vedaant00/opendot
A terminal AI agent that snapshots every file and shell change before it happens, so you can undo it all cleanly
opendot is a terminal AI agent that edits files and runs shell commands, but it snapshots the workspace before every action so anything it does can be exactly undone with a single command. Commands that escape the workspace, like network calls, sudo, or git push, are flagged and require confirmation first since they can't be reversed. It works with any LLM through LiteLLM and can reach over 1000 external app tools via MCP servers or Composio.
What it does
- Snapshot-before-action: every file write or shell command is snapshotted into a content-addressed store first, so opendot undo/redo/diff can restore or preview the exact previous state.
- Risk classification: a classifier separates commands contained within the workspace (auto-run, undoable) from ones that escape it, like network access, sudo, or deleting outside the working directory (confirmed first, marked irreversible).
- Model-agnostic: through LiteLLM, opendot connects to OpenAI, Anthropic, Google, DeepSeek, Groq, and Hugging Face models with just an API key, plus fully local, free models via Ollama. Default model is gpt-5.1, auto-switching if that key is missing.
- Budget caps: flags like --usd or --tokens let you set a spend or token ceiling, and the run stops the moment accumulated cost or tokens cross that limit.
- External connectors: MCP servers and Composio (with your own API key) add access to app tools like Gmail, Slack, and GitHub, but every such external call is treated as irreversible and always confirmed before running.
Why it matters
For anyone hesitant to let an AI agent freely touch files or run shell commands, opendot offers a real safety net: mistakes can be precisely reversed rather than just hoped away. Its budget caps and permission policy (--yes, --allow, --deny) also make it usable for unattended automation like CI, where someone needs guardrails without a human watching every step.
Terms in this repo
- MCP · a standardized protocol letting AI models talk to external tools and servers
- LiteLLM · a library that unifies calls to many different LLM providers under one interface
- content-addressed store · a storage method that saves each unique file only once, keyed by its content
- append-only ledger · a log of actions that can only be added to, never edited or erased, so history stays intact
- Composio · a service that bundles tools from many external apps like Gmail and Slack behind one API
Repository description (English)
A terminal AI agent you can fully undo - every file and shell action is snapshotted and reversible. Model-agnostic (any LLM), and connects to 1000+ app tools and MCP servers.
Open on GitHubTrending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
- block/buzzAn open-source workspace where humans and AI agents chat, code, and review in the same rooms
- NousResearch/hermes-agentNous Research's Hermes is an AI agent that gets smarter the more you use it
- virgiliojr94/book-to-skillA tool that turns technical book PDFs into on-demand reference skills for AI coding agents
- VoltAgent/awesome-agent-skillsA single hub collecting over 1000 'how-to' manuals that make AI coding assistants act like experts
- anthropics/claude-plugins-communityA shared shelf where anyone's Claude add-ons get listed for install