Open Orchestrators
Back to directory
Rv

Player profile

Raven

An Apache-2.0 host agent that plans a task across built-in research, code, design, and on-call agents, and across third-party agents reached over ACP, a CLI, or an OpenAI-compatible API.

6.3k GitHub stars
Raven WebUI new task page: a sidebar with New task, Connectors, Knowledge, Task board, Import conversation, and Settings, and a composer set to deepseek-flash.
Raven WebUI from the EverMind-AI/Raven README, retrieved November 13, 2026; not a local runtime test. Raven GitHub README

Profile

Where it fits

Raven is a host agent from EverMind. You install it on Linux, macOS, or WSL2 with the curl installer at raven.evermind.ai/install.sh, or on Windows with install.ps1. A piped install uses the published wheel. Running install.sh as a file from a checkout installs that checkout in editable mode. Docker Compose from docker/docker-compose.yml serves the page at http://localhost:18793 and needs Git and Docker on the host. raven web opens the WebUI and leaves Raven running in the background. raven web --stop stops that service. raven tui is the terminal UI. raven agent -m runs one task from the shell. Setup asks, for each built-in product, whether it uses the model it is tuned for or this Raven's LLM.

Four agents ship in that tree. Raven-Research writes a sourced report. Raven-Code implements, debugs, refactors, and does data work. Raven-Design makes slide decks, charts, diagrams, and web pages. Raven-Oncall runs a workflow for hours and stops for a person only when the workflow asks. The host can hand one task to one of them, or run several inside one workflow. Third-party agents connect over ACP, a CLI, or an OpenAI-compatible API. The README says there are presets for 15 third-party agents and nine Mirasim-routed presets, and it names MiniMax Code on mcode acp and DSH on dsh --profile acp. The WebUI roster screenshot also shows Codex, OpenCode, Hermes Agent, OpenClaw, and GitHub Copilot.

An editor can launch Raven itself as an ACP server with raven acp. A2A is a separate HTTP JSON-RPC face: raven a2a serve listens on loopback port 8710 unless you change it. MCP is how Raven reaches an external tool or resource, not how it assigns an agent task. In the WebUI, Add on the Plugins page leads into chat, where Raven connects a server from the PlugHub catalogue. A server the catalogue does not list is written by hand into tools.mcpServers in config.json. raven plugin auth handles OAuth for a configured server. A package may ship an mcp.json, and Raven starts those servers only after you name the package.

EverOS memory keeps user context, agent experience, and world knowledge across sessions. SkillForge retrieves skills from local libraries, that memory, and SkillHub. The README states a SkillHub catalog of 114,190 skills. The agent loop is split into Memory, Planning, Capability, and Action. The Curator can replace those for one agent, including tools and skills, and it does not install a change that fails its check. The README says the Curator is experimental and is not in the installed package. The Evolver is a separate tool that consumes Raven as a library and scores harness changes against benchmarks. It does not run inside a session.

Builder web analytics

Measure outcomes from Raven-run agent work

Raven plans a task across built-in and third-party agents. Separately configured product analytics can show the observed user-facing results of a change those agents helped prepare.

This is an optional measurement workflow, not a native Raven integration with Agent Analytics. Configure product tracking and authorized analytics access separately; the listing does not install tracking or report Raven task activity.

First loop to measure

  1. Raven plans a task and runs it across built-in or third-party agents, and a person reviews the result
  2. the reviewed change is deployed to a separately instrumented project
  3. a human or an authorized agent reads observed product outcomes
  4. the team uses that evidence to choose the next task

How to install Agent Analytics on Raven

Copyable product analytics prompt for Raven with Agent Analytics

If product tracking and analytics access are already configured, inspect the deployed change over an appropriate observation window. Report measured outcomes, missing evidence and limitations. Do not infer causation from before-and-after movement or treat Raven task activity as product usage.