opencrabs

What OpenCrabs is

OpenCrabs is an autonomous, self-improving multi-channel AI agent built in Rust. The README describes it as a single Rust binary that runs on your terminal with no server, no gateway, and no infrastructure, making direct HTTPS calls to LLM providers from your machine. It is inspired by Open Claw and authored by Adolfo Usier. The project emphasizes a small footprint: the binary is 34 to 36 megabytes with zero dependencies, which the README contrasts with Node.js agent frameworks that pull in more than a gigabyte of transitive packages. On security, OpenCrabs opens no network listeners and only makes outbound HTTPS calls. API keys stay on the user's machine, and the code uses a zeroize pattern to clear them from memory on drop, with redaction in debug output. Data residency is described as fully local: chat sessions, tool executions, memory, embeddings, brain files, configuration, and API keys all live under a local directory, backed by a SQLite database and local vector search. The project states it has zero telemetry, with no analytics, tracking, usage statistics, remote logging, or crash reports, and the README frames this as an architectural decision rather than an opt-out setting. Native speech to text through whisper.cpp and text to speech are built in, totaling about 130 megabytes and running offline.

Providers and memory

OpenCrabs supports a long list of providers. The README names Xiaomi MiMo, Anthropic Claude, OpenAI, GitHub Copilot using a Copilot subscription, OpenRouter with more than 400 models, MiniMax, Google Gemini, z.ai GLM, Moonshot Kimi, Claude CLI, OpenCode CLI, Codex CLI, Qwen Native with free OAuth and multi-account rotation, Qwen Code CLI, and any OpenAI compatible API such as Ollama, LM Studio, and LocalAI. Model lists are fetched live from provider APIs. A user can configure a chain of fallback providers so a failed primary is retried in sequence. Per provider vision is supported by calling a vision model on the same provider through a tool, and prompt caching is used across caching capable providers, with the README citing about 87 percent cache efficiency in real use. The context window defaults to 200,000 tokens with transparent auto-compaction at 65 percent in the background and 90 percent synchronously, which the project says gives effectively unlimited session memory. Memory is organized in three tiers: a user curated brain memory file, daily logs from auto-compaction, and a hybrid memory search that combines full text search with vector embeddings through reciprocal rank fusion, with local, API, or full text only modes. A dynamic brain system assembles the system prompt from workspace markdown files that are editable live between turns.

Tools, voice, and self-improvement

Beyond chat, OpenCrabs includes multimodal and document features. It accepts image and video attachments, with video routed through Google Gemini's multimodal API when enabled. PDF and document parsing cover PDF, DOC, DOCX, the Excel and ODS spreadsheet variants, CSV, HTML, and common text formats, all parsed in native Rust with no external services. Document generation creates XLSX with live formulas, DOCX, and PDF natively, plus PPTX through python-pptx when present. Voice support includes speech to text through Groq Whisper, any OpenAI compatible STT endpoint, a self-hosted Voicebox stack, or local whisper.cpp, and text to speech through OpenAI, compatible endpoints, Voicebox, or local Piper. Multi-agent orchestration can spawn typed child agents for general, explore, plan, code, and research work, each with a role specific prompt and a filtered tool registry. A recursive self-improvement feature is labeled experimental: a feedback ledger tracks tool executions, user corrections, and provider errors, and a self-improve tool can apply brain file changes without human approval, logging them to a local file with daily archives. An upstream template sync detects new releases and merges only new sections of brain file templates, taking backups before each merge. The project ships as a terminal UI, a command line, and a daemon, and connects to messaging channels. The README also documents skill templates for security audit, cost estimate, repo audit, multi-agent, browser automation through CDP, agent to agent gateway, and dynamic tools.

Editorial conclusion

The repository is published under the MIT license and its most recent commit was recorded on 2026-08-26. The source lives at https://github.com/adolfousier/opencrabs.