Categories
AI Agents Self-Hosted AI Workflow Automation AI Assistant Multi-Agent Systems AI Chat MCP Client Deep ResearchStars
Forks
Watchers
Developer links
Eidon
Beyond a plain chat for quick questions, Eidon gives you a team of AI agents, each with its own browser, workspace, and memory, and a Chief of Staff that routes your requests. Ask for anything and it answers, hands the job to a specialist, or creates a new agent for it, and agents message each other and report back. Open an agent's live browser view and watch it work. At a sign-in, a two-factor code, a CAPTCHA, or a payment it stops and waits for you, and you hand control back with a note. Passwords come from an encrypted vault and are typed into the page, never passed to the model. Shell commands and MCP tools run only after you allow them once, always, or never, and emails and Slack messages arrive as drafts to edit. Chat lets you rewind, branch, edit, or regenerate any message, queue follow-ups while a reply streams, and share a read-only link. Deep research drafts an editable plan of up to twelve steps and returns a cited report. Memories wait for your approval. Automations run on an interval, daily, weekly, or once, and alert you through ntfy, Pushover, webhooks, or web push. Connect OpenAI-compatible or Anthropic providers, MCP servers, web search, image generation, and voice dictation, with a private account for each person. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. AGPL-3.0 licensed.
Benefits
- Delegate work to agent teammates
- A Chief of Staff routes jobs to specialist agents, up to 25 per user, each with its own chat, files, memory, and browser session, and any agent can message another directly.
- Step in when agents get stuck
- Watch an agent's browser live, take over for sign-ins, two-factor codes, or CAPTCHAs, then return control with a note; the agent waits up to 30 minutes for you.
- Approve actions before they run
- Allow shell commands and MCP tools once, always, or never; emails and Slack messages arrive as drafts to edit, and any single run or whole agent chain can be stopped.
- Keep chats, keys, and data yours
- One volume holds the database, attachments, and browser profiles; provider keys are encrypted with AES-256-GCM, every user's data stays private, and no Eidon-operated service is called.
Features
- Agents with live browsers
- Persistent agents run shell, Python, and a sandboxed Chromium through agent-browser, confined by Landlock or bubblewrap, with a WebSocket live view and a take-control stage.
- Streaming chat with branching
- Replies stream over WebSocket to every open device, with rewind, fork, edit-and-restart, regenerate, queued follow-ups, hierarchical context compaction, personas, and read-only share links.
- Editable deep research
- Drafts a plan of up to 12 editable steps, runs parallel web searches and full page reads, then writes a cited Markdown report with a sources list.
- Scheduled automations
- Run prompts every N minutes, daily, weekly, or once, inside a chat or an agent, with saved transcripts and ntfy, Pushover, webhook, or web push notifications.
- MCP servers and skills
- Connect stdio or streamable HTTP MCP servers with OAuth, load reusable skills shared across the team, and add web search through Exa, Tavily, or SearXNG.
- Memory with local recall
- Approve, edit, and pin memories, with semantic recall from a multilingual embedding model running on your server's CPU, so no embedding API is called.