Holbox.ai
Boxcode brings powerful open models directly into your terminal, giving developers an AI coding experience built for speed, control, and the way software is actually built.
curl -fsSL https://boxcode.sh/install.sh | bashMIT licensed · Written in Rust · macOS, Linux, Windows
One platform. Two layers.
Holbox gives developers the tools to build with AI and the infrastructure to run it.
An AI coding assistant that lives in your terminal. It reads your files and runs your commands — and waits for you before every one.
$ boxcode
› add rate limiting to the public API
✓ plan ready · 4 files · awaiting approval
/plan research first, write nothing
/rollback undo everything this session
Deploy and route open-source models across GPU infrastructure without managing the complexity yourself.
Boxcode
Boxcode brings an intelligent coding agent directly into the environment developers already live in: the terminal.
Boxcode works across your repository instead of treating every file as an isolated prompt. Run /init and it explores the project and writes a BOXCODE.md that rides along with every request.
Give Boxcode a task. Let it inspect, plan, edit and iterate. Every change arrives as a diff you read before it lands — and /rollback undoes everything written this session.
Changes shouldn't just look right. Boxcode can run your tests, read the failure and keep iterating — pausing for your approval before it runs anything.
Command showcase
Boxcode opens an interactive session in the directory you're already in. Describe the work at the prompt, or drive it with slash commands.
Developer experience
/init explores the project and writes a BOXCODE.md that rides along with every request.
Inspect, plan, edit, run, read the failure, iterate — without leaving the session.
Destructive actions wait for your yes. There is no allow-everything key, and dangerous commands are refused outright.
Point it at any OpenAI-compatible endpoint with /provider or BOXCODE_ENDPOINT. Self-hosting included.
Sessions render as ordinary terminal output, so scrollback and search keep working.
Every change lands as a diff you read first, and /rollback undoes everything written this session.
Configuration lives in ~/.boxcode/config.toml — workspace, approval policy, theme.
Open models
Holbox gives developers access to powerful open models without forcing them into a single proprietary AI stack.
Qwen 3.5 4B
Alibaba Qwen
GLM-5.2
Zhipu AI
GLM-4 9B Chat
Zhipu AI
Qwen 3.6 35B-A3B
Alibaba Qwen
Catalogue, GPU tiers and pricing as published by GPU Router.
Boxcode is not tied to a single provider. Any OpenAI-compatible endpoint works — a hosted provider, a model you deployed on GPU Router, or something running on your own machine.
# Boxcode speaks to any OpenAI-compatible endpoint.
# Point it at a model you deployed with GPU Router:
export BOXCODE_ENDPOINT="https://api.gpurouter.dev/v1"
export BOXCODE_API_KEY="grk_live_..."
export BOXCODE_MODEL="glm-4-9b-chat"
boxcodeOr switch interactively with /provider and /model. Provider keys such as DEEPSEEK_API_KEY are detected automatically.
GPU Router
GPU Router handles the complexity of deploying and routing open models across GPU infrastructure, so teams can focus on the application.
Choose a model, a GPU tier and a region. Set a time-to-live so the deployment cleans itself up. You get a URL and an API key back.
gpurouter deploy glm-4-9b-chat \
--tier medium \
--region us-east-1 \
--ttl 2hConnect an AWS account with CloudFormation, choose a model, deploy. One click, with GPU matching handled for you.
Automatic GPU matching sends each model to compute that fits it. Pick the tier and region at deploy time.
You get back a URL and an API key. Existing OpenAI client code points at it without changing a line.
Auto-terminate windows on every deployment, so idle GPUs stop billing instead of running all weekend.
Endpoints are protected with an API key and CIDR range restrictions rather than exposed to the open internet.
Built-in load testing measures time-to-first-token and throughput on your own deployment.
The Holbox platform
Developer layer
Boxcode is the session developers work in — repository context, plans, diffs, and an approval gate before anything runs.
Infrastructure layer
GPU Router is the endpoint underneath — model deployment, GPU matching, private access and cost guards.
The two connect through one boundary: an OpenAI-compatible endpoint. Point Boxcode at a model you deployed on GPU Router and the whole stack is yours.
Technical positioning
Open-weight models from the GPU Router catalogue, or any OpenAI-compatible endpoint you bring.
One base URL and one key. Existing OpenAI client libraries work unchanged.
Deployments land on GPU tiers matched to the model, in the region you choose.
Private endpoints, TTL guards and built-in load testing before anything serves traffic.
Boxcode is a native binary that renders as ordinary terminal output.
Requests are placed on the compute a model needs, without hand-managing pools.
Boxcode is MIT-licensed and developed in public. Read the source, file an issue, build it yourself, or run it against your own endpoint — nothing about it is a black box.
Get started
Install the binary and open a session. Or deploy a model and call it with the OpenAI client you already use.
# macOS and Linux
curl -fsSL https://boxcode.sh/install.sh | bash
# Windows (PowerShell)
irm https://boxcode.sh/install.ps1 | iex
# Then, in any project directory
boxcodefrom openai import OpenAI
client = OpenAI(
base_url="https://api.gpurouter.dev/v1",
api_key="grk_live_...",
)
resp = client.chat.completions.create(
model="glm-4-9b-chat",
messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)Build with open intelligence. Run it on infrastructure designed for the AI era.