Stub in v0.1 |
| ToolSearch | Search available tools by keyword | Always allowed |
All tool output is automatically truncated at 8 KB with a byte-count summary.
FetchRepo is the explicit network exception: it only handles repository-root https://github.com/<owner>/<repo> URLs, clones them with git clone --depth 1 into .zipcode-remote/, and is unavailable in read-only mode. Air-gapped deployments should pre-copy repositories into the workspace instead.
| Mode | Bash | Writes | Reads | Use case |
|---|---|---|---|---|
read-only |
Denied | Denied | Allowed | Safe exploration |
workspace-write |
Approval required | Allowed | Allowed | Default |
full-access |
Allowed | Allowed | Allowed | Trusted automation |
In
workspace-write, approval-gated tools fail closed when zipcode is running non-interactively (for examplepromptmode with EOF / no TTY). A real interactive terminal is required to approvebashorrepl.
+-------------------------------------------------+
| zipcode CLI |
| REPL / one-shot / doctor |
+-------------------------------------------------+
| Agentic Runtime |
| +--------+ +----------+ +---------+ |
| |Session | |Permission| | Config | |
| |Manager | | Policy | | Loader | |
| +--------+ +----------+ +---------+ |
| +-----------------+ |
| | Conversation |<--- tool results |
| | Loop | |
| +-------+---------+ |
| | tool calls |
| +-------v---------+ |
| | Tool Router | |
| +--+---------+----+ |
| | | |
| +-----+ +----+----+ |
| |Bash | |FileOps | ...11 tools |
| +-----+ +---------+ |
+-------------------------------------------------+
| Inference Engine |
| candle | llama-cpp | llama-server (SSE) |
| GPU offload | KV cache | Flash attention |
+-------------------------------------------------+
| Model Store |
| ~/.zipcode/models/*.gguf |
+-------------------------------------------------+
zipcode/
|- crates/
| |- inference/ Candle GGUF engine, Gemma 4 chat template, sampler
| |- tools/ 11 tool implementations + Tool trait + registry
| |- runtime/ Agentic loop, config, permissions, sessions
| '- cli/ REPL, one-shot, doctor, slash commands
|- scripts/ install.sh, download_model.sh, package.sh
|- models/ .gguf model files (gitignored)
'- docs/ Design specs + implementation plans
cli --> runtime --> inference
|
'--> tools
| Crate | Responsibility |
|---|---|
| inference | GGUF loading, tokenization, KV cache, SSE streaming, GPU offload. Supports candle, llama-cpp, and llama-server backends. |
| tools | 11 tool implementations. Tool trait, ToolRegistry, ToolResult truncation. |
| runtime | Agentic conversation loop. Config hierarchy, permission policy, session persistence. |
| cli | Binary entry point. REPL (rustyline), ANSI rendering (termimad), clap argument parsing. |
| Command | Description |
|---|---|
zipcode |
Smart entrypoint: REPL when ready, setup/repair guidance when not |
zipcode prompt "..." |
One-shot prompt, then exit |
zipcode setup |
Discover prerequisites, write config, optionally run a smoke prompt |
zipcode doctor |
Show readiness state and repair hints |
| Flag | Default | Description |
|---|---|---|
--model <PATH> |
~/.zipcode/models/ |
Path to .gguf file |
--backend <BACKEND> |
llama-cpp |
llama-cpp / llama-server / candle |
--permission-mode |
workspace-write |
read-only / workspace-write / full-access |
--ui <MODE> |
fullscreen |
Interactive UI mode: plain or fullscreen |
| Command | Action |
|---|---|
/help |
Show commands |
/status |
Session ID, message count, cwd |
/clear |
Reset conversation |
/quit |
Exit |
zipcode now launches the fullscreen UI by default when the install is ready. Use
zipcode --ui fullscreen (or zipcode repl --ui fullscreen) explicitly if you want to force it,
or zipcode --ui plain to return to the classic line-oriented REPL.
The fullscreen terminal UI MVP includes:
/help, /status, /clear, /quit)Key bindings:
Enter — submitCtrl+J — newlineArrow keys — move cursor / recall input historyPgUp / PgDn — scroll transcript inside the TUIF1 — help overlayEsc — clear input or close overlayCtrl+D — exit.zipcode.json)Loaded from the nearest ancestor project root. Overrides global ~/.zipcode/config.json.
Paths such as model_dir, model_file, and llama_server_bin support ~/... home expansion.
{
"permission_mode": "workspace-write",
"model_dir": "~/.zipcode/models",
"gpu_layers": 99,
"flash_attention": true,
"generation": {
"temperature": 0.7,
"top_p": 0.9,
"max_tokens": 4096
}
}.zipcode.md)Place in any project root. Contents are injected into the system prompt.
# My Project - Language: Rust 2021 - Build: cargo build --release - Test: cargo test --workspace - Never modify files under vendor/
~/.zipcode/
|- config.json # global config
|- models/ # GGUF model files
'- sessions/ # conversation history
'- <uuid>.json
| Variable | Description | Default |
|---|---|---|
ZIPCODE_LLAMA_SERVER_BIN |
Path to llama-server binary | auto-detect |
ZIPCODE_LLAMA_SERVER_CMAKE_ARGS |
Extra cmake flags for build_llama_server.sh |
none |
ZIPCODE_LLAMA_SERVER_CUDA_ARCHITECTURES |
Override detected CUDA arch list (example: 75) |
auto-detect when possible |
ZIPCODE_LLAMA_SERVER_CTX |
Context window size | 131072 (Gemma 4 128K native) |
ZIPCODE_GPU_LAYERS |
GPU layers to offload (e.g., 99 for all) | none (CPU) |
ZIPCODE_FLASH_ATTENTION |
Enable flash attention (1 or true) |
false |
| Component | Minimum |
|---|---|
| Model | Gemma 4 E2B IT (GGUF format) |
| VRAM | ~8 GB recommended for Q8_0 quantization |
| CUDA | 12.0+ (optional, CPU fallback available) |
| Disk | ~5 GB for Q8_0 model file |
# Option 1: helper script ./scripts/download_model.sh # Option 2: manual # Download from: huggingface.co/ggml-org/gemma-4-E2B-it-GGUF # Also download tokenizer.json from the matching model repo # Place both files in: ~/.zipcode/models/
CPU inference works but is significantly slower. zipcode auto-detects CUDA availability at startup.
llama-cpp-2 0.1.141 still cannot open general.architecture = gemma4.llama-server../scripts/package.sh --with-llama-server=/path/to/llama-server~/.zipcode/install_llama_server.sh /path/to/llama-server ~/.zipcodeZIPCODE_LLAMA_SERVER_BIN=/path/to/llama-serverAfter that, rerun zipcode (default entrypoint) or zipcode setup if you want zipcode to refresh its saved config first.
| Crate | Purpose |
|---|---|
| candle | Rust-native ML framework for GGUF inference |
| tokenizers | HuggingFace tokenizer |
| clap | CLI argument parsing |
| rustyline | REPL line editing |
| termimad | Markdown to ANSI rendering |
| tokio | Async runtime |
Contributions welcome. Please open an issue first for major changes.
# Development workflow cargo test --workspace # run all tests cargo clippy --workspace # lint cargo fmt --all # format cargo build --release -p zipcode # build binary
Inspired by claw-code by @instructkr.
Built with candle by Hugging Face.
devswha/zipcode is a public repo on GitHub by devswha. Local-only AI coding agent powered by Gemma 4 via candle. Runs entirely offline — packaged as a ZIP for air-gapped environments.
It declares no license on GitHub. Public does not mean open source: without a license, nobody is granted the right to reuse it.
Its maker has not tagged a system, so it is unknown here. Its README says what it needs.
Open "Take it" on its page: Take My Repo writes an install plan for your machine and hands it to your AI agent (Claude Code, Codex or Cursor), which asks you before each step. Take My Repo has not tested it on any machine.
Its last change on GitHub was on 2026-05-17. Take My Repo does not judge from that whether it is maintained.
Nobody has reviewed it on Take My Repo yet, which says nothing about its quality.
Ovis (formerly herdr web UI): free, open-source browser and phone client for herdr. Chat and live terminal for Claude Code, Codex and every agent pane, remote PCs over SSH, web push alerts

Desktop and self-hosted interface for Gajae Code agent sessions.
Community plugin marketplace for Gajae Code (gjc). Workflow skills + slash commands, Claude Code / Codex compatible.

ChatMux — provider-agnostic tmux chat web-terminal for coding agents (GJC/Codex/Claude/Cursor/OpenCode)
Benchmark suite: Claw Code (Rust) vs Claude Code (Node.js)
This page is also available as structured data, with the same sources and dates.
/api/v1/resources/gh/devswha/zipcodeTake My Repo has not tested this on your machine. Requirements come from the project's own documentation.