Privibe: Local-First Coding Agent Forked From Mistral Vibe

Privibe is an open-source CLI coding agent created as a fork of Mistral Vibe, built around the principle that development tools should not phone home. The creator's intent is clearly stated: "I want to be able to use it and know it will not be sending data of any kind anywhere." To achieve this, the project removes the cloud and account machinery entirely, strips out telemetry, tracing, update notifiers, plan offers, data-retention systems, and remote authentication. The Mistral SDK, previously a hard requirement, is now an optional extra installable via `uv sync --extra mistral`.

The tool is architected to work with a local llama.cpp server, which it communicates with via an OpenAI-compatible API. This design choice enables developers to run language models on their own machines. The backend adapters for Anthropic and Mistral that remain in the code are inherited from upstream Mistral Vibe but are not actively tested in this fork; the creator explicitly cautions users to treat them as unverified.

Privibe includes several optimizations tailored to local inference. Conversation state is stored in a ConversationList structure that keeps the prefix immutable, preserving the server's prompt and KV cache validity across turns. When a session is resumed using the `--resume` flag, the KV cache is restored from disk by reloading the original system prompt rather than regenerating it, which would invalidate the cache. The tool includes context-size auto-detection with a `/detect-context-size` toggle and an opt-in model warmup. A companion llama-server build enables per-message reasoning effort control via `/effort` commands (off/low/medium/xhigh) without invalidating the KV cache, though stock servers silently ignore these stamps.

For code editing, Privibe expands the toolkit inherited from Mistral Vibe with hashed-line file tools split into explicit single-line, block, and delete operations, plus a `find_symbol` tool. A per-agent file undo stack with a `restore_file` tool serializes writes. The tool is tuned for smaller local models with best-effort indent correction on edits, hashed line addresses re-pointed when earlier edits shift them, and naive reads of huge files that return a head preview plus guidance instead of flooding context. Cross-dialect path translation handles Windows, WSL, Git Bash, and Cygwin via a [paths] config section. File completion is backed by stateless git enumeration.

The user interface includes a console mode (`--console`) that runs the agent as a plain-text REPL without colors or TUI. A rewind feature (Alt+Up) lets users browse previous messages and fork conversations from any point, with optional file restoration gated behind a typed confirmation code. Messages can be queued mid-turn instead of cancelling the agent, and a double-press of Ctrl+C is required to exit. The `/resume` session picker shows folder context and a short preview of each session with search, filtering, and ranking. File edits display a diff in the TUI, and the context footer shows model name and tokens per second. Configurable toggles include `/preview-lines`, `/scrollback`, and `/llm-debug`.

Installation requires Python ≥ 3.12 and uv. On first run, Privibe writes a config to ~/.privibe/config.toml (relocatable via PRIVIBE_HOME environment variable), where users point model entries at their local llama.cpp instance. The project includes a bundled sample skill called `lets-document`, which auto-discovers and surfaces load errors. Model selection skips entries with missing API-key environment variables and falls back to a valid model; config upgrades append commented stubs for new keys. The project uses datetime-based version stamping and provides .deb and Windows-zip build scripts. Privibe is licensed under Apache-2.0 as a fork of Mistral Vibe (© Mistral AI) and is available at https://github.com/alainnothere/privibe.git.