Raven is a host agent harness that brings built-in and third-party agents together on complex tasks. It turns a task into a DAG, delegates focused work to a single agent or orchestrates several agents in a shared workflow, and carries memory and context across sessions through EverOS. The project is pre-alpha, so interfaces and configuration change quickly.
Four agents ship built in: Raven-Research for deep research reports with traceable sources, Raven-Code for feature work, debugging, refactoring, and data analysis, Raven-Design for slide decks, brand assets, charts, diagrams, and web interfaces, and Raven-Oncall for unattended workflows that run for hours or overnight. Raven also connects to third-party agents over ACP, CLI, or OpenAI-compatible APIs, with presets for 13 of them, including Claude Code, Codex, OpenCode, Hermes Agent, OpenClaw, GitHub Copilot, Qwen Code, Grok Build, and Kimi Code.
The harness is built for recursive self-improvement. Each agent loop splits into four strategy modules (Memory, Planning, Capability, and Action), and an experimental Curator rewrites those modules for a given agent, changing its tools, skills, and decision code round after round. Nothing is installed before it passes verification. A separate Raven Evolver tool evaluates candidate harness changes against benchmarks and keeps those that beat the baseline. Describing a desired assistant makes the Curator assemble a Persona: a lead role plus specialists drawn from the available agents.
Raven runs from a terminal-first CLI with a TUI, a browser WebUI for conversations, task progress, files, memory, and skills, or a Docker container.
Features
- Agent orchestration: coordinates agents, manages task dependencies and parallel execution, and turns multi-step collaboration into reusable workflows
- Built-in agents: Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall for research, coding, visual design, and unattended automation
- Third-party agents: presets for 13 external agents connected over ACP, CLI, or OpenAI-compatible APIs
- Evolver: diagnoses failures, tests candidate harness improvements, and keeps changes that outperform the baseline in reproducible evaluations
- Curator: rewrites an agent's memory, planning, capability, and action modules round by round, with verification before install (experimental)
- EverOS memory: keeps user context, agent experience, and world knowledge across sessions
- SkillForge: retrieves skills from local libraries, EverOS memory, and a SkillHub catalog of over 114,000 skills
- Proactivity: event monitoring and scheduled execution for reminders and follow-up work
- WebUI: browser view of conversations, multi-agent collaboration, workspace files, memory, and skills
