Show HN: Anytool – Give your LLM infinite tools with one API

[acoyfellow]

I built inbox.dog and hit diminishing returns stuffing agents with dozens of static tools.

Context got noisy, tokens bloated, iteration slowed.

So I built Anytool: a single meta‑tool your agent calls to generate exactly the tool it needs, on demand.

First run generates + caches the code; subsequent runs are instant.

Each tool comes with its own isolated sandbox, ephemeral filesystem, per‑tool KV state, dedicated SQLite DB, per‑user env (USER_), full npm in a Bun container, and outbound fetch for APIs.

How I built it:

    - Frontend: SvelteKit (Svelte 5)
    - Backend: Cloudflare Workers + Durable Objects
    - Execution: Bun container, full npm, auto‑install + global cache, no bundling
    - Storage: R2 for code/metadata (hash‑based), per‑tool KV + SQLite for state
    - Auth: Better Auth; per‑user encrypted env injected at runtime as USER_*
    - API: Hono + Zod OpenAPI + Scalar docs; MCP‑native surface
    - AI: Effect‑TS retries/self‑healing; version‑pinned imports for determinism

What it does:

    - Natural‑language → real npm tool, executed in isolation
    - Hash caching: first run (AI + first install); next runs reuse cache
    - Per‑tool state (KV + SQLite) for workflows, history
    - Credentials: OPENROUTER_APIKEY and other secrets injected as env.USER_*
    - Retrieval/Augmentation (how it “learns”):
    - Full NPM accesss; indexes package README + types into R2 libraries/; generation retrieves these for better code and fewer tokens.
    - Per‑user similarity: LLM ranks your existing tools for semantic similarity and suggests modify vs create to avoid duplicates. Improves as your library grows.
    - MCP surface (simple on purpose):
        `search_user_tools` — semantic search in your library
        `create_tool` — generate and store new tool
        `execute_tool` — run a tool with input

Try it: AnytoolHQ.com

    — 1,000 free runs/month. Bring your OpenRouter key (OPENROUTER_APIKEY).
    - Works with AI SDK/LangChain/MCP. 
    - One tool → infinite capabilities.

Privacy/Safety:

    - Per‑user encrypted env; no secrets in code
    - Isolated execution; audit/run history
    - Explicit versioned imports; reproducible runs

Would love feedback on:

    - The “one meta‑tool” DX (AI SDK/LangChain/MCP)
    - First‑run latency + cache behavior
    - Env/credential ergonomics (USER_ injection)
    - Templates you want first (QR, CSV, validators, API clients, etc.)

Happy to answer questions.. cheers!

- Jordan Coeyman (x.com/acoyfellow)

[SamTinnerholm]

Cool. Will try this with my agent.

Quick comment: the code on the website isn't using any sort of intellisense or highlighting

[SamTinnerholm]

I can't seem to create tools on https://anytoolhq.com/dashboard. Report back when it works, and I'll give it a try.

[acoyfellow]

Yes, thank you, I know it's not ideal now. It would make a much better impression.

I plan on adding it soon, after smoothing out the internals.