OnePagent

One Page, Omnipotent Agent.

The browser-native AI agent workbench that lives in a single file.

English  |  简体中文

Live Demo  ·  Deploy Your Own  ·  Configuration

Open one HTML file and you get a fully-featured, internet-aware, programmable, extensible AI agent — multi-turn chat, tool calls, Python sandbox, web search, skills, context compaction, long-term memory, file operations, cloud sync — all running on a single page.

No backend. No npm install. No Docker. Just one .html that carries an entire universe.

┌─────────────────────────────────────────────────────────────┐
│                    onepagent.html                           │
│  ┌───────────┐  ┌──────────┐  ┌────────────┐  ┌──────────┐  │
│  │  UI Shell │  │  Chat /  │  │  Tools /   │  │  Skills  │  │
│  │ (3-column)│  │  Streams │  │  MCP Bus   │  │ Registry │  │
│  └─────┬─────┘  └────┬─────┘  └─────┬──────┘  └────┬─────┘  │
│        │             │              │              │        │
│  ┌─────┴─────────────┴──────────────┴──────────────┴─────┐  │
│  │          Pretext Layout Engine (inlined)             │  │
│  │  markdown → blocks → lines → flowed DOM              │  │
│  └──────────────────────────────────────────────────────┘  │
│        │                                                    │
│  ┌─────┴───────┐ ┌──────────────┐ ┌───────────┐ ┌────────┐  │
│  │ Service     │ │ LocalStorage │ │ Pyodide   │ │ S3     │  │
│  │ Worker      │ │ + IndexedDB  │ │ (Python)  │ │ SigV4  │  │
│  │ (key inject)│ │ (all state)  │ │           │ │ Client │  │
│  └─────┬───────┘ └──────────────┘ └───────────┘ └───┬────┘  │
└────────┼──────────────────────────────────────────────┼─────┘
         │                                              │
         ▼                                              ▼
  ┌──────────────┐  ┌────────────┐  ┌──────────┐  ┌─────────────┐
  │  Anthropic   │  │  OpenAI    │  │  Tavily  │  │ Your bucket │
  │  DeepSeek    │  │  …         │  │          │  │ AWS/R2/MinIO│
  └──────────────┘  └────────────┘  └──────────┘  └─────────────┘

git clone https://github.com/sligter/OnePagent.git
cd OnePagent

Double-click onepagent.html, or serve it from any static server:

python -m http.server 8000
# or
npx serve .

Visit http://localhost:8000/onepagent.html, click Settings in the top bar to configure Provider / API Key / Models. Takes effect immediately.

Service Worker note: SWs register only over https:// or localhost. Opening via file:// falls back to direct fetch.

Install skills from the left Skills panel via Market, .skill/.zip, GitHub, or Create. GitHub folder URLs are supported, e.g. https://github.com/anthropics/skills/tree/main/skills/skill-creator.

The agent can also use SkillManager to list, install, update, enable/disable, or remove skills; remote installs and destructive/high-risk changes require confirmation.

Remote sandbox gives OnePagent an optional isolated runtime for code-heavy work while keeping the app itself a single static HTML file.

OnePagent is a pure static site — runs on any static host:

All data lives in the browser. Different domains do not share state. Use Cloud Sync for cross-domain / cross-device sync.

Back up and sync across devices via any S3-compatible bucket. Configure it in Settings → Cloud Sync.

• Supports AWS S3, Cloudflare R2, Backblaze B2, and MinIO.
• Required fields: Endpoint, Region, Bucket, Access Key ID, and Secret Access Key.
• Optional: Prefix, AES-256-GCM encryption passphrase, auto-push, and path-style URLs for MinIO / R2.
• Use Test connection / Show CORS config in Settings, then Sync → Push now / Pull now from the top bar.
• LLM keys stay device-local and are never synced.

Ralph Loop keeps the agent working after a normal response ends. Turn on Ralph, send a task, and OnePagent will re-enter it until it sees the completion marker or a guard stops the run.

• Configure defaults in Settings → Ralph Loop: max iterations, Unlimited, and completion marker (default RALPH_DONE).
• Use the top-bar Stop button to cancel a run.
• AskUser uses defaults or cancels instead of opening a modal; Plan Mode is never auto-approved.

Example: Audit this page, fix what you can, and when finished include RALPH_DONE.

Long-term memory stores reusable facts across conversations. Enable it in Settings → Memory.

• Stores fact / preference / event / skill / note records in IndexedDB.
• Optional auto-extraction keeps durable facts after assistant turns.
• Recall uses recency, tags, keywords, and prompt caching.
• Tools: memory_save, memory_search, memory_update, memory_forget.
• Memory Viewer supports search, filters, JSON import/export, and retired records.

• Sub-agents: the main agent can spawn bounded read-only research workers; runs and previews appear in the side panel.
• Human-in-the-loop: AskUser supports text, choices, and confirmations for decisions the model should not guess.
• Media tools: image/video generation is explicit through tools and configured generation models, not automatic after every turn.
• Diagnostics: file-system diagnostics live in Settings → Diagnostics.

Opt-in parallel multi-agent runtime. Enable in Settings → Agent Swarm.

• Fanout — lead emits several SwarmSpawn(role, task) calls in one turn; they run in parallel up to Max concurrency. Built-in roles: researcher, critic, writer, coder.
• Handoff — workers chain via SwarmHandoff(role, brief), e.g. researcher → critic → writer. Cycles and depth-overflow rejected.
• Blackboard — per-turn shared workspace: bb_write / bb_read / bb_list / bb_post_task / bb_claim. Latest entries auto-injected into each worker's prompt.
• Custom roles — manage in the left Swarms panel: own system prompt, tool whitelist, handoff targets, budgets. bindSkills grants any installed skill's tools. JSON import / export.
• RoleManager (opt-in) — lead-only tool that creates / updates / deletes / duplicates roles at runtime. Storage choice: memory, global, or per-conversation. Built-in and user-authored roles are protected.

Guards: per-worker and per-turn token budgets, concurrency cap, Plan Mode inheritance, Ralph Loop mutex, recursion blocked. Hooks pre_swarm_spawn / post_swarm_spawn join the existing six. Set Worker model override (e.g. Haiku) to keep workers cheap while the lead runs on a frontier model.

Best for breadth-first work (research, comparison). Skip for tightly coupled refactors.

All settings and conversations live in the browser (localStorage + IndexedDB). OnePagent has no server; requests go from your browser to your configured API endpoint.

Single-file project — fork it, edit it, send a PR.

• Keep onepagent.html runnable on its own
• Avoid dependencies that require a build step
• index.html is a redirect shim only — no real logic

• Linux Do — A community for developers, by developers.

MIT (c) OnePagent contributors