OpenCrabs is a self‑hosted AI agent written in Rust that can learn, adapt, and fix its own problems. The latest release, v0.3.17, adds video vision, smarter error handling, and a new skill system that lets it run on any channel—Slack, Discord, Telegram, or a terminal. In this article we’ll walk through what makes the OpenCrabs self‑healing agent so powerful, how it works, and why it’s a game‑changer for developers and teams that need reliable, on‑prem AI.

What Is OpenCrabs?

OpenCrabs is a single binary that runs in your terminal or chat app. It uses the Ratatui library to give you a clean, text‑based interface. The agent can talk to any LLM provider (OpenAI, Claude, Gemini, etc.) and can be extended with custom skills written in YAML. The core idea is that the agent can self‑heal: if a provider fails or a tool crashes, it automatically retries, falls back, or even learns a new tool on the fly.

History and Vision

The project started in early 2025 as a hobby to build a lightweight, privacy‑first chatbot. Over time it grew into a full‑featured agent that can run on a Raspberry Pi or a cloud VM. The vision is to give teams a single, self‑contained AI that never needs a cloud connection, never stores data in the cloud, and can fix itself when something goes wrong.

New Features in v0.3.17

The latest release brings several game‑changing additions. Below we break them down.

Video Vision (Phase 1)

OpenCrabs can now analyze short video clips (up to 18 MB) using Gemini’s vision model. You can attach a video with the <<VID:path>> marker, and the agent will return a text summary or answer questions about the content. This opens up new use cases like automated video captioning or quick visual debugging.

Partial JSON Repair

When the agent receives malformed JSON, it now auto‑fixes common errors: missing commas, unbalanced brackets, or trailing commas. The tool is wired into five provider drop sites, so you never see a broken response again.

TCP Keepalive (15 s)

The agent now checks for silent TCP drops every 15 seconds instead of the default 300 seconds. If a connection drops, it reconnects automatically, keeping your workflow smooth.

Health‑Aware Sticky Fallbacks

OpenCrabs reads a provider_health.json file that tracks which providers are healthy. If the primary provider fails, it automatically switches to a backup without user intervention. This makes the agent more resilient in flaky network conditions.

RetryAttempt Progress Events

The terminal UI shows “⏳ Retry 2/3 — stream dropped” during recovery. You can see exactly how many attempts remain, so you know when the agent is giving up.

Phantom Detection Expanded

The agent now catches “Now ” file‑ops and build/deploy completion claims that were previously missed. This helps prevent false positives in CI pipelines.

RSI Escalation

Instead of deduplicating repeated violations, the agent now bumps a counter. This gives you a clearer picture of how often a problem occurs.

OpenRouter Response Caching

If you use OpenRouter, identical requests are cached for free. This saves tokens and speeds up repeated queries.

Built‑in Skills

The release ships with four built‑in skills: opencli, browser-cdp, a2a-gateway, and dynamic-tools. These cover common tasks like running shell commands, browsing the web, and converting data between formats.

Thinking Content Persisted

The agent captures intermediate text and final responses, storing them in a local database. This lets you review the agent’s reasoning later.

Approval Policy at Runtime

You can change the approval policy for each tool request on the fly by editing a config file. The policy is read at runtime, not at startup, so you can tweak it without restarting.

How It Works Under the Hood

OpenCrabs is built on a few core concepts that make it self‑healing.

Ratatui TUI

The terminal UI is powered by Ratatui, a Rust library that gives you a responsive, scrollable interface. It shows the conversation, tool calls, and progress events in real time.

Provider Health

The agent keeps a health score for each LLM provider. If a provider’s score drops below a threshold, the agent stops sending requests to it and switches to a backup. The health score is updated after every request.

Skill System

Skills are defined in SKILL.md files with YAML frontmatter. The agent loads all skills at startup and lets you add new ones by dropping a file into the skills/ folder. Skills can call external APIs, run shell commands, or even spawn other agents.

OpenRouter Integration

OpenCrabs can route requests through OpenRouter, which aggregates multiple LLM providers behind a single API. The new caching feature means you only pay for unique prompts.

Self‑Healing Loop

When a tool fails, the agent logs the error, updates the provider health, and retries. If the retry fails again, it escalates to the next provider or falls back to a local model via Ollama. This loop runs continuously, so the agent rarely stops working.

Use Cases and Benefits

OpenCrabs is useful for a wide range of teams. Here are some real‑world scenarios.

DevOps Automation

Article supporting image

The opencli skill lets you run shell commands directly from chat. Combine this with the new video vision to debug CI logs or view build videos. The agent can also run cargo audit or npm audit automatically and report vulnerabilities.

Security Audits

The /security-audit skill scans code for CVE patterns, injection risks, and crypto issues. It scores the code from 0‑100 and can be integrated into a CI pipeline to block insecure merges.

Code Review

The agent can read a pull request, run static analysis, and suggest changes. With the new JSON repair, it can handle malformed diff files without crashing.

Multi‑Channel Support

OpenCrabs can be added to Slack, Discord, Telegram, or a terminal. The same agent works across all channels, so your team can chat with AI wherever they prefer.

AI Agent Automation

Because the agent can learn new skills on the fly, you can build a chain of agents that hand off tasks. For example, a “data‑cleaning” agent can hand off a cleaned dataset to a “model‑training” agent.

Comparison to Other Agents

Feature OpenCrabs OpenAI CLI Claude Code
Self‑healing ✔ ❌ ❌
Video vision ✔ (Phase 1) ❌ ❌
Built‑in skills ✔ ❌ ❌
Local deployment ✔ ❌ ❌
OpenRouter caching ✔ ❌ ❌

OpenCrabs stands out because it is fully self‑hosted, can fix itself, and runs on any channel. Other agents either require a cloud connection or lack the ability to recover from failures automatically.

Getting Started

Installation

# On Linux or macOS
curl -L https://github.com/adolfousier/opencrabs/releases/download/v0.3.17/opencrabs-linux-amd64 -o opencrabs
chmod +x opencrabs
sudo mv opencrabs /usr/local/bin/

Configuration

Create a config.yaml file in your home directory:

providers:
  - name: openai
    api_key: YOUR_OPENAI_KEY
  - name: claude
    api_key: YOUR_CLAUDE_KEY
skills_path: ./skills

Custom Skills

Add a new skill by creating skills/my_skill/SKILL.md:

name: my_skill
description: A custom skill
commands:
  - name: echo
    description: Echoes back the input
    script: echo "{{input}}"

Running on Windows

Download the Windows binary from the releases page and run it from PowerShell. The UI works the same as on Linux.

Future Roadmap

OpenCrabs is still young, but the roadmap is clear:

  1. Full Video Analysis – Expand to longer videos and real‑time streaming.
  2. AI‑Generated Skills – Let the agent write new skills from prompts.
  3. Graph‑Based Orchestration – Connect multiple agents in a workflow.
  4. Community Plugins – Open a marketplace for user‑created skills.

The team is actively listening to user feedback. If you have ideas, join the discussion on the GitHub repo or the community forum.

Conclusion

OpenCrabs self‑healing agent is a powerful, privacy‑first AI that can learn, adapt, and fix itself. With new video vision, smarter error handling, and a robust skill system, it’s ready for real‑world use in DevOps, security, and beyond. If you’re looking for a reliable, on‑prem AI that never stops working, OpenCrabs is worth a try.