# openwarp

Community fork of [Warp](https://www.warp.dev) that puts the AI layer behind a Bring-Your-Own-Provider (BYOP) interface — paste a Base URL and API key in settings, pick the wire protocol, and Warp now talks to that endpoint instead of Warp Inc's hosted models. Everything else stays Warp: blocks, AI commands, workflows, keymaps, themes. Upstream changes get continuously merged.

The marketing line is *"Bring any AI model into your terminal"*; the technical line is *"the AI layer is opened up, the rest of Warp isn't reimplemented."* Tracked here as the GPU-terminal-with-baked-in-AI category's first credible community fork.

## Why this exists

Warp ships a great terminal but locks you into its hosted AI subscription, with a Cloud Agent / Computer Use surface that uploads context. OpenWarp keeps the upstream codebase verbatim and patches the AI client so:

1. Custom providers are configurable from Settings.
2. Cloud Agent / Computer Use is **disabled by default**.
3. Credentials live only in `~/.config/openwarp.toml` — no relay, requests go straight to the configured Base URL.
4. UI strings get a real i18n layer (Chinese + English shipping, with Japanese / Spanish previewed).

It's the same pattern as VSCodium-vs-VS Code or Cursor-vs-Continue: take an opinionated commercial product, strip the lock-in surface, keep the upstream merges flowing.

## How the provider layer works

Built on the [genai](https://crates.io/crates/genai) Rust crate as a multi-protocol client. **6 native API protocols, not just an OpenAI-compat shim**:

- OpenAI
- OpenAI Responses
- Anthropic
- Gemini
- Ollama
- DeepSeek

Why the distinction matters: an OpenAI-compat shim flattens reasoning models down to plain chat completions and drops provider-specific fields. OpenWarp's *native* protocols pass through:

- DeepSeek `reasoning_content`
- Claude `thinking`
- Gemini reasoning blocks

…across multi-turn conversations, so the agent loop sees the full chain-of-thought a reasoning model produces. OpenAI-compatible endpoints (Qwen, Groq, Together, OpenRouter, SiliconFlow, LM Studio, etc.) plug in by selecting the OpenAI protocol and pointing Base URL at them.

Provider config is explicit — protocol is named, no model-name guessing:

```toml
# ~/.config/openwarp.toml
name        = "DeepSeek (Personal)"
api_protocol = "OpenAI"          # or "Anthropic", "Gemini", "Ollama", "DeepSeek", ...
base_url    = "https://api.deepseek.com"
api_key     = "sk-……"
model       = "deepseek-reasoner"
```

## minijinja system prompts

System prompts are templates rendered per call from the current cwd, locale, and role. The engine is [minijinja](https://github.com/mitsuhiko/minijinja) (Mitsuhiko's Rust port of Jinja2):

```jinja
{% if user.role %}
You are an expert {{ user.role }}.
{% endif %}
Reply in {{ locale }}.
```

Same context-engineering instinct as a Claude Code skill or a Cursor rules file, but baked into the terminal so every shell session gets a context-aware system prompt without project-level config.

## Install

No public release yet — has to be built from source:

```
git clone -b openWarp https://github.com/zerx-lab/warp
cd warp
cargo build --release
./target/release/openwarp
# then add a custom provider in Settings
```

The fork lives on the `openWarp` branch (capital W). Upstream `main` is reserved for sync from `warpdotdev/warp`.

## Status

3 days old (created 2026-04-29) and already at **422 stars / 23 forks**. The site banner says *"Early development — no public release yet"* and the README points at `cargo build --release` as the only install path. Repo size is ~155 MB because it carries the full upstream Warp source tree.

License is AGPL-3.0 in the GitHub field; site says AGPL-3.0 / MIT dual, mirroring upstream Warp.

Distinct from the unrelated [`runs-on/warp`](https://github.com/runs-on/warp) (CI/Docker tool) — same name collision the GitHub redirect briefly serves the wrong page for.

## Limitations to think about before adopting

- **No release binaries.** Cargo build of a Tauri-shaped Rust desktop app on macOS is a 10-minute affair plus toolchain setup.
- **Upstream merge cadence is unproven.** Warp ships fast; a fork that promises *"continuously merging upstream"* needs to actually keep up, otherwise the value proposition collapses.
- **Single-vendor fork.** `zerx-lab` is a 2024-vintage GitHub user with no other notable projects — the bus factor is exactly one.
- **AGPL-3.0 means** any modifications you serve to others (e.g., behind a proxy) must be open-sourced.
- **Cloud Agent / Computer Use are disabled, not removed.** Upstream merges will keep dragging that surface back in.
- **Branch naming (`openWarp` vs `OpenWarp` vs `openwarp`)** is inconsistent across the site, README, and repo — small but annoying friction.

Filed on [[toolbox/watchlist]] — re-check 2026-08-02. The interesting question is whether a binary release lands and whether anyone other than the original author starts merging upstream.

## Related

- [[gpu-accelerated-terminal]] — Warp is the canonical example; battery cost discussion in [[terminal-gpu-battery-cost]] applies.
- [[gomodel]] — Go AI gateway with the same OpenAI-compatible-unified-API instinct, different layer of the stack (process-level proxy vs. client-side library).
- [[mcp-vs-skills]] — minijinja templates here are the Skills-style "load knowledge into context" model, baked into the terminal config.
- [[deepseek]] — reasoning_content passthrough is one of the most concrete OpenWarp benefits over a plain chat-completions client.
- BYOP (Bring Your Own Provider) — same pattern as [[ctx]], [[pi-coding-agent]], [[claude-defuddle]]; OpenWarp generalises it to the whole terminal.

## Repo

[github.com/zerx-lab/warp](https://github.com/zerx-lab/warp) · 422★ · AGPL-3.0 / MIT · Rust · `openWarp` branch

Sources: [[openwarp-landing]], [GitHub repo](https://github.com/zerx-lab/warp).