Trytet

Trytet

[](https://github.com/bneb/trytet/releases) [](https://www.npmjs.com/package/trytet-client) [](https://pypi.org/project/trytet-client/)

A WebAssembly sandbox engine that runs AI-generated code without crashing. Agents invoke deterministic Wasm components ("cartridges") inside fuel-bounded sandboxes — infinite loops trap in microseconds instead of hanging until a timeout, memory bombs get capped instead of OOMing.

Installation

# macOS (Apple Silicon)
curl -sL https://github.com/bneb/trytet/releases/latest/download/tet-darwin-arm64.tar.gz | tar xz
./tet doctor
./tet mcp --list-tools

# Docker
docker pull ghcr.io/bneb/trytet:latest
docker run -p 3000:3000 ghcr.io/bneb/trytet:latest

Also available via SDK: npm install trytet-client | pip install trytet-client

Quickstart

Claude Desktop / Cursor — add to claude_desktop_config.json:

{
  "mcpServers": {
    "trytet": {
      "command": "tet",
      "args": ["mcp"]
    }
  }
}

Then tools appear automatically. Run tet mcp --list-tools to verify.

SDK — call Trytet from your own agent code:

npm install trytet-client        # TypeScript
pip install trytet-client        # Python

API server:

tet up                            # Starts on port 3000
curl http://localhost:3000/health # Health check

Performance

| Operation | Latency | |---|---| | Cached cartridge call | <500µs | | First cartridge call (Cranelift compilation) | ~400ms (one-time, cached) | | MCP server boot | ~50ms | | Infinite loop trap | <100µs (instruction-level fuel exhaustion) |

MCP Tools

Trytet exposes 5 tools via the Model Context Protocol (3 ship with compiled .wasm, 2 experimental):

| Tool | Status | |---|---| | trytet_js_evaluator | shipped — Execute JavaScript with fuel and memory limits | | trytet_regex_evaluator | shipped — Run regex patterns safely (ReDoS-protected) | | trytet_jmespath_evaluator | shipped — Query JSON with JMESPath expressions | | trytet_scraper | experimental — Parse HTML with CSS selectors | | trytet_structured_data | experimental — SQLite-powered queries over JSON arrays |

Architecture

Three layers:

1. Sandbox — Wasmtime engine with instruction-level fuel metering. Each Wasm instruction deducts from a fuel budget. Exhaustion produces a deterministic trap — no wall-clock timeouts, no OS process overhead.

1. Cartridge Substrate — Wasm Components are loaded, compiled, and executed in sub-sandboxes with independent fuel and memory limits. The host controls all resources; cartridges own nothing.

1. Hive Mesh (experimental) — Agent snapshot, fork, and teleport between nodes for state migration.

Measured Performance

• MCP Server boot: ~50ms
• Cached cartridge call: <500µs
• First cartridge call: ~400ms (Cranelift compilation, one-time)
• Test suite: 0 failures across 64 test files

Detailed benchmarks: [BENCHMARKS.md](BENCHMARKS.md)

CLI

tet up          # Start the API server
tet mcp         # Start the MCP server (for Claude Desktop, Cursor)
tet ps          # List running agents
tet metrics     # Run benchmark suite
tet init [name] # Scaffold a new agent project

Full reference: [CLI.md](CLI.md)

SDKs

npm install trytet-client     # TypeScript
pip install trytet-client     # Python

Project Status

The core sandbox engine, cartridge substrate, and MCP server are functional. The project is under active development. Production deployments should pin to a specific release.

License

MIT

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bneb
• Source: bneb/trytet
• License: MIT

Install and usage instructions live in the source repository linked above.