Raven

EverMind-AIApache-2.0活跃维护

一个可持久运行、能自我进化的多智能体协作框架,让多个大模型智能体跨领域分工协作,适合搭建长期在线的复杂任务编排系统。

The Harness of Harnesses • Super Agent: a trusted, persistent, self-evolving multi-agent ecosystem for all-domain collaboration.

v0.2.7发布于 2026-11-03 16:25 UTC6,142star +44 / 7 天250fork同步于 2026-11-10 11:34 UTC

怎么上手

由 README 提炼,执行前请确认
README 未说明
硬件要求按 README 说明

README 未说明

有 Docker 镜像
下载:源码 zip·GitHub Releases(安装包)本站不镜像安装包

项目说明

由 README 自动翻译提炼

多智能体编排框架,用 DAG 调度内置与第三方 Agent 完成复杂任务

一个可持久运行、能自我进化的多智能体协作框架,让多个大模型智能体跨领域分工协作,适合搭建长期在线的复杂任务编排系统。

核心能力

  • 生成 DAG 并编排多个专用 Agent 协同执行任务
  • 内置 Research、Code、Design、Oncall 四个 Agent
  • 用 EverOS 跨会话保留记忆与上下文
  • 轨迹视图以瀑布图展示多 Agent 每步耗时
  • Raven Evolver 评估并采纳 harness 改进

容易踩的坑

README 未给出安装命令,需查阅官方文档;内置 Agent 依赖 EverOS 提供记忆与上下文

内容由模型根据 README 生成,可能不完整或不准确,以官方文档为准。看原文 README

最近发版

5 个版本v0.2.7
  1. v0.2.7Raven 0.2.7 (2026-11-03)最新正式版2026-11-03 · 2026-11-03 16:25 UTC

    Raven 0.2.7 is a patch release. Each conversation now keeps a Task Board of the matters it is pushing, with a Task Board page across conversations; the trajectory page is redrawn as an agent waterfall that shows an external agent's own steps and can replay a session from its first instant; a stuck DAG node can be handed to another sub-agent mid-run; Raven's own product agents share memory with their host, so five of them cost about half of what they did; and the trace store compresses what it writes. Also in this release: a conversation can ask for more or less reasoning than the configured default, DeepSeek joins the web search and fetch vendors, Skill Hub retrieval is on for onboarded installs with a switch, a self-signed model gateway can be reached with a per-section tls_verify opt-out, and an owner can name a registered machine by address, port or nickname.

    Review the upgrade notes: product agents now run forked from a template and relay their model calls through the host, Skill Hub retrieval is on by default once onboarded and writes a key an older build refuses, new trace artifacts are written compressed and cannot be read by older code, the trajectory page drops several controls, a reply that is only a silence marker is not delivered, the research high tier closes a dry search earlier, and the tool_failure_count field leaves the tasks-list wire.

    Highlights

    Task Board

    Raven keeps a Task Board for each conversation: the matters it is pushing (task groups), each turn's progress on them, the sub-agents and DAG runs a turn started for them, and where each matter stands -- needs you, running, idle or completed. The web UI draws it in the conversation's tasks tab and on a new Task Board page across conversations, and the model reads it back with a task_board tool (#1157).

    • One card per open group, in columns by category or as a list, with completed groups behind a switch. A card unfolds its progress steps in place; clicking one opens its conversation on that turn. A group that needs you frames each open question with Ignore and Answer; Move on puts a bar over the composer so the next message carries the group with it (#1157).

    • How it is kept: one extra model call after every turn that answered or dispatched work, run in the background on the turn's own cached prefix. The board lives in task_board.json beside the session, follows undo, branch, clear and delete, and only the Raven host keeps one. agents.defaults.taskBoard.enabled (on by default) turns it off, which also removes the tool (#1157).

    Trajectory waterfall and replay

    • The trajectory page is an agent waterfall: every entry on one time axis, grouped by agent, then turn, then span, with idle stretches over 60 s folded, a header of six figures, an activity strip of calls per bucket of real time, and a clock ruler. Fit, 3h, 1h and 15m windows, drag on the strip to window, Ctrl/Cmd + wheel to zoom, Shift + wheel to pan. The rows are a keyboard tree, the details pane floats as a card, and a dispatching row's door arrow leads to the worker it started. Gone: the duration threshold, the hidden-entries switch, the toolbar and reset button, the now line, the dispatch rules and the scope tab strip (#1147).

    • An external agent's own steps are recorded: what an ACP agent said, thought and ran is written as model and tool spans under its run, at most 400 per run, live while the run lasts. raven trajectory backfill-frames writes those steps for runs recorded before, from the wire journal, which is now kept 90 days instead of 7. The details pane gains a Reasoning tab on a model call (#1147).

    • Replay in the header draws the conversation again fro

  2. v0.2.6Raven 0.2.6 (2026-10-27)2026-10-27 · 2026-10-27 13:24 UTC
    查看更新说明

    Raven 0.2.6 is a patch release. This release changes how a long conversation stays inside the model's context window: the new compacting history strategy is the default, and the previous Curator is one setting away. It also reads portable Agent Plugins packages, strengthens multi-task management, ships the EverOS Cloud memory plugin with Raven, and falls back to mirrors when a download host cannot be reached. MiniMax image and video, Keenable as a keyless web search and the Opper gateway join the model and provider choices.

    Review the upgrade notes: the context history default and its reply budget change, the Claude Code, Codex and Pi adapter pins move, Shift+Enter and Ctrl+Shift keys change in the TUI on native Windows, Raven's own ACP server reports a running cost total, and a playbook route that names an undeclared param is refused.

    Highlights

    Compacting context history

    context.history.strategy now defaults to compacting. A long session is kept verbatim until the window nears a trigger that scales with its size; then large tool results are shortened to a one-line record, older messages are dropped, and past that a summary is written that survives a restart. Set context.history.strategy to "curator" to restore the previous behaviour (#1031).

    • Read a shortened result back: the new context_retrieve tool reads a shortened tool result in full, and runs without asking. Tool results are kept up to context.toolResultMaxChars (default 1,000,000 characters, previously 16,000), so session files grow. Copies that leave the session -- the session.resume transcript, a memory backend's turn and the consolidator's prompt -- carry at most 16,000 characters of each result, and the model reads the rest back with context_retrieve (#1031).

    • Memory and skills move: recalled memory and routed skills now go at the end of the turn's user message instead of the system prompt (context.cache.turnScratch, default true), which keeps the cached system prompt stable (#1031).

    • A reply budget beside the prompt: a request near the window asks for no more reply than fits beside its prompt, never below context.headroom.minTokens (4,096), so a reply near the window can be shorter (#1031).

    • A stable context ring: the web composer's context ring shows the same number after a reload or a conversation switch as it did when the turn finished, read from the window the last model call measured. Conversations from before the upgrade keep the old estimate until their next reply (#1119).

    • Tuning and rollback: context.history.compacting.summary.mode = "off" keeps the fold and writes no model summary, context.history.compacting.prune.enabled = false turns the one-line tool records off, and context.cache.turnScratch = false returns memory and skills to the system prompt (#1031).

    Agent Plugins packages

    Raven reads portable Agent Plugins 1.0 packages: a directory with a plugin.json, plus optional skills/ and mcp.json (#1110).

    • raven agent-plugins list | inspect | import | remove | check (each with --json) validates a package and imports its skills into <workspace>/skills/<package>/. The import is tracked in agent-plugins.json in the Raven home, so remove deletes only what was imported and check reports a package that changed since. Imported skills lose their always flag, and a skill on skillForge.blocklist or one whose body points at another product's home is refused (#1110).

    • List package locations in agentPlugins.packages and opt-outs

  3. v0.2.5Raven 0.2.5 (2026-10-20)2026-10-20 · 2026-10-20 14:21 UTC
    查看更新说明

    Raven 0.2.5 is a patch release. This release strengthens multi-agent coding, adds remote agent execution on registered machines, introduces conditional branches in DAG workflows, and improves native Windows support. Raven can delegate tasks to connected agents on any registered machine, using each agent's capabilities and working environment. Multi-agent coding can use this capability to coordinate local and remote coding agents. Document support is also expanded.

    The trajectory view is now available to everyone, alongside conversation export and import and an optional EverOS Cloud memory backend. Review the upgrade notes for prompt customization, DAG behavior and time-zone changes.

    Highlights

    Native Windows support

    • Installation: install.ps1 correctly finds the latest release in Windows PowerShell 5.1. The direct installer URL below remains available when PowerShell cannot follow the website redirect (#863).

    • Agent startup: Bundled ACP agents start correctly, saved launch commands preserve backslashes, and Check again, Connect and Test discover agents installed after Raven started (#864, #880, #908, #912).

    • Task cleanup: Agents terminate with their host, CLI sub-agents stop on timeout or cancellation, and connecting agents with multiple processes, such as Claude Code, no longer fails with WinError 32 (#927, #951, #974).

    • File handling: File tools reject reserved Windows device names such as nul, aux and con.txt anywhere in a path, preventing writes that appear to succeed without creating a file (#918).

    • Remote tools and operations: Windows hosts can lend MCP servers to remote ACP agents, and Raven-Oncall campaign jobs run on Windows (#972, #1016).

    For Windows users: raven upgrade starts a separate helper; wait for its completion message before restarting Raven. After installing or upgrading, run raven resources once to download the knowledge reader's PDF models and word dictionary (about 110 MB). The Windows installer does not download them.

    Remote agents on registered machines

    Remote agent execution extends Raven's agent roster across registered machines. Raven can invoke connected agents for the work they support, run tasks in their remote working directories, and collect the results. From the Machines tab, register a machine, install its Raven node, then connect an agent and choose its working directory on that machine (#983, #1007).

    • Run ACP, CLI and OpenAI agents on registered machines over SSH by setting the entry's machine field. A Raven node owns each run, so a dropped SSH connection does not end the task. raven ops connection install-node installs or updates the node (--dir chooses its location; --prune removes other versions). raven ops connection remove unregisters the machine and leaves its node installed. Installation uses the machine's configured package index without exposing its credentials on command lines (#893, #920, #963, #965).

    • File tools accept remote paths such as user@host:/path or scp://user@host:port//path. read_file, list_dir, find and grep ex

  4. v0.2.4Raven 0.2.4 (2026-10-03)2026-10-03 · 2026-10-03 16:36 UTC
    查看更新说明

    Raven 0.2.4 is a patch release. Raven can now read and change its own configuration from the conversation: a new raven_config tool reads any setting without asking, asks before every write, and never takes a key through the chat (#833). Four things behave differently once you upgrade. The approval sheet on the page no longer answers to 1, 2 and 3: Esc denies, Cmd+Enter (Ctrl+Enter) allows once, and Shift+Cmd+Enter gives the broader grant (#817). Channels have left the Settings dialog: the rail row that was Agent Connector is now Connectors, with Agents and Channels as its two tabs (#830). Every turn the gateway runs (the page, chat apps, cron) now takes the per-turn workspace checkpoint, a shadow git repository at .raven/shadow.git in the workspace, which is what lets a file an exec command wrote open as a diff; runtime.checkpoint.policy: "never" turns it off (#829, #849). And the main agent answers in plainer, shorter prose, using headers, bullet menus and emoji only when you ask for them (#853). No command, flag or config key is removed or renamed.

    Highlights

    Raven configures itself

    • The raven_config tool has seven actions: describe (every setting with its value, or a search), get, set (one path, or several under one confirmation), unset, add (connect preset agents, several at once), test and restart (reload or restart once pending changes are gathered). Its catalog covers 14 sections and about 93 settings, each saying when a change takes effect: next turn, at once, on reload, on restart, or on the memory server. Writes go through the same methods the Settings page uses, so both are checked the same way. Setting session.model switches only the current conversation (#833).

    • Every write asks first, on a config card that shows what changes, the value before and after, and when it takes effect. Full access, your own allow rule or the smart-mode reviewer can let a change through, but only in a turn someone is at, and the reviewer never approves a sensitive setting (approval mode, sandbox, deny patterns, MCP servers, provider endpoints, who may instruct Raven on a channel, lent keys and the like). Cron and other unattended turns cannot change the configuration (#833).

    • A key never goes through a tool call. The agent names the setting and a credential card on the page takes the key, which the host writes; the model only learns whether it was saved or skipped. A key passed to the tool directly is not written, and the agent is told to ask you to rotate it. In the terminal or a chat app, where no card can be shown, you are told where to enter it. Tool output is scrubbed of the keys Raven holds. The credential card answers to the same keys as the approval sheet: Esc skips, Cmd+Enter saves (#833, #849).

    • Raven can connect preset agents, test them and work out why one does not answer. On the Agents page, a refused connect or a failed test offers "Hand it to Raven", which opens a new conversation naming the agent and the reason, without sending it. An ACP agent can be started with one of Raven's provider keys (lendKeys; Pi only for now), and the connect card says which key it lends (#833, #849).

    • An EverOS memory LLM left unset now follows the main model, and the memory server restarts when the main model changes (#833).

    The page

    • Connectors:
  5. v0.2.3Raven 0.2.3 (2026-09-27)2026-09-27 · 2026-09-27 04:05 UTC
    查看更新说明

    Raven 0.2.3 is a patch release with two things worth reading before you upgrade. Bare raven now opens the browser page, the way raven web does, and falls back to the terminal UI only where no browser can be opened; raven tui still opens the TUI outright. And the page's own update row now works on a normal raven web install and shows the download's progress, but only once the page is served by this release: a page served by 0.2.2 or earlier still refuses it, so this one upgrade goes through the command line as before (#806). No command, flag or config key is removed or renamed. Two smaller removals on the page: the file viewer no longer offers a download button, and HTML previews run their scripts by default instead of asking (#805).

    Highlights

    Upgrade from the page

    • The update row at the bottom of the sidebar upgrades a raven web install: the gateway hands the page its own graceful stop on every mount, exits cleanly on upgrade, and the helper waits for the supervisor, installs, and relaunches raven web --supervise on the same port with the same token and cookie, so the page reloads itself still signed in. It is refused, each with its own reason, while a turn, a sub-agent or a question waiting for an answer is running (including in an IM channel), and for a Raven started by hand with raven gateway or raven web --foreground, which has nothing to restart it (#806).

    • The page shows what is moving. The helper downloads the release itself so the bytes can be counted, then answers /upgrade/status on the page's own port with phase, bytes and rate while the old Raven is gone; the card measures the bar while bytes are counted and says when the new Raven is starting. A failed install is held for the page to read before the old version comes back, instead of relaunching it silently. The twenty-minute ceiling now counts silence, not the whole upgrade, so a slow download that is visibly moving is not given up on (#806).

    • The terminal raven upgrade prints each asset's size and the total before downloading, then a live progress line, so a large release on a slow link no longer reads as a stalled install. The downloaded wheels stay under ~/.raven/cache/upgrade/<version> (about 75 MB for a release with its engines); earlier versions' files are cleared (#806).

    • raven gateway refuses a half-written install the way raven serve already did: a gateway that starts inside an upgrade window waits it out and exits instead of serving the placeholder page until restarted by hand. The gateway also closes its websockets on shutdown, so raven web --stop with a page open no longer waits on the page's own socket (46 seconds down to 4 ms in the PR's run) (#806).

    • Bare raven opens the page, leaving the resident gateway that raven web always leaves. Where no browser can be opened it runs the TUI instead, decided before anything starts. raven web and raven tui are unchanged (#806).

    Desk panes

    • Every desk pane has one header that names what is on screen. The file path bar, the task status line and the node card's back-and-title row fold into it: a glyph or a back chevron, the title, a muted facts line, then the controls that act on the object. A picked node reads as "task > node"; narrow panes drop the parent crumb and keep the facts on their own line. Dragging a pane by its title works again (#805).

    • Stop leaves the palette's task list and lives in the task pane header, shown only while the task or the picked node is running. The runtime has no per-node cancel for a graph, so on a graph node it st

用户点评

登录用户的真实使用记录,可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译

    Raven banner

    Website · Documentation · 中文

    ❯❯ What's New

    ❯ Multi-Agent Trajectory Observability

    Every agent, every turn, every step — on one timeline. Raven's new trajectory view renders a running multi-agent session as a live waterfall, from the whole run down to a single step.

    https://github.com/user-attachments/assets/abb0a5e9-1c37-48bd-9be0-1c9267293337

    What is Raven

    Raven is the harness of harnesses, built for recursive self-improvement (RSI). As a Host Agent, it brings built-in and third-party agents together to carry out complex tasks. Its modular architecture supports iterative improvement of Raven's own harness: proposing changes to how agents plan and act, evaluating those changes, and adopting improvements that pass validation. Powered by EverOS, Raven carries memory and context across sessions to support this process.

    Built-in Agents: Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall support research, coding, visual design, and unattended workflow automation.

    ❯❯ Showcase

    These are three complete projects delivered by Raven. In each case, Raven drove a team of specialized agents from the initial brief or objective through execution to a complete set of final deliverables.

    ❯ THRESHOLD: a complete game development project

    The brief came from a person; Raven completed the entire project. Working autonomously for about 4 days, Raven completed 42 rounds of planning, development, and verification to build a playable first-person shooter in Godot 4, centered on an arena boss fight. The full deliverable includes the game, its poster, presentation, and website, all produced by Raven.

    https://github.com/user-attachments/assets/44724e3e-6564-467a-b422-473aa8f474bd

    ❯ Raven RSI: recursive self-improvement in practice

    AI that improves AI, with the entire project completed by Raven. Given a task and evaluation criteria it cannot modify, Raven RSI independently plans each round, writes code, runs experiments, and evaluates the results. In nanochat pre-training experiments, it completed 172 training runs across 7 rounds without a single crash, reducing val_bpb by 5.8% within the same 20-minute, single-GPU budget. The same process reduced overshoot in a dam-break simulation by three orders of magnitude and completed an FEA limit-load search in 8 rounds of bisection. The complete deliverable includes the experimental results, visualizations, poster, presentation, and project website, all produced by Raven.

    ❯ Raven: a complete product launch project

    One raven, a whole flock of specialists. Raven completed the entire project. Its launch kit brings together a browser-based physics mini-game, a 16-slide product overview, posters in English and Chinese, and the README you are reading now, all produced by Raven.

    https://github.com/user-attachments/assets/dd186b6d-3752-4c59-89c7-eeea5c6fa962

    More showcases

    ❯❯ Built-in Agents

    Raven's modular architecture is designed for harness self-evolution and subagent creation. Its four built-in agents deliver state-of-the-art (SOTA) performance in their respective domains, combining reusable harness components with domain-specific tools, skills, and agent loops. Raven can delegate a focused task to a single agent or orchestrate multiple agents within a shared workflow. The harness they share is refined by the Raven Evolver, a separate tool that consumes Raven as a library and evaluates candidate harness changes against benchmarks; it develops the agents rather than running inside them.

    All four agents are built in and ready for orchestration out of the box.

    ❯ Raven-Research

    Raven-Research enables autonomous deep research for complex questions, literature reviews, and technical analysis. It delivers clear, structured reports with traceable sources, helping users understand unfamiliar domains, compare alternatives, and make informed decisions.

    ❯ Raven-Code

    Raven-Code enables agentic software development, turning requirements into working, tested code. It supports feature implementation, debugging, refactoring, data processing, and data analysis, helping users build new capabilities, resolve issues, and improve code quality while following their project's conventions.

    ❯ Raven-Design

    Raven-Design performs visual design, turning ideas and content into polished visual deliverables. It creates PowerPoint slide decks, brand assets, charts, diagrams, and web interfaces, refining layout, typography, and visual consistency to help users communicate clearly and bring their ideas to life.

    ❯ Raven-Oncall

    Raven-Oncall enables unattended workflow automation for experimentation, optimization, and continuous monitoring. It autonomously manages workflows from start to completion, sustaining progress over hours or overnight, delivering results, and involving users only when human judgment is needed.

    ❯❯ Runtime Self-Evolution

    Raven is built for this from the ground up. Its agent loop is split into four decoupled strategy modules — Memory for what a turn gets to see, Planning for how it approaches the work, Capability for which tools an iteration exposes, Action for what to do next and for judging it before it runs. A Curator keeps rewriting those four seats: a setting, or a small piece of judgement code written for that agent. What it changes belongs to that agent alone, and once installed the agent runs on it.

    A Curator changes more than the prompt — the tools and outside services it reaches for, the skills and procedures it follows, and its own judgement at each point can all be replaced. It keeps going round after round — you put it to work, you say what was wrong, it reworks — until you are satisfied, until it has nothing left worth changing, or until the round budget runs out. Two things hold the quality: nothing is installed before it is verified, and a failed check sends it back; and a round's signals normally reach it with the reference answer stripped out, which limits how directly the answer is exposed. The Curator is experimental: it ships with the repository rather than the installed package.

    A Persona is the first thing it builds. Describe the assistant you want and the Curator assembles one: a lead role that talks to you, and specialists drawn from the agents you already have. What you asked for lands in that assistant's harness — its division of work, the tools it may reach for, and the checks it must pass before it acts.

    Describe what you need once. Raven assembles the assistant, keeps improving it while you work, and afterwards a sentence is enough to put it to work again.

    ❯❯ Connect Third-Party Agents

    Raven can connect to and orchestrate agents via ACP, CLI, or OpenAI-compatible APIs, with presets for 13 third-party agents to simplify setup, task delegation, and coordination across shared workflows. Try these agents in Raven through a unified interface!

    ❯❯ News

    • 📣 2026-09-30 — Raven's technical report reached #1 on Hugging Face Daily Papers — daily (2026-09-30), weekly (2026-10-04) and monthly (2026-10-10).

    ❯❯ Quick Start

    🤖 Install with Your Agent

    Let your own agent install Raven for you. Copy this prompt into any agent that can read a web page and run shell commands, such as Claude Code or Codex:

    Read https://evermind-ai.github.io/Raven/quick-start/ and follow it to install Raven, or to update it if it is already installed.
    

    📦 Install

    Linux, macOS, or WSL2:

    curl -fsSL https://raven.evermind.ai/install.sh | bash
    

    Native Windows PowerShell:

    irm https://raven.evermind.ai/install.ps1 | iex
    

    Windows PowerShell 5.1, the version built into Windows, cannot follow that URL's redirect and stops with (308) Permanent Redirect. If you see that error, use the direct installer URL instead:

    irm https://raw.githubusercontent.com/EverMind-AI/Raven/refs/heads/main/install.ps1 | iex
    

    Or run it in a container, with nothing on the host but Git and Docker:

    git clone https://github.com/EverMind-AI/Raven.git
    cd Raven
    docker compose -f docker/docker-compose.yml up
    

    Then open http://localhost:18793 in your browser. Prerequisites are Git and Docker with Docker Compose. See docker/README.md for what comes up, how the page signs itself in, and how to run it behind a remotely exposed port.

    Or install from a source checkout, to develop against the code or to run what has not been released yet:

    git clone https://github.com/EverMind-AI/Raven.git
    cd Raven
    ./install.sh
    

    Run as a file, install.sh installs that checkout in editable mode: raven and its bundled plugins link back to your tree, and the TUI bundle and the served page are built from it. A piped run installs the published wheel even from inside a clone, so that a one-line install never picks up whatever a working tree happens to contain. Set RAVEN_LOCAL_SRC=<dir> to force the editable install through a pipe.

    The agent products ship with raven itself: a wheel carries the agents/ product tree and copies it out to your raven home on first use, and a source checkout reads the tree in place. Setup asks, for each product, whether it runs on the model it is tuned for, which needs a key of its own, or on this raven's LLM. Nothing is registered: a folder is on the roster because it is there. See agents/README.md.

    Learn more about Raven on the documentation site.

    Read the documentation

    ❯❯ Core Systems

    SystemWhat it adds
    Agent OrchestrationCoordinates agents, manages task dependencies and parallel execution, and turns multi-step collaboration into reusable workflows.
    EvolverDrives harness self-evolution by diagnosing failures, testing candidate improvements, and retaining changes that outperform the baseline in reproducible evaluations.
    EverOS MemoryPreserves user context, agent experience, and world knowledge across sessions, recalling relevant memories and reusable skills for future tasks.
    SkillForgeRetrieves relevant skills from local libraries, EverOS memory, and SkillHub's catalog of 114,190 skills, giving agents specialized expertise on demand.
    ProactivityCombines event monitoring and scheduled execution to anticipate user needs, deliver timely reminders, and initiate follow-up work.

    ❯❯ Launch WebUI

    Raven's WebUI brings conversations, multi-agent collaboration, and workspace management into your browser. Chat with agents, follow task progress, inspect files and outputs, and browse memory and skills in one place.

    raven web
    

    The command opens the WebUI in your browser and keeps Raven running in the background. Use raven web --stop to stop the background service.

    ❯❯ EverMind Ecosystem

    EverMind connects memory research, production-ready products, and practical integrations into one open-source ecosystem.

    Together, these projects form EverMind's research-to-runtime stack: methods and benchmarks become reusable memory infrastructure, products, and agent integrations.

    ❯❯ Contributing

    Issues and pull requests are welcome. Start with the developer workflow, follow AGENTS.md for repository rules, and use GitHub Discussions for design conversations.

    ❯❯ License

    Apache License 2.0

    ❯❯ Citation

    If you use Raven in your research, please cite the technical report:

    @misc{evermind2026raven,
      title         = {{Raven: The Harness of Harnesses for Composable Agentic Intelligence}},
      author        = {{EverMind AI}},
      year          = {2026},
      eprint        = {2609.33439},
      archivePrefix = {arXiv},
      primaryClass  = {cs.AI},
      url           = {https://arxiv.org/abs/2609.33439}
    }
    

    ❯❯ Contributors

    Thanks to everyone who has helped make Raven better.