Zero: open-source terminal AI agent supports 25+ model providers

The agent inspects and edits files inside a repository, runs shell commands, and can interact with web content, all from a local session that never uploads history to the cloud. Vision support is available too, depending on which model is connected on the backend at the time, and the same agent can also run in a headless mode built for scripting, using stream-JSON input and output instead of a live interactive terminal.

Nothing risky happens without asking first. Reading files in the workspace is allowed by default, but writing outside it requires an explicit --add-dir flag, and shell commands, network access, or anything destructive sit behind a permission gate that has to be cleared first. Secrets get redacted automatically from logs and tool output before anything gets written down, and unsafe or fully autonomous modes stay off unless a user explicitly opts in.

More than 25 providers plug into the same agent: commercial APIs like OpenAI, Anthropic, and Google Gemini, alongside DeepSeek, Mistral, Qwen, Kimi, and others, plus any custom endpoint that speaks the OpenAI or Anthropic API format. Anyone who wants to skip cloud models entirely can run it fully local through Ollama or LM Studio instead, with no API calls leaving the machine at all.

The extensibility stack runs deep: MCP servers connect through a dedicated command, skills are just markdown instructions the agent can follow, and plugins add custom tools, hooks, and prompts through a manifest file. Sessions themselves are searchable and resumable, stored locally on disk, with no account or login tying them to a specific provider. For riskier experiments, Zero can also spin up an isolated git worktree, keeping changes separate from the main working copy until they’re reviewed.