Real-time monitoring, orchestration, and self-healing system for AI coding agent fleets across Claude, Codex, and Gemini.
vibe_cockpit is a fleet management layer for large-scale AI coding agent scenarios, positioned as a "command tower" rather than a code executor. It collects metrics from 15+ data sources via SSH (session health, token consumption, git commits, system resources, rate limits, network activity, etc.) to centrally monitor 30+ AI coding agent accounts across multiple Linux machines.
Data Collection Integrates with ntm (session orchestration), caut (account usage), cass (session metadata), ru (git commits), br (task status), bv (triage metrics), dcg (blocked commands), rano (network activity), process_triage (zombie processes), rch (remote builds), mcp_agent_mail (inter-agent messaging), cloud_benchmarker (VPS benchmarks), and more.
Multi-Interface Access
- TUI dashboard (ratatui): 12 screens (Overview/Machines/Repos/Accounts/Sessions/Mail/Alerts/Guardian/Oracle/Events/Beads/Settings)
- Web dashboard (Next.js + Axum): JSON API + WebSocket real-time push + static file serving
- CLI Robot mode: Structured JSON/TOON output for AI agent consumption
- MCP Server: Exposes tools and resources via Model Context Protocol
Prediction & Self-Healing
- Oracle prediction engine: Rate limit forecasting, Agent DNA behavioral fingerprinting, evolutionary optimization, experiment management
- Guardian self-healing system: Playbook execution, autopilot mode, approval workflows, automatic account switching
- Alert system: Rule-based threshold detection, pattern matching, cooldown tracking, multi-channel delivery
Knowledge Management Knowledge Base supports CRUD for solutions, patterns, prompts, and debug logs, with feedback scoring and keyword search.
Architecture Highlights
Storage uses DuckDB for analytical aggregate queries; collection employs incremental cursors, fail-soft, and idempotent insert strategies; SSH remote collection requires no agent installation on target machines. Built with Rust (95%) as a 12-crate workspace (vc_config / vc_collect / vc_store / vc_query / vc_oracle / vc_guardian / vc_knowledge / vc_alert / vc_tui / vc_web / vc_cli / vc_mcp), enforcing #![forbid(unsafe_code)] and Clippy pedantic lints project-wide. Async runtime uses Asupersync (structured concurrency + Cx capability tokens) with Tokio compatibility bridge.
Build & Deployment
Currently at v0.1.0 early development with no published releases. Requires Rust nightly (edition 2024) for source builds. Cargo.toml references local path dependencies ../frankensqlite and ../frankentui from the author's unpublished companion repos, making full builds inaccessible to general users. Configuration is managed via TOML files (machine inventory, collectors, alert rules, web server settings) with environment variable overrides and path expansion, but no example config is publicly available.
Special Note MIT license with restrictive clause: prohibits use by OpenAI, its affiliates, Anthropic, and its affiliates.