The tool hub for your local AI.

---
name: send-to-myllm
description: Package web apps and documents so the user can save them into the MyLLM iOS app (MyLLMos home screen) via the share sheet.
---

# Send creations to MyLLM

Use when the user says "send this to MyLLM", "make this a MyLLM app", "save this to MyLLMos", or similar.

## Interactive apps
Produce ONE complete, self-contained HTML document:
- Inline CSS/JS only, no external resources; include <meta name="viewport" content="width=device-width, initial-scale=1">.
- Give it a real <title> — MyLLM uses it as the app's name on the home screen.
- Respect dark mode (prefers-color-scheme).
- For data that must survive relaunches use MyLLM's async bridge, feature-detected: if (window.myllmStorage) { await myllmStorage.setItem(key, stringValue); const v = await myllmStorage.getItem(key); } — localStorage does NOT persist there.
- Optional, also feature-detect: await myllmAsk(prompt) calls the user's own model and resolves to text; await myllmFetch(url, {method, headers, body}) for network; myllmHaptic('light'|'medium'|'success'|'error') for vibration.

Then tell the user: select the whole code block → Share → MyLLM → Save as App in MyLLMos.

## Documents
Produce clean Markdown with a # title heading — MyLLM uses it as the document name.

Then tell the user: select the text → Share → MyLLM → Save to Documents.

{
  "name": "your_tool_name",
  "primaryTheme": "research",
  "tags": ["research", "example"],
  "description": "What the tool does and when the model should call it. Be specific — the LLM uses this to decide whether to invoke it.",
  "timeout": 15,
  "allowsNetwork": true,
  "parameters": {
    "type": "object",
    "properties": {
      "query": {
        "type": "string",
        "description": "Describe each parameter clearly, with examples."
      }
    },
    "required": ["query"]
  },
  "code": "var q = (args.query || '').trim();\nif (!q) return { error: 'Missing query' };\n\n// Use fetch() to call public APIs. Return a plain JS object.\nvar resp = fetch('https://api.example.com/search?q=' + encodeURIComponent(q));\nif (!resp || resp.status < 200 || resp.status >= 300) return { error: 'Request failed' };\n\nvar body = typeof resp.body === 'string' ? JSON.parse(resp.body) : resp.body;\nreturn { query: q, results: body };"
}

• name — snake_case identifier the model uses to call the tool. Must be unique.
• primaryTheme — one of: research, science, finance, news, developer, productivity, fun.
• tags — short string array for filtering. Include primaryTheme as the first tag.
• description — written for the LLM. Explain what it does and when to use it.
• timeout — seconds before the call is aborted. Typical range: 5–30.
• allowsNetwork — true if the code calls fetch(), otherwise false.
• parameters — JSON Schema for the tool's inputs. Use required for mandatory fields.
• code — JavaScript body. Receives args (parsed parameters) and the global fetch(). Must return a JSON-serialisable value (or { error: "..." }).

MyLLM feeds the on-screen controls into specific nodes of your graph. Match these node IDs / input keys and the sliders will drive your workflow:

• Positive prompt → 4.text
• Negative prompt → 5.text
• Width / height → 7.width / 7.height (multiples of 32)
• Frame count → 7.length — must be 4·N + 1 (e.g. 25, 49, 121)
• Seed / steps / CFG → 8.seed / 8.steps / 8.cfg
• Frame rate → 10.frame_rate

If your graph uses different node IDs, the video still generates — it just ignores the sliders and uses the values baked into your JSON. So either match the IDs above, or set your preferred values inside the graph before exporting. (The same screen documents the image-workflow fields too.)

• GitHub — read your repos, browse code, open/comment on issues, search across orgs. The single most useful install for any developer. OAuth on first connect.
• Linear — query, create, and triage issues from your team's board. OAuth.
• Atlassian (Jira / Confluence) — ticket and wiki access. Useful if your work lives in Atlassian.
• Notion — read and write pages in a workspace; great for note-taking and personal knowledge bases. OAuth.
• Stripe — query customers, charges, subscriptions, and invoices. Read-only mode available; recommended for safety on a phone.
• Sentry — pull production errors and issue details. Lets the model answer "why did production break?" without leaving the app.
• Cloudflare — Cloudflare hosts several remote MCP servers (Docs, Workers AI, browser rendering, R2, KV). The Docs server is a low-stakes first install.
• Hugging Face — search models/datasets/spaces and run light inference. Useful for ML and AI questions.
• Context7 (Upstash) — up-to-date documentation for thousands of programming libraries (React, Next.js, Django, FastAPI, Tailwind, and many more). Free, no account required. Perfect companion to the GitHub MCP for coding questions.
• DeepWiki (Cognition) — ask natural-language questions about any indexed public GitHub repository ("how does authentication work in repo X?"). Free, no account required.

No GPU needed here. Unlike running a model (see Connect) or image generation, MCP servers are lightweight — the heavy lifting happens elsewhere. Any always-on machine does the job: an old laptop, a Raspberry Pi, or a small VPS with 1–2 GB RAM is plenty.

• Brave Search — replace hallucinated web answers with real search results. Free API tier; run the server on a Mac or VPS and expose to the iPhone via Tailscale or Cloudflare Tunnel.
• Filesystem / Obsidian / Joplin — point the LLM at your local notes. Most are stdio-only, so run them behind a stdio → HTTP bridge such as mcp-proxy.

Every install shows a confirmation sheet with the tool's name, description, and whether it needs network access. Nothing is added or executed without your tap. Tools run in MyLLM's sandboxed JavaScript runtime — they cannot read your files, contacts, photos, or other apps' data unless MyLLM itself exposes a specific API for that, and you can revoke any tool from Settings → Tools at any time.

Only if the tool itself calls fetch() to a remote URL — and in that case the install sheet flags it with a "needs network" warning. Tools without allowsNetwork: true are pure on-device computation. The MyLLM app does not phone home and does not have analytics; everything else stays on your device or your own LLM server.

Yes — every tool without allowsNetwork works fully offline. That includes the calculator, algebra solver, statistics summary, prime factorization, JSON/CSV/Base64/URL encoders, markdown → HTML, JWT decode, regex tester, password generator, and more. Network tools only fire if you have connectivity.

A tool is a self-contained JavaScript snippet that lives inside MyLLM and runs on-device. An MCP server is an external program (often hosted by a service like GitHub, Notion, or Linear) that MyLLM connects to over HTTP. Use tools for fast, private, deterministic operations. Use MCP for live access to external services and your own self-hosted resources.

Yes — in MyLLM go to Settings → Tools, tap the tool, and edit its name, description, code, or parameters. You can also export it back to JSON to share, or delete it entirely.

Yes — MyLLM creates images (FLUX.2-klein) and video (Wan 2.2 TI2V-5B) on your own ComfyUI server, with no cloud and no per-image fees. Add the server's URL and token under Settings → Image & Video Generation, tap Test ComfyUI, and generate from the in-app Media screen. You can also paste your own ComfyUI workflow (separate slots for image and video) for a fully custom pipeline. See Image & video generation for the full walkthrough.

Write the JSON object following the spec in the Submit a tool section above, then email it to tools@opticell-limited.com. We test on-device, then publish to the catalogue with credit to you.

Yes — the tool runtime uses Apple's JavaScriptCore framework, which is explicitly sanctioned for in-app interpreted JS. There's no remote code download in the App Store sense: tools are JSON payloads describing parameters and a JS body, installed only after your explicit confirmation, and run only when the on-device LLM decides to call them.