███████╗██████╗ ██╗ █████╗ ███████╗██╗ ██╗
██╔════╝██╔══██╗██║ ██╔══██╗██╔════╝██║ ██║
███████╗██████╔╝██║ ███████║███████╗███████║
╚════██║██╔═══╝ ██║ ██╔══██║╚════██║██╔══██║
███████║██║ ███████╗██║ ██║███████║██║ ██║
╚══════╝╚═╝ ╚══════╝╚═╝ ╚═╝╚══════╝╚═╝ ╚═╝
A production-ready, self-improving, multi-provider autonomous agent engine built for reliability, precision, and agency workflows.
Splash Core (v3.1.1) is an autonomous execution engine—not a chatbot. It is a strictly contract-based, multi-modal AI system that executes real-world tasks with precision, automatically audits its own work, and evolves its memory layers from every operation. Built for developers and agencies looking for deterministic outcomes over conversational fluff.
- Contract-First Execution: Every multi-step task generates a typed JSON execution contract outlining steps, token budgets, and verification criteria before execution begins.
- Advanced Memory Architecture: A robust 4-layer memory system (Working, Episodic, Semantic, Procedural) with a dedicated Trash Memory quarantine to prevent adversarial noise from polluting core skills.
- Safety as a First-Class Concern: Four configurable safety tiers (permissive, standard, strict, paranoid) physically block destructive actions and leakages, validated by independent safety filters.
- Instant Fast-Paths: Deterministic sub-100ms bypasses for common operations, sidestepping the LLM loop entirely to save tokens and time.
- Multi-Provider Arbitration: Native routing to OpenRouter, Claude, OpenAI, Gemini, DeepSeek, and local Ollama deployments with automatic rate-limit fallbacks.
You can install Splash with a single curl command on macOS and Linux:
curl -fsSL https://raw.githubusercontent.com/splash-agent/splash/main/scripts/install.sh | bashFor Windows PowerShell and manual installation methods, see docs/INSTALL.md.
Initialize Configuration:
splash initSplash natively supports execution directly from the command line:
# Bypass the LLM entirely for deterministic system commands
splash "list files in src/"# Engages the full contract builder and execution loop
splash "debug the memory leak in packages/core/src/agent-loop.ts"# Detaches execution to the daemon so you can track it later
splash "build a FastAPI todo service" --background- Contract Builder: Generates deterministic execution plans (
packages/core/src/contract-builder.ts). - State Machine: Strongly-typed execution lifecycle (
packages/core/src/state-machine.ts). - Reflector: Evaluates executed steps and extracts permanent learnings, scoring quality via an automated heuristic.
- Semantic Memory: Long-term storage of validated facts.
- Episodic Memory: Context tracking across sessions.
- Trash Memory: Quarantines failed, noisy, or adversarial prompts to prevent skill pollution.
- 36 Built-in foundational skills loaded dynamically via registry.
- Heavy dependencies dynamically imported to preserve <100ms startup times.
- Real: Filesystem, Webhook, Terminal, Browser, HTTP.
- Configuration Available / Coming Soon: Gmail, SMS, Telegram, Discord, Slack, WhatsApp. (Note: these gateway connectors currently validate config but await full real-world REST API mappings).
- Fully-featured non-interactive mode.
- Interactive
ink-based Terminal User Interface (splash tui) for real-time state monitoring.
- Native API connectors for Anthropic, OpenAI, Google Gemini, Groq, Mistral, Together, DeepSeek, and Ollama.
- Done: V3 Engine (Contract Execution, Memory Layers, Safety Engine, Trash Memory).
- In Progress: Transitioning simulated gateway connectors (Discord, Telegram, Slack) to real REST/WebSocket API implementations.
- Planned: Full MCP (Model Context Protocol) plugin integration for cross-agent compatibility.
- docs/COMMANDS.md — Complete CLI reference manual.
- docs/INSTALL.md — Detailed OS-specific installation instructions.
- docs/VERIFICATION.md — Test suite and operational proof records.
MIT License. See LICENSE for details.
Built with precision for the autonomous future.