Plurality

AI agentic assistant you can see and steer in real time

A free, self-hosted AI assistant that you can see in action, steer in real time, and easily debug.

Sort emails, messages, calendar... Virtually connect anything to it, with no config: just ask the AI to set it up for you

• Conversation and agents workflow in one UI
• Fully open-source, free, and local
• (almost) Any models from (almost) any providers
• Sandboxed agents with full control over their virtual environment (They can setup software in their VM)
• Tool calls, MCP, Image generation, Shell execution, skills and more (compatible with openClaw skills)
• Custom harness with tool approval
• Attach a folder to any conversation — code in it like Claude Code, draft documents, or let it organize your desktop
• Remote clients (vibe code on your PC from your phone)
• Native application (mobile and desktop) and web
• More features such as: CRON Scheduler, Webhook interception, Agent preset library, etc...

A chatbot conversation happens on the spot. You type, it answers, and when you close the tab, everything stops — nothing runs without you. Agents are different in kind, not degree: you give one a goal and it keeps working after you walk away, for minutes, hours, or days. It calls tools, monitors things, spawns sub-agents that work in parallel. That's real capability — but most of the work now happens while you're not there, in a place you can't see. Platforms like OpenClaw bolt a chat on top, but the chat is just a window the agent occasionally waves through.

Plurality covers both. Need a quick answer? Spawn a plain conversation — instant, disposable, nothing running behind it. Need work done over time? The long-running tasks unfold inside a fully featured conversation you can open at any moment to watch, stop, steer, or debug. Other platforms let the agent text you updates; Plurality shows you every inference the agent made, as a conversation. Even sub-agents appear as nested sub-conversations under the agent that spawned them.

Highlights

• See every action. Each tool call shows up inline with its arguments. File edits render as git-style diffs. Long jobs show a live checklist. Timers count down. Click any step to open the raw call and its result.
• Steer in real time. Hit stop to cancel an in-flight run. Require approval before a tool executes. Switch models mid-conversation.
• Watch the whole tree. Agents can launch parallel sub-agents; the UI nests them under their parent with live "typing / running tool" status so you can follow a fan-out without losing the thread.
• Know the cost. Every message carries its token counts and dollar cost. No guessing where the bill came from.
• Any model, one config. Routes through a local LiteLLM proxy, so OpenAI, Anthropic, Gemini, Fireworks, and others all work from a single YAML file — no provider lock-in, no hardcoded model list, keys stay on your box.
• Real agents. A server-side tool loop lets the assistant run shell commands, read and write files, search the web, generate images, and spin up parallel sub-agents.
• Background work. Schedule prompts with cron, or trigger them over HTTP with signed webhooks — the assistant keeps working when you're not watching.
• Multimodal. Text, vision, image generation/editing, speech-to-text, and TTS, gated per-model by capability.
• Presets / Mini-Apps. Save a system prompt + model + tool set as a reusable one-tap assistant.
• Hybrid search. Full-text (BM25) and semantic vector search over your whole history, merged into one ranking.
• MCP support. Plug in any Model Context Protocol server to add tools.
• One client, everywhere. A single Flutter app runs on web, desktop (Windows/ macOS/Linux), and mobile (iOS/Android).
• Auth your way. Local accounts or OpenID/OIDC (Authentik, Keycloak, Auth0, Google, …) via PKCE — see openid.md.
• Self-contained. Per-user SQLite, no external database to run.

Gallery

🗂️  Conversations and agents, one UI	⏰  Schedule prompts with cron
🤖  Spawn background sub-agents	🎨  Generate & edit images inline	✨  One-tap preset mini-apps
⚙️  Pick a model preset by cost	🔐  Per-tool approval (Off / Ask / On)	🚀  Guided preset launch	🌿  Nested sub-conversations

Quickstart

From Cosmos cloud, install in one click: install

Or from your server's terminal (This assumes you are using Open Router, see deployments and litellm.md for details on how to setup your AI models accesses)

docker run -d --name plurality \
  -p 8090:8090 \
  -v plurality-users:/app/users-data \
  -v plurality-data:/app/data \
  -e OPENROUTER_API_KEY=xxxxx \
  -e INIT_ADMIN_USER=admin \
  -e INIT_ADMIN_PASSWORD=change-me \
  ghcr.io/aseracorp/plurality

services:
  plurality2:
    image: ghcr.io/aseracorp/plurality
    container_name: plurality2
    ports:
      - "8090:8090"
    volumes:
      - plurality-data:/app/data
      - plurality-users:/app/users-data
    environment:
      - OPENROUTER_API_KEY=xxxxx
      - INIT_ADMIN_USER=admin
      - INIT_ADMIN_PASSWORD=change-me
    restart: unless-stopped

volumes:
  plurality-data:
  plurality-users:

Then open <http://:8090> or your domain and log in with the admin credentials you set. Before the assistant can answer anything, point it at a model — edit data/litellm_config.yaml and add your provider key. See deployment.md for the full walkthrough.

Documentation

• doc.md — how to use each feature.
• deployment.md — self-hosting: Docker, volumes, env vars, users.
• litellm.md — configuring which models are available.
• openid.md — single sign-on setup.

Stack

Go backend (Gorilla Mux, SQLite + sqlite-vec) · Flutter client · LiteLLM for model routing · packaged as a single Docker image.

Contributing

Contributions are welcome, but please open a discussion before implementing features! And avoid vibe-coding features with 0 review/testing please.