---

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.

---

Preview

---

Highlights

Capability	Description
Single-file deployment	Drop onepagent.html on any static host or open it locally
Multi-provider LLM	Anthropic / OpenAI / DeepSeek with custom endpoints; BYOK keys are stored locally and sent directly to the configured API endpoint
Reasoning levels	Inline selector with OpenAI none / low / medium / high / xhigh / max tiers (plus Auto)
Long context compaction	Per-model context window with automatic LLM-driven summary compression
Long-term memory	Opt-in persistent facts / preferences / events / skills across sessions — auto-extraction, agent tools, manual CRUD, tag & keyword search
MCP Servers	Paste mcpServers JSON to import (streamable_http / sse), Bearer auth, optional CORS proxy
Plan Mode	Agent investigates with read-only tools, drafts a Markdown plan for approval, then executes
Ralph Loop	Fully unattended continue-until-done mode with marker, max/unlimited iterations, Stop, and no-progress guard
Sub-agents	Delegate bounded read-only research tasks and monitor runs in the side panel
Agent Swarm	Opt-in parallel orchestrator-worker fanout: lead emits multiple SwarmSpawn calls in one turn, role-scoped workers (researcher / critic / writer / coder) run concurrently with token budgets
Human-in-the-loop	Agent can ask for text, choices, or confirmations when a task needs user input
TodoWrite	Agent maintains a visible task list (pending / in-progress / completed)
Hooks	User-defined JS handlers on 6 agent lifecycle events
Python sandbox	Execute Python in-browser via Pyodide
Remote sandbox	Optional isolated runtime for heavier code tasks
Web Search	Tavily integration with basic / advanced depth modes
Skills system	Install .skill / .zip packs, pull from GitHub, create in-page, or let AI manage skills with SkillManager
Conversation management	Multiple sessions, folders, drag-and-drop, IndexedDB persistence, one-click export
Cloud Sync	Incremental sync to S3-compatible buckets with optional AES-256-GCM encryption

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    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  │  │
│  │ (PWA cache) │ │ (all state)  │ │           │ │ Client │  │
│  └─────┬───────┘ └──────────────┘ └───────────┘ └───┬────┘  │
└────────┼──────────────────────────────────────────────┼─────┘
         │                                              │
         ▼                                              ▼
  ┌──────────────┐  ┌────────────┐  ┌──────────┐  ┌─────────────┐
  │  Anthropic   │  │  OpenAI    │  │  Tavily  │  │ Your bucket │
  │  DeepSeek    │  │  …         │  │          │  │ AWS/R2/MinIO│
  └──────────────┘  └────────────┘  └──────────┘  └─────────────┘

---

Getting Started

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: the bundled sw.js is only a PWA/offline cache for the app shell and static assets. LLM and Tavily requests are browser direct fetches with the configured BYOK credentials; the Service Worker does not proxy, inject, or hide API keys.

---

Skills

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

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

---

Deploy

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

Platform	Steps
Vercel	Click button, sign in, Deploy — done in ~10 seconds
Zeabur	Click button, select your fork, auto-detected as static, Generate Domain
Cloudflare Pages	Connect to Git, Framework: None, Build command: empty, Output: /, Deploy
GitHub Pages	Settings, Pages, Source: GitHub Actions, push to deploy

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

---

Cloud 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

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.

---

Memory

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.

---

Agent Workflow

• 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.

---

Agent Swarm

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.

---

Configuration

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

---

Contributing

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

---

Star History

---

Friends

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

---

License

MIT (c) OnePagent contributors

---