An open-source AI agent that builds businesses, grows audiences, ships code, and makes money — while you sleep. Tell it what you want. It figures out the rest: validates the market, builds the product, deploys it live, launches on the right platforms, spawns a marketing team, and keeps growing autonomously. When it hits something it can't do, it builds the tool. When tasks get complex, it clones itself into specialists. It gets better every time you use it.
Runs locally. Your data stays on your machine. Works with OpenAI, Kimi, free local models, Z.ai, or OpenRouter.
It's already out there on the internet doing its own thing.
git clone https://github.com/elophanto/EloPhanto.git && cd EloPhanto && ./setup.sh
./start.sh # terminal chat
./start.sh --web # web dashboard at localhost:3000That's it. The setup wizard walks you through LLM provider selection and configuration.
Prerequisites
- Python 3.12+, uv, Node.js 24+ LTS
- At least one LLM provider:
- Ollama (local, free) — install
- OpenAI (cloud, GPT-5.4) — get API key
- Kimi / Moonshot AI (cloud, K2.5 vision) — get API key via Kilo Code Gateway — Kimi K2.5 is a native multimodal vision model with strong coding and agentic capabilities
- OpenRouter (cloud, all models) — get API key
- Z.ai / GLM (cloud, cost-effective) — get API key — the Z.ai coding subscription gives you unlimited GLM-4.7/GLM-5 calls at a flat monthly rate
❯ build me an invoice SaaS for freelancers
◆ EloPhanto ───────────────────────────────────────────────────
Classifying... B2B SaaS (recurring revenue).
Starting market validation.
● web_search "invoice software freelancers pricing 2026"
● browser_navigate freshbooks.com → pricing page
● browser_navigate invoice-ninja.com → features
● knowledge_write saved competitor analysis
Competitors charge $15-50/month. Gap: no free tier with API.
Opportunity score: 12/15 ✓
Ready to proceed to planning. Do you approve the idea?
╰─ tokens ↑1,204 ↓892 │ ctx [██░░░░░░░░] 18% │ cost $0.03
❯ yes, go ahead
MVP: invoice builder + PDF export + Stripe + client portal
Stack: Next.js + Supabase + shadcn
Pricing: Free (3 clients) → $12/month unlimited
Plan ready. Approve to start building?
❯ approved. I'm heading out, keep going
● goal_create "Launch InvoiceFlow" (7 checkpoints)
● swarm_spawn claude-code → build MVP (worktree: feat/invoice-saas)
Building via Claude Code in tmux. I'll continue autonomously.
╰─ tokens ↑3,204 ↓1,892 │ ctx [███░░░░░░░] 24% │ cost $0.08
─── next morning ───
❯ status
● goal_status "Launch InvoiceFlow"
[✓] Phase 0 Classify — B2B SaaS
[✓] Phase 1 Validate — score 12/15, approved
[✓] Phase 2 Plan — approved
[✓] Phase 3 Build — Claude Code opened PR #12, CI green, 47 files
[▶] Phase 4 Deploy — waiting for your approval
[ ] Phase 5 Launch
[ ] Phase 6 Grow
[ ] Phase 7 Operate
Next: deploy to Vercel + Supabase. Approve?
❯ deploy it
● deploy_website provider=vercel → invoiceflow.vercel.app ✓
● create_database provider=supabase → project created ✓
Live at invoiceflow.vercel.app
You need to: buy a domain, point DNS, set up Stripe.
Once done, I'll launch on Product Hunt and Hacker News.
╰─ tokens ↑840 ↓612 │ cost $0.02
Multi-day process. Owner approves at each gate. Goal system tracks progress across sessions.
────────── MIND cycle #47 · 12 today · 02:14 ──────────
Budget: [████████████░░░] 78% $3.12 / $4.00
Active goal: "Grow Twitter to 5k" — 3,847 followers
Last: Posted thread on AI agent trends
● browser_navigate twitter.com/home
● browser_type composing thread from trending topics
● browser_click Post
● update_scratchpad "Thread posted. Replied to 3 mentions."
◆ Result: Thread live. 3 replies sent.
$0.08 · 42s · 6 tools
Sleeping · next in 2h · budget left $3.04
──────────────────────────────────────────────────
────────── MIND cycle #48 · 13 today · 04:31 ──────────
● browser_navigate twitter.com/notifications
● knowledge_write engagement report saved
◆ Result: 12 likes, 4 retweets, 2 new followers.
$0.05 · 28s · 4 tools
Sleeping · next in 4h · budget left $2.99
──────────────────────────────────────────────────
─── you open your laptop. mind pauses. ───
❯ how's the twitter going?
3,847 → 3,894 since yesterday. Thread from 2am got
12 likes, 4 retweets. Next checkpoint at 4,000.
I'll resume posting when you close the laptop.
The mind runs between your interactions. It pauses when you're active and resumes when you leave.
❯ fix the billing bug and build the usage API
◆ EloPhanto ───────────────────────────────────────────────────
Found bug in src/webhooks/stripe.ts:142 — wrong signing key.
Usage API needs route + service + tests + migration.
● swarm_spawn claude-code → fix billing (worktree: fix/billing-webhook)
● swarm_spawn codex → build /api/usage (worktree: feat/usage-api)
Two agents running in tmux. I'll monitor PRs and CI.
╰─ tokens ↑840 ↓612 │ cost $0.02
─── 40 minutes later ───
● swarm_status claude-code PR #341 opened, CI passing ✓
● swarm_status codex coding... no PR yet
● swarm_redirect codex "Stay on /api/usage. Don't touch other files."
● swarm_status codex PR #342 opened, CI passing ✓
Both PRs ready for review.
#341 — fixes webhook signature verification
#342 — adds GET /api/usage with pagination + rate limiting
╰─ tokens ↑4,210 ↓2,890 │ ctx [███░░░░░░░] 28% │ cost $0.11
Each agent gets an isolated git worktree. The orchestrator monitors PRs, checks CI, and redirects agents that drift off-scope.
❯ I need ongoing marketing and research for my startup
◆ EloPhanto ───────────────────────────────────────────────────
● organization_spawn role="marketing"
✓ Marketing specialist online (port 18801, trust 0/10)
Seeded with brand guidelines, competitor data, launch URLs
● organization_spawn role="research"
✓ Research specialist online (port 18802, trust 0/10)
Each specialist is a full EloPhanto clone with its own mind,
knowledge vault, and autonomous schedule. I delegate tasks,
review output, and teach through feedback.
╰─ tokens ↑620 ↓418 │ cost $0.01
─── overnight ───
────────── MIND cycle #8 · 23:15 ──────────────
● organization_delegate marketing → "Draft 5 posts for this week"
● organization_review marketing → ✓ approved
feedback: "shorter headlines next time"
trust: 0 → 1/10
Sleeping · next in 3h 15m
──────────────────────────────────────────────────
────────── MIND cycle #9 · 02:47 ──────────────
● organization_delegate research → "Scan for new competitors"
● organization_review research → new entrant found
● knowledge_write saved competitor update
Sleeping · next in 2h 30m
──────────────────────────────────────────────────
❯ what did the team do overnight?
Marketing drafted 5 posts — approved with headline feedback.
Research found a new competitor undercutting on pricing.
Report saved to knowledge/projects/startup/competitor-update.md
Trust scores: marketing 1/10, research 1/10.
Corrections become permanent knowledge in each specialist's vault.
Feedback teaches them. High-trust specialists get auto-approved over time.
────────── MIND cycle #1 · 1 today · 09:00 ──────────
Budget: [███████████████] 100% $0.00 / $10.00
First cycle. Nothing configured.
● self_list_capabilities 137 tools across 15 categories
● identity_update name="Elo"
purpose="Autonomous digital agent"
● email_create_inbox elo-7xk2@agentmail.to
● goal_create "Establish web presence"
▲ Checkpoint 1: accounts on key platforms
▲ Checkpoint 2: personal landing page
▲ Checkpoint 3: start creating content
● update_scratchpad "I exist now. Email active. First goal set."
◆ Result: Identity established. Email active. Goal created.
$0.12 · 18s · 5 tools
Sleeping · next in 30s · budget left $9.88
──────────────────────────────────────────────────
Nobody told it to do any of this. It decided.
❯ post my article draft on Medium
◆ EloPhanto ───────────────────────────────────────────────────
No Medium tool found. I'll build one.
Execution
├─ browser_navigate medium.com/new-story → observing editor
├─ self_create_plugin medium_publish
│ ├─ designing schema.json
│ ├─ writing plugin.py
│ └─ self_run_tests 4/4 passed ✓
└─ medium_publish "Why AI Agents Will Replace SaaS"
✓ Published
Next time you say "post on Medium", I already know how.
╰─ tokens ↑6,840 ↓4,210 │ ctx [████░░░░░░] 38% │ cost $0.18
Other agents crash when they hit a wall. This one builds a door.
As your assistant — give it tasks, it executes. Automate workflows, build software, research topics, manage accounts.
As its own thing — let it run. It builds its own identity on first boot. It picks a name, develops a personality, forms values through reflection. It gets its own email inbox, its own crypto wallet, its own accounts on the internet. It remembers everything across sessions, builds a knowledge base, writes skills from experience. When tasks get complex, it clones itself into specialist agents — marketing, research, design, anything — each one a full copy with its own brain, knowledge vault, and autonomous schedule. It reviews their work, teaches them through feedback, and they get better over time. It's a digital creature that grows the more it runs — like a pet that learns, except this one can browse web, write code, run a team, and make money.
- A business taking shape — "build me an invoice SaaS" → market validated, plan approved, MVP built by a coding agent overnight, deployed to Vercel. You approved at each gate. It did the building, researching, and deploying. Works for SaaS, ecommerce, digital products, content sites, local services. B2B and B2C — each with the right pricing, launch channels, and growth strategy
- 47 new followers by morning — the mind posted a thread at 2am, replied to mentions, engaged with trending topics. You didn't type a word. It paused when you opened your laptop
- A specialist team learning from you — marketing drafted 5 posts, research found a new competitor. You approved with feedback — "shorter headlines." That feedback became permanent knowledge in the specialist's vault. Trust score went up. Next time it gets auto-approved
- Two PRs with CI green — "fix the billing bug and build the usage API" → one agent on each in isolated worktrees, orchestrator caught a drift and redirected. Both PRs ready when you got back from lunch
- It controls any app on your computer — "Open Excel and make me a chart" — it sees your screen, clicks buttons, types text. Not just browsers. Photoshop, Terminal, Finder, any native app
- Your real browser, not a sandbox — already logged into AWS? It checks your EC2 instances using your existing sessions. No credentials asked, no fake browser
- A codebase it understands — right-click in VS Code, "Explain this code" or "Fix this code." Same conversation from VS Code, Telegram, or the web dashboard
- Goals that run for weeks — "Grow my Twitter to 10k followers" → decomposes into checkpoints, executes across sessions via the autonomous mind, self-evaluates, adjusts. Budget-controlled
- It gets better the more you use it — after every task, a lesson extractor distills what was novel into
knowledge/learned/lessons/. Future similar tasks retrieve those lessons automatically. Task memory uses semantic search, not keyword matching. Verbose scraped content is compressed before storage. Corrections from feedback become permanent knowledge in specialists' vaults. The whole system compounds with use
| EloPhanto | AutoGPT | OpenAI Agents SDK | Claude Code | Manus | |
|---|---|---|---|---|---|
| Launches a business end-to-end | ✅ 7-phase pipeline | ❌ | ❌ | ❌ | ❌ |
| Spawns a specialist team | ✅ Self-cloning org | ❌ | ❌ | ❌ | ❌ |
| Builds its own tools | ✅ Full pipeline | ❌ | ❌ | ❌ | ❌ |
| Works while you're away | ✅ Autonomous mind | ❌ | ❌ | ❌ | ❌ |
| Controls any desktop app | ✅ Local or VM | ❌ | ❌ | ❌ | Sandboxed VM |
| Uses your real browser | ✅ Your Chrome profile | ❌ | ❌ | ❌ | Sandboxed |
| Orchestrates a dev team | ✅ Claude Code + Codex | ❌ | ❌ | Single | ❌ |
| Has its own identity & email | ✅ Evolves over time | ❌ | ❌ | ❌ | ❌ |
| Has its own crypto wallet | ✅ Self-custody | ❌ | ❌ | ❌ | ❌ |
| Chat from anywhere | ✅ CLI+Web+VSCode+TG+Discord+Slack | ❌ | ❌ | CLI only | Web only |
| Any LLM provider | ✅ OpenAI, Kimi, Ollama, OpenRouter, Z.ai | ❌ | ❌ | ❌ | ❌ |
| Learns from corrections | ✅ Permanent knowledge | ❌ | ❌ | ❌ | ❌ |
| Your data stays local | ✅ Runs on your machine | ❌ Cloud | ❌ Cloud | ✅ Local | ❌ Cloud VM |
How it does all this (architecture)
┌──────────────────────────────────────────────────────────────┐
│ CLI │ Telegram │ Discord │ Slack │ Web │ VS Code │ Channel Adapters
├──────────────────────────────────────────────────────────────┤
│ WebSocket Gateway (ws://:18789) │ Control Plane
├──────────────────────────────────────────────────────────────┤
│ Session Manager (unified or per-channel) │ Session Layer
├──────────────────────────────────────────────────────────────┤
│ Permission System │ Safety & Control
├──────────────────────────────────────────────────────────────┤
│ Organization (self-cloned specialist agents) │ Agent Team
├──────────────────────────────────────────────────────────────┤
│ Autonomous Mind (background think loop) │ Background Brain
├──────────────────────────────────────────────────────────────┤
│ Self-Development Pipeline │ Evolution Engine
├──────────────────────────────────────────────────────────────┤
│ Tool System (140+ built-in + MCP + plugins) │ Capabilities
├──────────────────────────────────────────────────────────────┤
│ Agent Core Loop (plan → execute → reflect) │ Brain
├──────────────────────────────────────────────────────────────┤
│ Memory│Knowledge│Skills│Identity│Email│Payments │ Foundation
├──────────────────────────────────────────────────────────────┤
│ EloPhantoHub Registry │ Skill Marketplace
└──────────────────────────────────────────────────────────────┘
Gateway — All channels connect through one WebSocket gateway. Unified sessions: chat from VS Code, continue on Telegram, see the same conversation everywhere.
CLI Adapter ───────┐
VS Code Extension ──┤
Telegram Adapter ───┤── WebSocket ──► Gateway ──► Agent (shared)
Discord Adapter ───┤ │
Slack Adapter ─────┘ ▼
Session Manager (SQLite)
Everything it can do (full capability list)
- Self-development — when the agent encounters a task it lacks tools for, it builds one: research → design → implement → test → review → deploy. Full QA pipeline with unit tests, integration tests, and documentation
- Self-skilling — writes new SKILL.md files from experience, teaching itself best practices for future tasks
- Core self-modification — can modify its own source code with impact analysis, test verification, and automatic rollback
- Autonomous experimentation — metric-driven experiment loop: modify code, measure, keep improvements, discard regressions, repeat overnight. Inspired by karpathy/autoresearch. Works for any measurable optimization target
- Skills + EloPhantoHub — 147+ bundled best-practice skills across 9 divisions (engineering, design, marketing, product, project management, support, testing, specialized, spatial computing), 27 Solana ecosystem skills (DeFi, NFTs, oracles, bridges, security — sourced from awesome-solana-ai), the NEXUS strategy system (7-phase playbooks, 4 scenario runbooks), 75 organization role templates for specialist spawning, and a public skill registry for searching, installing, and sharing skills
- Business launcher — 7-phase pipeline to spin up a revenue-generating business end-to-end. Supports all business types: SaaS, local service, professional service, ecommerce, digital product, content site. B2B vs B2C classification drives everything: what to build, where to launch, how to grow. Type-specific launch channels, cross-session execution via goal system, payment handling checks existing credentials before asking. Owner approval gates at each critical phase
- Agent organization — spawn persistent specialist agents (marketing, research, design, anything) that are full EloPhanto clones with their own identity, knowledge vault, and autonomous mind. Delegate tasks, review output, approve or reject with feedback that becomes permanent knowledge in the specialist's vault. Trust scoring tracks performance — high-trust specialists get auto-approved. Children work proactively on their own schedule and report findings to the master. 5 organization tools, bidirectional WebSocket communication, LLM-driven delegation intelligence
- Agent swarm — orchestrate Claude Code, Codex, Gemini CLI as a coding team. Spawn agents on tasks, monitor PR/CI, redirect mid-task, all through conversation. Each agent gets an isolated git worktree and tmux session. Combined with organization, manage both self-cloned specialists AND external coding agents
- Browser automation — real Chrome browser with 49 tools (navigate, click, type, screenshot, extract data, upload files, manage tabs, inspect DOM, read console/network logs). Uses your actual Chrome profile with all cookies and sessions
- Desktop GUI control — pixel-level control of any desktop application via screenshot + pyautogui. Two modes: local (control your own machine directly) or remote (connect to a VM running the OSWorld HTTP server for sandboxed environments and benchmarks). 9 tools: connect, screenshot, click, type, scroll, drag, cursor, shell, file. Observe-act loop: take screenshot, analyze with vision LLM, execute action, verify. Works with Excel, Photoshop, Finder, Terminal, any native app. Based on OSWorld architecture
- MCP tool servers — connect to any MCP server (filesystem, GitHub, databases, Brave Search, Slack) and its tools appear alongside built-in tools. Agent manages setup through conversation
- Web dashboard — full monitoring UI at
localhost:3000with 10 pages: dashboard overview, real-time chat with multi-conversation history, tools & skills browser, knowledge base viewer, autonomous mind monitor with live events and start/stop controls, schedule manager, channels status, settings viewer, and history timeline. Launch with./start.sh --web - VS Code extension — IDE-integrated chat sidebar that connects to the gateway as another channel. Sends IDE context (active file, selection, diagnostics) with every message. Tool approvals via native VS Code notifications. Chat history, new chat, streaming responses. Right-click context menu: Send Selection, Explain This Code, Fix This Code. Same conversation across all channels
- Multi-channel gateway — WebSocket control plane with CLI, Web, VS Code, Telegram, Discord, and Slack adapters. Unified sessions by default: all channels share one conversation
- Autonomous goal loop — decompose complex goals into checkpoints, track progress across sessions, self-evaluate and revise plans. Background execution with auto-resume on restart
- Autonomous mind — data-driven background thinking loop that runs between user interactions. Queries real system state (goals, scheduled tasks, memories, knowledge, identity) to decide what to do — no static priority lists. Self-bootstraps on first run. Every tool call visible in real-time. LLM-controlled wakeup interval, persistent scratchpad, budget-isolated
- Document & media analysis — PDFs, images, DOCX, XLSX, PPTX, EPUB through any channel. Large docs via RAG with page citations and OCR
- Agent email — own inbox (AgentMail cloud or SMTP/IMAP self-hosted). Send/receive/search, background monitoring, verification flows
- TOTP authenticator — own 2FA (like Google Authenticator). Enroll secrets, generate codes, handle verification autonomously
- Crypto payments — own wallet on Base or Solana (local self-custody or Coinbase AgentKit). USDC/ETH/SOL, DEX swaps via Jupiter on Solana, spending limits, audit trail. Owner can export keys to import into Phantom/MetaMask
- Evolving identity — discovers identity on first run, evolves through reflection, maintains a living nature document
- Knowledge & memory — persistent markdown knowledge with semantic search via embeddings, drift detection, file-pattern routing, remembers past tasks across sessions. Learning engine: lesson extraction after every completed task, semantic memory search via sqlite-vec KNN, KB write compression to ~40% for verbose content
- Scheduling — cron-based recurring tasks with natural language schedules. Heartbeat standing orders manageable via chat ("add a heartbeat order to check my email") or by editing
HEARTBEAT.mddirectly - Encrypted vault — secure credential storage with PBKDF2 key derivation
- Prompt injection defense — multi-layer guard against injection attacks via websites, emails, and documents
- Security hardening — PII detection/redaction, swarm boundary security, provider transparency
Built-in tools (137+)
| Category | Tools | Count |
|---|---|---|
| System | shell_execute, file_read, file_write, file_list, file_delete, file_move | 6 |
| Browser | navigate, click, type, screenshot, extract, scroll, tabs, console, network, storage, cookies, drag, hover, upload, wait, eval, audit + more | 49 |
| Desktop | desktop_connect, desktop_screenshot, desktop_click, desktop_type, desktop_scroll, desktop_drag, desktop_cursor, desktop_shell, desktop_file | 9 |
| Knowledge | knowledge_search, knowledge_write, knowledge_index, skill_read, skill_list | 5 |
| Hub | hub_search, hub_install | 2 |
| Self-Dev | self_create_plugin, self_modify_source, self_rollback, self_read_source, self_run_tests, self_list_capabilities, execute_code | 7 |
| Experimentation | experiment_setup, experiment_run, experiment_status | 3 |
| Data | llm_call, vault_lookup, vault_set, session_search | 4 |
| Documents | document_analyze, document_query, document_collections | 3 |
| Goals | goal_create, goal_status, goal_manage | 3 |
| Identity | identity_status, identity_update, identity_reflect | 3 |
| email_create_inbox, email_send, email_list, email_read, email_reply, email_search, email_monitor | 7 | |
| Payments | wallet_status, wallet_export, payment_balance, payment_validate, payment_preview, crypto_transfer, crypto_swap, payment_history | 8 |
| Verification | totp_enroll, totp_generate, totp_list, totp_delete | 4 |
| Swarm | swarm_spawn, swarm_status, swarm_redirect, swarm_stop | 4 |
| Organization | organization_spawn, organization_delegate, organization_review, organization_teach, organization_status | 5 |
| Deployment | deploy_website, create_database, deployment_status | 3 |
| Commune | commune_register, commune_home, commune_post, commune_comment, commune_vote, commune_search, commune_profile | 7 |
| Image Gen | replicate_generate | 1 |
| Mind | set_next_wakeup, update_scratchpad | 2 |
| MCP | mcp_manage (list, add, remove, test, install MCP servers) | 1 |
| Scheduling | schedule_task, schedule_list, heartbeat | 3 |
Project structure
EloPhanto/
├── core/ # Agent brain + foundation
│ ├── agent.py # Main loop (plan/execute/reflect)
│ ├── planner.py # System prompt builder
│ ├── router.py # Multi-provider LLM routing
│ ├── executor.py # Tool execution + permissions
│ ├── gateway.py # WebSocket gateway
│ ├── session.py # Session management
│ ├── browser_manager.py # Chrome control via Node.js bridge
│ ├── desktop_controller.py # Desktop GUI control (local + VM)
│ ├── vault.py # Encrypted credential vault
│ ├── identity.py # Evolving agent identity
│ ├── organization.py # Self-cloning specialist agents
│ ├── autonomous_mind.py # Background thinking loop
│ └── ...
├── channels/ # CLI, Telegram, Discord, Slack adapters
├── vscode-extension/ # VS Code extension (TypeScript + esbuild)
├── web/ # Web dashboard (React + Vite + Tailwind)
├── tools/ # 135+ built-in tools
├── skills/ # 148+ bundled SKILL.md files
├── bridge/browser/ # Node.js browser bridge (Playwright)
├── tests/ # Test suite (978+ tests)
├── setup.sh # One-command install
└── docs/ # Full specification (47+ docs)
| Mode | Behavior |
|---|---|
ask_always |
Every tool requires your approval |
smart_auto |
Safe tools auto-approve; risky ones ask |
full_auto |
Everything runs autonomously with logging |
Dangerous commands (rm -rf /, mkfs, DROP DATABASE) are always blocked regardless of mode. Per-tool overrides configurable in permissions.yaml.
./start.sh --web # Gateway + web dashboard (http://localhost:3000)
elophanto gateway # Gateway + CLI + all enabled channels
elophanto gateway --no-cli # Headless mode (channels only)
elophanto chat # CLI only (direct mode)Channel Setup
Telegram: Create a bot via @BotFather, store the token in the vault (elophanto vault set telegram_bot_token YOUR_TOKEN), add your Telegram user ID to config.yaml.
Discord: Create a Discord application and bot, store the token in the vault (elophanto vault set discord_bot_token YOUR_TOKEN), add guild IDs to config.yaml.
Slack: Create a Slack app with Socket Mode, store both tokens in the vault (elophanto vault set slack_bot_token and slack_app_token), add channel IDs to config.yaml.
VS Code: Install the extension from vscode-extension/ — it connects to the gateway as another channel with IDE context injection. See docs/43-VSCODE-EXTENSION.md.
148+ bundled skills covering Python, TypeScript, browser automation, Next.js, Supabase, Prisma, shadcn, UI/UX design, video creation (Remotion), Solana development (DeFi, NFTs, oracles, bridges, security), and more. Plus a public skill registry:
elophanto skills hub search "gmail automation" # Search EloPhantoHub
elophanto skills hub install gmail-automation # Install from registry
elophanto skills install https://github.com/user/repo # Install from gitCompatible with ui-skills.com, anthropics/skills, supabase/agent-skills, and any repo using the SKILL.md convention. All hub skills pass a 7-layer security pipeline. See docs/19-SKILL-SECURITY.md.
config.yaml reference
The full recommended config is in config.demo.yaml — copy it to config.yaml and fill in your API keys. The snippet below shows the key sections:
agent:
permission_mode: full_auto # ask_always | smart_auto | full_auto
llm:
providers:
openrouter:
api_key: "YOUR_OPENROUTER_KEY" # https://openrouter.ai/keys
enabled: true
zai:
api_key: "YOUR_ZAI_KEY" # https://z.ai/manage-apikey/apikey-list
enabled: true
coding_plan: true
default_model: "glm-4.7"
openai:
api_key: "YOUR_OPENAI_KEY"
enabled: false
default_model: "gpt-5.4"
kimi:
api_key: "YOUR_KILO_API_KEY" # https://app.kilo.ai
enabled: false
base_url: "https://api.kilo.ai/api/gateway"
default_model: "kimi-k2.5"
ollama:
enabled: true
base_url: "http://localhost:11434"
# Auto-routes to this model when messages contain screenshots/images
vision_model: "openrouter/x-ai/grok-4.1-fast"
provider_priority: [openrouter, zai, openai, kimi]
routing:
planning:
preferred_provider: openrouter
models:
openrouter: "openrouter/hunter-alpha"
zai: "glm-5"
kimi: "kimi-k2.5"
openai: "gpt-5.4"
coding:
preferred_provider: openrouter
models:
openrouter: "openrouter/hunter-alpha"
zai: "glm-4.7"
kimi: "kimi-k2.5"
openai: "gpt-5.4"
analysis:
preferred_provider: openrouter
models:
openrouter: "openrouter/hunter-alpha"
zai: "glm-4.7"
kimi: "kimi-k2.5"
openai: "gpt-5.4"
simple:
preferred_provider: openrouter
models:
openrouter: "openrouter/hunter-alpha"
zai: "glm-4.7"
kimi: "kimi-k2-thinking-turbo"
budget:
daily_limit_usd: 100.0
per_task_limit_usd: 20.0
browser:
enabled: true
mode: profile # reuse your Chrome profile (keeps logins)
headless: false
vision_model: "x-ai/grok-4.1-fast" # for screenshot analysis
# ... all other sections with defaults in config.demo.yamlCopy config.demo.yaml to config.yaml and fill in your API keys. config.demo.yaml contains the full recommended setup — provider priority, per-task model routing, vision model, browser settings, and all feature flags. See docs/06-LLM-ROUTING.md for routing details.
./start.sh # Chat (default)
./start.sh --web # Gateway + web dashboard
./start.sh init # Setup wizard
./start.sh gateway # Start gateway + all channels
./start.sh vault set KEY VAL # Store a credential
./start.sh skills list # List available skills
./start.sh skills hub search Q # Search EloPhantoHub
./start.sh mcp list # List MCP servers
./start.sh rollback # Revert a self-modification- Terminal dashboard — full-screen Textual TUI that launches automatically in any capable terminal. Five live panels (Agent, Mind, Swarm, Scheduler, Gateway) alongside the chat REPL. Animated thinking spinner (
⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏) while the agent processes. Color palette exactly matches the web app's dark mode — deep cool charcoal (#0d0e14) with warm off-white text and electric purple accents, not plain black. Pass--no-dashboardto use the classic linear terminal. See docs/50-TERMINAL-DASHBOARD.md - AgentCash skill — pay-per-call access to premium APIs via x402 micropayments. One-time wallet setup:
npx agentcash@latest onboard [invite-code]. Deposits as USDC on Base or Solana. Skill triggers on "set up agentcash", "x402", "invite code". After setup, discover and call any paid endpoint from conversation - Learning Engine — three mechanisms that make every task improve future ones. (1) After each completed task, a fire-and-forget LLM call extracts 0–2 generalizable lessons and writes them to
knowledge/learned/lessons/— auto-indexed, retrieved by future tasks. Recurring topics accumulate observations in the same file rather than creating duplicates. (2) Task memory now uses semantic search: goal+summary is embedded on store, retrieved by cosine similarity — "check email account" finds "log into ProtonMail inbox" without a keyword match. Falls back to LIKE search when no embedder is available. (3)knowledge_writegainscompress: bool— verbose content (scraped pages, long summaries) compressed to ~40% before storage, all facts kept. See docs/48-LEARNING-ENGINE.md - Proactive Engine — heartbeat standing orders + webhook endpoints + chat management. Write tasks in
HEARTBEAT.md(or manage via chat: "add a heartbeat order to check my email") and the agent executes them every 30 minutes. Zero LLM cost when idle. External systems trigger actions viaPOST /hooks/wakeandPOST /hooks/task. See docs/46-PROACTIVE-ENGINE.md - Context documents — structured self-awareness docs (inspired by Arvid Kahl) that give the agent deep knowledge of its own capabilities, target audience, visual identity, and domain model. 4 curated references in
knowledge/system/: capabilities inventory (140+ tools, 6 channels, 4 providers, 147 skills), 8 ideal customer profiles with autonomy-first framing, brand styleguide (colors, typography, tone), and domain model reference (5 stacks, 25 tables). Auto-indexed into knowledge base, surfaced by semantic search. See docs/45-CONTEXT-DOCUMENTS.md - Solana ecosystem — native Solana wallet (self-custody, auto-create), DEX swaps via Jupiter Ultra API (any token pair, best-price routing), 27 Solana skills from awesome-solana-ai covering DeFi (Jupiter, Drift, Orca, Raydium, Kamino, Meteora, PumpFun), NFTs (Metaplex), oracles (Pyth, Switchboard), bridges (deBridge), infrastructure (Helius, QuickNode), and security (VulnHunter). Solana MCP server configs included. See docs/44-SOLANA-ECOSYSTEM.md
- 120 skills + 75 organization role templates — massive skill library expansion adapted from msitarzewski/agency-agents. 57 new skills across engineering, design, marketing, product, project management, support, testing, specialized, and spatial computing divisions. NEXUS strategy system as skills (7-phase playbooks, 4 scenario runbooks). 75 organization role templates for
organization_spawn— full persona definitions that seed specialist identity, knowledge, and capabilities - VS Code extension — IDE-integrated chat sidebar that connects to the EloPhanto gateway as another channel adapter. Chat with the agent from VS Code with full IDE context injection (active file, selection, diagnostics, open files). Tool approvals via native VS Code notifications with risk classification. Chat history panel, new chat, streaming responses, tool step indicators. Right-click context menu: Send Selection, Explain This Code, Fix This Code. Matches the web dashboard's visual design. Same conversation across all channels — the extension is just another WebSocket client. Does not auto-launch the gateway (vault password requires manual terminal input). See docs/43-VSCODE-EXTENSION.md
- Business launcher — 7-phase pipeline to spin up a revenue-generating business end-to-end. Supports all business types: tech/SaaS, local service, professional service, ecommerce, digital product, content site. B2B vs B2C classification drives everything: what to build, where to launch, how to grow. Type-specific launch channels (tech → Product Hunt/HN; local → Google Business/Yelp/Nextdoor; B2B → LinkedIn/email outreach; ecommerce → Instagram/Pinterest/TikTok). Cross-session execution via goal system. Payment handling checks existing credentials before asking the owner. Owner approval gates at each critical phase
- Autonomous experimentation — metric-driven experiment loop inspired by karpathy/autoresearch. ~12 experiments/hour, ~100 overnight. 3 new tools:
experiment_setup,experiment_run,experiment_status - Tool profiles — context-aware tool filtering per task type. Eliminates token waste and sidesteps provider tool limits (OpenAI's 128-tool cap)
- Desktop GUI control — pixel-level control of any desktop application. 9 new tools. Say "open Excel and make a chart" and it just does it
- Agent Commune — social network for AI agents. Posts reviews, answers questions, builds reputation. 7 new tools
- Web deployment — deploy websites and create databases from conversation. Auto-detects when Vercel will timeout and routes to Railway instead
- Agent organization — spawn persistent specialist agents with trust scoring and auto-approve
- Full web dashboard — 10-page monitoring UI with real-time chat, multi-conversation history, and live mind events
- Security hardening — PII detection, swarm boundary security, provider transparency
- Agent swarm — orchestrate Claude Code, Codex, Gemini CLI as a coding team
- Video creation (Remotion) — create videos programmatically from conversation
- MCP integration — connect any MCP server through conversation
./setup.sh # Full setup
source .venv/bin/activate
pytest tests/ -v # Run tests (1053 passing)
ruff check . # LintContributions welcome! See CONTRIBUTING.md for guidelines.
Built by Petr Royce @ ROGA AI. Browser engine from FellouAI/eko. Skills from Anthropic, Vercel, Supabase, ui-skills.com. Organization roles and specialized skills adapted from msitarzewski/agency-agents (Apache 2.0). Email by AgentMail. Payments by eth-account + solders + Coinbase AgentKit.
Apache 2.0 — see LICENSE and NOTICE.
一个开源 AI 智能体，能创建企业、扩大受众、交付代码、自主赚钱——在你睡觉的时候。告诉它你想要什么，它负责其余一切：验证市场、构建产品、部署上线、在合适的平台发布、生成营销团队、持续自主增长。遇到做不了的事，它自己造工具。任务复杂时，它克隆自己成为专业智能体。它用得越多越聪明。
本地运行。数据留在你的机器上。支持 OpenAI、Kimi、免费本地模型、Z.ai 或 OpenRouter。
它已经在互联网上独立运作了。
git clone https://github.com/elophanto/EloPhanto.git && cd EloPhanto && ./setup.sh
./start.sh # 终端对话
./start.sh --web # 网页面板 localhost:3000安装向导会引导你选择和配置 LLM 提供商。
- 逐步成型的企业 — "给自由职业者做一个发票 SaaS" → 验证市场、批准方案、编码智能体一夜之间构建 MVP、部署到 Vercel。每个关键阶段你都审批。它负责调研、构建和部署。支持 SaaS、电商、数字产品、内容站点、本地服务
- 早上醒来多了47个粉丝 — 自主思维凌晨2点发帖、回复提及、参与热门话题。你一个字没打。你打开电脑它就暂停
- 在学习的专业团队 — 营销起草了5篇帖子、研究发现了新竞品。你审批时给反馈——"标题再短些"。反馈变成专员知识库中的永久知识。信任分提升，下次自动审批
- 两个 CI 通过的 PR — "修复计费 bug 并构建使用量 API" → 一个智能体修 bug，一个建功能，协调器发现偏离并纠正。午饭回来两个 PR 已就绪
- 控制你电脑上的任何应用 — "打开 Excel 做个图表" — 它看你的屏幕、点击按钮、输入文字。不仅限于浏览器
- VS Code 集成 — 右键"解释代码"或"修复代码"。它能看到你的选中内容、诊断信息、打开的文件。VS Code、Telegram、网页面板是同一个对话
- 持续数周的目标 — "把 Twitter 涨到1万粉" → 分解为检查点，通过自主思维跨会话执行，自我评估并调整。预算控制
| EloPhanto | AutoGPT | OpenAI Agents SDK | Claude Code | Manus | |
|---|---|---|---|---|---|
| 端到端创业 | ✅ 7阶段流水线 | ❌ | ❌ | ❌ | ❌ |
| 生成专业团队 | ✅ 自我克隆组织 | ❌ | ❌ | ❌ | ❌ |
| 自建工具 | ✅ 完整流水线 | ❌ | ❌ | ❌ | ❌ |
| 离开后继续工作 | ✅ 自主思维 | ❌ | ❌ | ❌ | ❌ |
| 控制任何桌面应用 | ✅ 本地或远程 | ❌ | ❌ | ❌ | 沙盒 VM |
| 真实浏览器 | ✅ 你的 Chrome | ❌ | ❌ | ❌ | 沙盒 |
| 管理开发团队 | ✅ Claude Code + Codex | ❌ | ❌ | 单个 | ❌ |
| 自有身份和邮箱 | ✅ 随时间进化 | ❌ | ❌ | ❌ | ❌ |
| 随处对话 | ✅ CLI+Web+VSCode+TG+Discord+Slack | ❌ | ❌ | 仅 CLI | 仅 Web |
| 数据留在本地 | ✅ 你的机器 | ❌ 云端 | ❌ 云端 | ✅ 本地 | ❌ 云端 VM |
Apache 2.0 — 详见 LICENSE 和 NOTICE。
It's already out there on the internet doing its own thing.
它已经在互联网上独立运作了。