The Personal AI with a Secure Perimeter.

Experience uncompromised intelligence. Clambot leverages a WASM sandbox for a security-first approach to local execution and cloud integration.

A futuristic 3D clam shell made of dark matte metal and glowing emerald green circuits, containing a glowing AI core — representing Clambot's secure sandbox architecture.

[ Functional Core ]

Engineered for Technical Sovereignty

lock

Secure Execution

Isolation is non-negotiable. Every line of code runs in a high-performance WASM sandbox, ensuring zero-access to host files without explicit cryptographic signatures.

send

Telegram Native

Full control via Telegram with real-time approval modals and typing indicators. No new app needed.

terminal

Technical CLI

Optimized for power users. Interactive REPL mode and single-turn commands via our custom shell wrapper.

memory

Intelligent Memory

Long-term recall through vector-embedding of your interaction history. Instantly searchable and contextually aware.

// Retrieval Sequence
> Search memory...
> Results: 14 matches

[ How It Works ]

The Clam Lifecycle

From user request to sandboxed execution — every step is explicit, auditable, and secure.

chat

User Request

"What is the price of BTC?"
recycling

Pre-selection

Exact match against existing clams — if a clam already handles this task, it's reused instantly at zero LLM cost.

Match found → reuse clam No match → continue
call_split

Selector LLM

A lightweight LLM decides the action:

generate_new select_existing chat
bolt

Generator LLM

Produces a self-contained JavaScript clam:

btc_price.js
async function run(args) {
  const res = await http_request({
    method: "GET",
    url: "https://api.coinbase.com/v2/prices/BTC-USD/spot"
  });
  return JSON.parse(res.content).data;
}
shield

WASM Sandbox Execution

The clam runs in an isolated WASM environment. External calls pass through an explicit approval gate before the Python host dispatches them.

http_request Approval Gate Host Dispatch Result
search

Post-Runtime Analyzer

The output is evaluated and the clam is triaged:

Accept → promote to clams/ for reuse Self-fix → retry with fix instructions (up to 3×) Reject → return error
neurology

Response & Memory

The answer is delivered to the user. In the background, a fire-and-forget process extracts key information into long-term memory for future context.

Built for Developers

Clambot is open-source and architected for extensibility. Join hundreds of contributors building the future of secure personal AI.

--
GitHub Stars
--
Forks
--
Open Issues
WASM
Core Architecture
star Star on GitHub