Most AI tools give you one model and one answer. Agentis gives you a team.
Deploy fleets of specialized agents — researchers, coders, analysts, writers, and more — across 12 LLM providers simultaneously. Each agent works its angle, shares findings, and hands off to the next. Watch it unfold live with hexagonal agent nodes, curved edges, real-time thought bubbles, and per-agent token tracking. Open source, provider-agnostic, and built for tasks that are too big for a single prompt.
Agentis is a browser-native multi-agent AI platform. You describe a task — Agentis spawns a coordinated team of specialized AI agents across multiple LLM providers, visualises their live thinking on an animated canvas, and synthesizes everything into one clean answer.
No backend. No Docker. No infra. Clone, npm install, go.
You → "Research the competitive landscape for AI coding tools"
Agentis orchestrator plans:
├── Researcher (claude-sonnet-4-6 · Anthropic) → market sizing, key players
├── Analyst (gemini-2.5-flash · Google) → feature comparison matrix
├── Coder (gpt-4.1-mini · OpenAI) → API/SDK landscape
└── Reviewer (llama-3.3-70b · Groq) → fact-check & critique
↓ ~45 seconds
One comprehensive, synthesized report.
- Live animated canvas — agents spawn as orbiting nodes, spin while working, send visible messages to each other
- 12 LLM providers simultaneously — mix Anthropic, OpenAI, Google, Groq, Mistral, DeepSeek, OpenRouter, Cohere, xAI, Together, Ollama, LM Studio in one run
- Exact model names on every node —
claude-sonnet-4-6,gpt-4.1-mini,gemini-2.5-flash - Smart tier selection — orchestrator assigns
simple → fast model,complex → frontier modelper task - Auto-failover — when a provider goes down mid-task, agents switch to the next available one automatically, zero data loss
- Persistent universe — follow-up questions recall relevant old agents and add new ones; knowledge compounds across turns
- Hexagonal agent nodes — distinct visual identity per agent with role-based color coding
- Curved bezier edges — animated particle flow along curved connections between agents and tools
- Live thought bubbles — active agents display their last output snippet in a glassmorphism overlay in real time
- Token progress bars — visual indicator under each agent node showing output token usage
- Tool call diamonds — web search, LLM calls, and browser actions rendered as animated diamond nodes
- Horizontal timeline — shows every agent's start/end as a color-coded bar
- Tool call markers — overlaid on each agent's track showing exactly when web searches, LLM calls, and browser actions fired
- Live duration counter — active agents show elapsed time in seconds
- Per-agent token counts — real input/output token data captured from Anthropic's SSE stream
- Cost estimation — per-agent and session-total cost calculated from live token counts and model pricing
- Header metrics bar — total tokens and estimated cost (
~$X.XXX) displayed live in the canvas header
- Orchestrator plans agent topology, delegates subtasks, then merges all outputs
- Final answer is a clean direct response — no meta-commentary about which agent said what
- Export as Markdown, plain text, or save key insights to persistent memory
- Providers — configure all 12 providers with live connection testing + model recommendations per complexity tier
- Models — browse all available models with pricing, context window, and availability status
- Memory — IndexedDB-backed persistent memory with importance scoring, decay, export/import
- Migration — one-click OpenClaw → OpenFang migration (auto-detect, YAML→TOML conversion, tool remapping)
| Provider | Best Models | Strength |
|---|---|---|
| Anthropic | Claude Opus 4.6, Sonnet 4.6, Haiku 4.5 | Reasoning, writing, long context |
| OpenAI | GPT-4.1, GPT-4.1 Mini, o4-mini | Code, structured output, tools |
| Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash | 1M context, multimodal | |
| Groq | Llama 3.3 70B, 3.1 8B, Mixtral | Fastest inference on the planet |
| Mistral | Large 2, Small 3.1, Codestral | European data, code generation |
| DeepSeek | V3, R1 Reasoner | Math, logic, best cost/quality ratio |
| OpenRouter | 200+ models | Single API for everything |
| Cohere | Command R+, Command R | Enterprise RAG, retrieval |
| xAI | Grok 3, Grok 3 Mini | Real-time web knowledge |
| Together AI | Llama 405B, Qwen 2.5 72B | Best open-source models |
| Ollama | Any model you pull | Local, private, free |
| LM Studio | Any GGUF model | Local GUI + OpenAI-compatible API |
- Node.js 18+
- At least one LLM provider API key (or Ollama running locally — free)
git clone https://github.com/Dhwanil25/Agentis.git
cd Agentis
npm install
npm run devOpen http://localhost:5173, paste any API key, and launch your first agent team.
# .env.local
VITE_ANTHROPIC_API_KEY=sk-ant-...| Feature | What's needed |
|---|---|
| Web search | Free Tavily API key |
| Browser agent | npm install -g pinchtab && pinchtab server |
| Local models | Ollama or LM Studio running locally |
| Feature | Status | Notes |
|---|---|---|
| Agent Universe (multi-agent) | ✅ Works | Core feature, no extra setup |
| Chat (single agent) | ✅ Works | All 12 providers |
| Workflow Templates | ✅ Works | Pre-built multi-step pipelines |
| Analytics (tokens + cost) | ✅ Works | Stored in IndexedDB |
| Sessions & Memory | ✅ Works | IndexedDB, importance decay, export/import |
| Skills (skills.sh) | ✅ Works | 90K+ skills, live search, role assignment |
| Scheduler | ✅ Works* | *Runs in-browser only — stops if tab is closed |
| Channels (test connectivity) | ✅ Works | Telegram, Discord, Slack, Webhook, and more |
| Browser Agent (Hands) | ⚙️ Requires setup | Needs PinchTab running at localhost:9867 |
| Channels (agent-triggered send) | 🔜 Coming soon | Config and test work; auto-send from agents not yet wired |
| Server-side Scheduler | 🔜 Coming soon | Persistent jobs that run without the browser open |
| Agent approval gates | 🔜 Coming soon | Pause mid-task for human review |
| Visual workflow builder | 🔜 Coming soon | Drag-and-drop pipeline editor |
| Semantic memory search | 🔜 Coming soon | Find memories by meaning, not keywords |
| Cross-device memory sync | 🔜 Coming soon | Sync without a server via export token |
┌──────────────────────────────────────────────────────────────────┐
│ Browser (React + Vite) │
│ │
│ UniversePage ──────► multiAgentEngine.ts │
│ │ │ │
│ │ ┌──────▼───────┐ │
│ │ │ Orchestrator │ claude-sonnet-4-6 │
│ │ └──────┬───────┘ │
│ │ │ JSON plan │
│ │ ┌────────────┼────────────┐ │
│ │ ▼ ▼ ▼ │
│ │ Worker A Worker B Worker C │
│ │ Groq/llama Gemini/flash GPT-4.1-mini │
│ │ researcher analyst coder │
│ │ │ │ │ │
│ │ └────────────┼────────────┘ │
│ │ │ outputs │
│ │ ┌──────▼───────┐ │
│ └───────────────►│ Synthesis │ claude-sonnet-4-6 │
│ └─────────────┘ │
│ │
│ Vite proxy routes → 12 provider APIs (CORS-free streaming) │
│ IndexedDB (Dexie) → memory, analytics, sessions (no server) │
└──────────────────────────────────────────────────────────────────┘
| File | Purpose |
|---|---|
src/lib/multiAgentEngine.ts |
Orchestration, streaming, provider failover, token tracking, synthesis |
src/components/pages/UniversePage.tsx |
Canvas rendering, agent cards, metrics bar, post-analysis panel |
src/components/FlowGraph.tsx |
2D canvas flow graph — hexagonal nodes, bezier edges, thought bubbles |
src/components/TimelinePanel.tsx |
Horizontal timeline — agent bars + tool call markers |
src/components/pages/SettingsPage.tsx |
All 12 providers, model browser, memory, migration UI |
src/lib/db.ts |
IndexedDB schema (Dexie) — memories, analytics, sessions |
src/lib/memory.ts |
Advanced memory — importance scoring, decay, categories, export/import |
vite-plugin-agentis.ts |
Vite middleware: engine daemon, migration endpoints |
vite.config.ts |
Proxy routes for all 12 provider APIs |
| Doc | Description |
|---|---|
| Getting Started | Install, first run, quick start |
| Providers | All 12 LLM providers, keys, local models |
| Universe | Multi-agent system, roles, visualization |
| Memory | IndexedDB memory, decay, export/import |
| Skills | skills.sh integration, install, assign to roles |
| Workflows | Template system, custom templates |
| Channels | Messaging integrations, coming soon features |
| Scheduler | Recurring tasks, limitations |
| Architecture | Codebase structure, data flow, contributing |
agentis/
├── public/
│ ├── favicon.png # App icon
│ └── agentis-hero.png # Hero image
├── src/
│ ├── components/
│ │ ├── pages/
│ │ │ ├── UniversePage.tsx # Multi-agent canvas + controls
│ │ │ ├── SettingsPage.tsx # Providers, models, memory, migration
│ │ │ ├── ChatPage.tsx # Single-agent chat
│ │ │ ├── WorkflowsPage.tsx
│ │ │ ├── SchedulerPage.tsx
│ │ │ └── ... # Analytics, Logs, Sessions, etc.
│ │ └── Sidebar.tsx
│ ├── lib/
│ │ ├── multiAgentEngine.ts # Core engine
│ │ ├── db.ts # IndexedDB schema (Dexie)
│ │ ├── memory.ts # Advanced memory layer
│ │ └── claude.ts # Anthropic streaming
│ └── hooks/
│ └── useAgent.ts # Agent state machine
├── vite-plugin-agentis.ts # Engine plugin + migration endpoints
├── vite.config.ts # 12 provider proxies
└── index.html
Go to Settings → Migration and click Auto-Detect OpenClaw.
Agentis scans ~/.openclaw on disk and shows you exactly what it found. Click Migrate Now to:
- Convert
agent.yaml→agent.tomlwith proper capabilities - Remap tools:
read_file → file_read,execute_command → shell_exec,http_request → http_fetch, etc. - Merge channel configs into
config.toml - Copy workspace files and memory data to
~/.openfang
Or use Enter Path Manually to specify a custom source directory.
- Agent Flow view — hexagonal nodes, bezier edges, thought bubbles
- Timeline panel — per-agent activity bars + tool call markers
- Token & cost tracking — live per-agent token counts and cost estimation
- Advanced memory — IndexedDB, importance scoring, decay, export/import
- Skills system — skills.sh integration, 90K+ skills, role assignments
- Scheduler — recurring agent tasks
- Channels — 15 messaging integrations with connection testing
- Agent-triggered channel messaging (Slack/Discord/Telegram on task complete)
- Server-side persistent scheduler (runs without browser open)
- Agent approval gates — pause for human review mid-task
- Visual workflow builder — drag-and-drop pipeline editor
- Semantic memory search
- Shareable agent universes (export + replay)
- Native desktop app (Tauri)
PRs welcome. Please open an issue first for large changes.
npm run dev # start dev server
npm run build # production build
npm run lint # ESLint check- Agent Flow view with hexagonal nodes, quadratic bezier curved edges, and animated particles
- Live thought bubbles — active agents show real-time output snippets on the canvas
- Token progress bars under each agent node
- Timeline panel — horizontal activity bars with tool call markers per agent
- Per-agent token & cost tracking via Anthropic SSE usage events
- Header metrics bar — total tokens and estimated session cost shown live
- Claude PR review workflow with automated code review on every pull request
- Multi-provider Agent Universe (12 providers, exact model names on canvas)
- Provider auto-failover — silent switch when a provider goes down
- Persistent universe — follow-up questions extend the existing agent graph
- Model recommendations in Settings → Providers
- Memory viewer in Settings → Memory
- OpenClaw migration with real filesystem scanning + YAML→TOML conversion
- 9 new Vite proxy routes for all external providers
- New favicon and hero image
- Streaming pipeline agent with Claude API
- Skill-based persona orchestration
- Sidebar navigation, Overview, Analytics, Logs pages
- Three.js landing page
MIT © Dhwanil25
Your next big idea deserves more than one model. Give it a team. If you find Agentis useful, please ⭐ star the repo — it helps a lot!
Built with React · TypeScript · Vite