A standalone agent harness in Rust: a provider-agnostic loop, MCP tools, a path jail and prompt-injection interlock, sandboxed shell, scheduled...
mecha is a local-first agent harness built in Rust and MIT licensed, designed to give a local open-weight model context, permissions, and a safe way to reach the world. It is built for local open-weight models first — llama-server, vLLM, Ollama — with Anthropic available as a ceiling to measure against. It ships as one binary with five front ends: four in a terminal and one in a browser. Personal context arrives over MCP, so adding a source is configuration rather than a code change. Mail and calendar sit behind one surface, alongside a personalized knowledge graph for who people are and what happened when. The mail and calendar layer is provided by mecha-mail, a library plus four thin MCP binaries holding Gmail and Google Calendar v3, Outlook mail and calendar over Microsoft Graph, both OAuth flows, and the token lifecycle. The harness is built for the lethal trifecta: a personal assistant holds private data, reads other people's words, and can send. The interlock is structural — tools declare capabilities, the conversation carries the taint, and a send to a destination the model chose is refused before the human is ever asked. Naming a tool in the outbox stages its calls as drafts instead of executing them. Evaluation is graded on the trace rather than the claim: the eval rig checks the tool calls first and the prose second, runs a verify command for ground truth, and reports pass^k beside pass@k. mecha serve puts the same agent behind a web app on your tailnet, bound to loopback and fronted by Tailscale.
Teams use mecha to run a local open-weight model against their own context and permissions.
Users use mecha to triage an inbox overnight and review staged drafts rather than sent mail.
Users use mecha to connect every mail account behind one provider-neutral surface.
Users use mecha to answer scheduling questions from merged busy intervals across accounts.
Teams use mecha to evaluate models on tool calls and ground truth rather than prose.
Users use mecha to serve the same agent behind a web app on their tailnet.
Users use mecha to add personal context sources over MCP as configuration rather than code.
Users use mecha to stage sends as drafts by naming tools in the outbox.
Lynote combines AI-draft rewriting and AI-text detection with note-taking, transcription, video summaries, and study tools.
LMSYS Chatbot Arena is a crowdsourced open platform for LLM evals. Collected over 1,000,000 human pairwise comparisons to rank LLMs with the Bradley-Terry model and display the model ratings in Elo-scale.
Zhipu Qingyan is a GLM-based AI assistant whose official description emphasizes understanding goals, breaking down tasks, and using tools.
字节跳动AI助手