OpenCrabs is a lightweight, single‑binary tool that lets you run AI agents on your own servers. The latest 0.3.19 update brings several new features that make it easier to connect to OpenAI’s Codex, use external embedding services, and run the agent with minimal memory. In this article we’ll walk through what’s new, how to set it up, and why self‑hosted AI agents are a good fit for developers, researchers, and small businesses that want full control over their data.

What is OpenCrabs?

OpenCrabs is built with the Rust programming language and a terminal‑first user interface. It can run on Linux, macOS, and Windows, and it supports a wide range of AI models, from local Llama‑based models to OpenAI’s GPT‑5 series. The core idea is to give you a single command‑line tool that can:

  • Run agents that can browse the web, read documents, and generate code.
    Swap models on the fly without restarting the process.
  • Store memory in a local database or use an external vector store.

Because it is self‑hosted, you keep all data on your own hardware, which is a big advantage for privacy‑conscious teams.

Key New Features in 0.3.19

Codex OAuth Provider

The biggest addition is a native OAuth flow for OpenAI’s Codex. Instead of having to store an API key in a file, you can now authenticate through a browser once and let OpenCrabs keep the token for you. The flow uses the device‑code PKCE protocol, so it works on headless servers too. The new provider shows up in the /models and /onboard:provider menus, and it automatically refreshes the token in the background.

OpenAI‑Compatible Embedding API

If you prefer not to download a 300 MB GGUF embedding model, the new embedding API lets you point OpenCrabs at any /v1/embeddings endpoint. You can use OpenAI’s text‑embedding‑3‑small, Ollama’s nomic‑embed‑text, Jina, LM Studio, or any custom service. The configuration is simple:

[memory.embedding]
url = "https://api.openai.com/v1/embeddings"
model = "text-embedding-3-small"
api_key = "YOUR_KEY"
dimensions = 1536

OpenCrabs will detect the vector size automatically and use it for similarity search.

FTS5‑Only Memory Mode for VPS

Running a full vector engine can be heavy on memory. The new vector_enabled = false flag in the [memory] section turns off the GGUF engine and the vector search, leaving only a lightweight FTS5 keyword search. This mode is ideal for virtual private servers (VPS) where RAM is limited. OpenCrabs will auto‑detect a VPS environment and add the flag for you.

Other Fixes and Improvements

  • Fixed a bug where the Codex OAuth flow used wrong field names in the response.
  • Corrected the verification URL for Codex OAuth.
  • Made the /models dialog show the correct provider list.
  • Added platform‑specific test variants for Windows CI.
  • Improved error handling for cron jobs that use mismatched provider/model pairs.
  • Added a health check before taking a screenshot in the browser tool.

How to Get Started with the New Release

Installing OpenCrabs 0.3.19

  1. Download the binary for your OS from the OpenCrabs GitHub releases page.
  2. Place the binary in a directory that is in your PATH.
  3. Run opencrabs --version to confirm you have 0.3.19.

Configuring the Codex OAuth Flow

  1. Run opencrabs onboard:provider codex.
  2. A browser window will open with a short code. Enter the code on the page.
  3. Once authenticated, the token is stored in ~/.opencrabs/auth/codex.json.
  4. You can now select Codex models from the /models menu.

Article supporting image

Using the Embedding API

Add the [memory.embedding] section to your opencrabs.toml file. If you want to use OpenAI’s embeddings, copy the snippet from the previous section. After saving, restart OpenCrabs. The agent will now use the external service for vector search.

Switching to FTS5‑Only Mode

If you’re on a VPS and want to save memory:

[memory]
vector_enabled = false

Restart OpenCrabs and the agent will use only keyword search. You can still use the /search command, but it will be faster and lighter.

Real‑World Use Cases

Small Business AI Automation

A local bakery can run an OpenCrabs agent that reads inventory spreadsheets, predicts restock dates, and sends email reminders. Because the agent is self‑hosted, the bakery keeps all customer data on its own servers, avoiding third‑party cloud storage.

Research Labs and Data Privacy

A university research group can use OpenCrabs to run experiments on sensitive data sets. The agent can process PDFs, extract key points, and generate summaries without sending the data to external APIs. The new embedding API lets the group use a private embedding service that complies with institutional policies.

Developer Tooling and Testing

Developers building AI‑powered IDE extensions can use OpenCrabs to test code‑generation agents locally. The new browser tool can open a local web page, run JavaScript, and capture screenshots, making it easier to debug web‑based agents.

Comparison with Other Self‑Hosted AI Platforms

OpenAI Agents SDK vs OpenCrabs

The OpenAI Agents SDK is still experimental, whereas OpenCrabs is a stable, production‑ready binary. OpenCrabs offers a terminal UI, built‑in OAuth, and a lightweight memory mode that the SDK does not provide.

OpenCrabs vs OpenAI’s Codex CLI

OpenCrabs now uses the Codex CLI under the hood for authentication, but it adds a higher‑level agent interface. The CLI is great for raw model calls, but OpenCrabs lets you chain tools, browse the web, and store memory in one place.

Future Roadmap and Community Involvement

Upcoming Features

  • Tool‑loop reasoning markers: The agent will now keep track of its own thoughts across tool calls.
  • Cron job validation: Future releases will add stricter checks for provider/model pairs.
  • Health‑check improvements: More robust checks before taking screenshots.

How to Contribute

OpenCrabs is open source. You can:

  • Submit pull requests on GitHub.
  • Report bugs in the issue tracker.
  • Join the Discord community to discuss ideas.

The project welcomes contributions from developers, researchers, and hobbyists.

Conclusion

OpenCrabs 0.3.19 brings a set of practical improvements that make self‑hosted AI agents easier to set up and run. The new Codex OAuth provider removes the need to manage API keys manually, the external embedding API gives you flexibility, and the FTS5‑only mode saves memory on VPS deployments. Whether you’re a small business, a research lab, or a developer, these features help you keep control of your data while still enjoying powerful AI capabilities.