mcp-v8 — a JavaScript/TypeScript runtime for AI agents

mcp-v8 is a Model Context Protocol server, written in Rust, that lets an AI agent run JavaScript and TypeScript in a sandboxed V8 isolate. Instead of wiring up dozens of narrow tools, you give the agent one tool — run_js — and it writes code: looping, branching, transforming data, and calling other tools, often with far fewer tokens than equivalent tool-call chains.

In its default stateful mode the V8 heap is saved as a content-addressed snapshot, so an agent can build up state across many turns. Host capabilities (network, filesystem, subprocess, WebAssembly, module imports, and calls to other MCP servers) are all off by default and unlocked only by explicit OPA/Rego policies.

Why mcp-v8

• One tool, unbounded capability. The agent runs a program, not a fixed menu of tools.
• Durable state. Heap snapshots persist variables and objects across calls.
• Secure by default. fetch, filesystem, subprocess, and external imports are denied until you grant them via policy.
• Production-ready. stdio / Streamable HTTP / SSE transports, a REST sidecar, async execution with pagination, JWKS auth, and Raft-replicated clustering.

Documentation

Full documentation lives at https://r33drichards.github.io/mcp-js/ (built from site-docs/) — tutorials, how-to guides, concept explanations, and complete reference for the CLI flags, HTTP API, and MCP tools.

Quick start

Install

# Server
curl -fsSL https://raw.githubusercontent.com/r33drichards/mcp-js/main/install.sh | sudo bash

# Optional CLI client
curl -fsSL https://raw.githubusercontent.com/r33drichards/mcp-js/main/install-cli.sh | sudo bash

Installs to /usr/local/bin. Supported platforms: Linux x86_64/arm64 and macOS Apple Silicon. You can also nix run github:r33drichards/mcp-js, use Docker (see the docker-compose.*.yml stacks), or build from source.

Connect an MCP client

# Claude Code (stdio)
claude mcp add mcp-v8 -- mcp-v8 --directory-path /tmp/mcp-v8-heaps   # stateful
claude mcp add mcp-v8 -- mcp-v8 --stateless                          # stateless

For Claude Desktop / Cursor, add to the client's mcpServers config:

{ "mcpServers": { "js": { "command": "mcp-v8", "args": ["--stateless"] } } }

Then ask the agent: "Run this JavaScript: console.log([1,2,3].map(x => x*2))".

Run over HTTP

mcp-v8 --stateless --http-port 8080
# MCP endpoint: POST http://localhost:8080/mcp
# REST sidecar: POST http://localhost:8080/api/exec

See the Quick Start tutorials and the transports guide for more.

Features

• JavaScript & TypeScript in an isolated V8 engine (via deno_core); TypeScript types are stripped with SWC (type removal, not type checking).
• Async/await & timers — Promises and the event loop, plus setTimeout/clearTimeout.
• Console capture — console.log/info/warn/error/debug/trace, streamed to storage and readable with line- or byte-based pagination.
• Async execution model — run_js returns an execution ID; poll status and stream output; cancel running work.
• Content-addressed heap snapshots — persist/restore V8 state across calls (local FS, S3, or S3 + write-through cache), or run stateless.
• WebAssembly — the standard WebAssembly API, plus pre-loaded modules (--wasm-module) exposed as globals and advertised to clients as runjs__wasm__<name> stub tools.
• ES module imports — optional npm:, jsr:, and URL imports fetched at runtime (policy-gated).
• Policy-gated capabilities — fetch, filesystem (fs), and subprocess access, each checked against a Rego policy per operation; plus header/OAuth injection for fetch.
• Compose other MCP servers — connect upstream MCP servers and call them from JS via mcp.callTool() / mcp.listTools().
• Customizable surface — override the server instructions and the run_js description (--instructions, --run-js-description).
• Auth & clustering — JWKS-based JWT verification, and optional Raft clustering with replicated session metadata and horizontal scaling.
• Multiple transports — stdio, Streamable HTTP (MCP 2025-03-26+), and SSE, with a REST sidecar and OpenAPI spec.

What the agent's code can do

These globals are available inside run_js (capability globals require a policy):

See Concepts → Security policies for the policy model.

MCP tools

Full parameters: MCP tools reference.

Configuration

mcp-v8 is configured entirely through CLI flags — storage backend, transport, execution limits, policies, fetch-header injection, WASM modules, clustering, JWKS auth, and the prompt/tool-description overrides. The complete, always-current list is the generated CLI flags reference.

mcp-v8 --help            # all flags
mcp-v8 --print-openapi   # print the REST OpenAPI spec

CLI client (mcp-v8-cli)

A fully-typed client for the REST API, generated from the OpenAPI spec via progenitor:

mcp-v8 --stateless --http-port 3000 &
mcp-v8-cli exec "console.log('hello'); 1 + 1"
mcp-v8-cli executions get <execution_id>
mcp-v8-cli executions output <execution_id>
export MCP_V8_URL=https://my-server.example.com   # point at a remote server

Rust client (mcp-v8-client)

[dependencies]
mcp-v8-client = { git = "https://github.com/r33drichards/mcp-js" }

use mcp_v8_client::Client;
let client = Client::new("http://localhost:3000");
let body = mcp_v8_client::types::ExecRequest {
    code: "1 + 1".to_string(),
    heap: None, session: None,
    heap_memory_max_mb: None, execution_timeout_secs: None, tags: None,
};
let resp = client.exec_handler(&body).await?;
println!("execution_id: {}", resp.into_inner().execution_id);

Build from source

The repo is a Nix flake (it wires up the prefetched V8 archive so the build stays offline-friendly):

nix build github:r33drichards/mcp-js   # → ./result/bin/server
# or for development:
nix develop      # then: cargo build -p server

A plain cargo build --release inside server/ also works if your toolchain can build deno_core/V8.

Limitations

• setInterval is not available — use a loop with awaited setTimeout.
• No DOM or browser APIs — there is no window/document.
• TypeScript is type-stripped, not type-checked — invalid types are removed, not reported. JSX/TSX is not supported (it parses to a clear error).

License

GNU AGPL-3.0.

MCP-V8 Load Test Benchmark Report v0.1.0

Comparison of single-node vs 3-node cluster at various request rates.

Results

ran on railway gha runners on pr

P95 Latency

Notes

• Target Rate: The configured constant-arrival-rate (requests/second k6 attempts)
• Actual Iter/s: Achieved iterations per second (each iteration = 1 POST /api/exec)
• HTTP Req/s: Total HTTP requests per second (1 per iteration)
• Dropped: Iterations k6 couldn't schedule because VUs were exhausted (indicates server saturation)
• Topology: single = 1 MCP-V8 node; cluster = 3 MCP-V8 nodes with Raft