Wasmagent Js

What WasmAgent/wasmagent-js MCP server does

WasmAgent/wasmagent-js MCP server is a JavaScript toolkit for agent runtimes that need isolated execution, tool governance, evidence records, or shared state between an agent and a user interface. The repository presents these capabilities as separate packages rather than as one documented server binary.

The execution path runs agent-created JavaScript inside a WASM kernel with no host-process access. The quickstart uses @wasmagent/aisdk and @wasmagent/kernel-quickjs, and the repository also lists remote, Pyodide, and Wasmtime-related isolation options in its kernel documentation. The material does not establish that every listed kernel is exposed through an MCP server.

Governance features cover tool inspection, per-call policy decisions, consent records, taint observations, and signed evidence export. Shared-state support gives an agent a projected view of application state and lets it dispatch semantic actions through reducer-backed tools.

How it works

For sandboxed execution, create a QuickJSKernel and pass it to sandboxedJsTool. The resulting tool can be added to an AI SDK, LangChain, or OpenAI Agents setup according to the repository’s quickstart. The documented isolation boundary prevents the executed code from accessing the host process.

The MCP firewall path wraps another MCP server. It can snapshot a tool descriptor at registration, perform a static vetting step, evaluate a policy against the requested arguments and consent records, and attach taint information to returned data. The README characterizes this firewall as a first-line filter rather than an adversarial-grade defense.

For evidence, AEPEmitter collects actions during a run and builds a versioned evidence record. The example records a tool name, outcome, and exit code. Shared state is defined with an initial state function, reducer, projection, and affordances; stateTools then supplies read_state and dispatch_action for a session.

Setup and configuration

Install only the package path that matches the required capability. Sandboxed execution uses npm install @wasmagent/aisdk @wasmagent/kernel-quickjs; firewall protection uses npm install @wasmagent/mcp-firewall; evidence export uses npm install @wasmagent/aep; and shared state uses npm install @wasmagent/core.

It also does not provide setup instructions for a named desktop or editor client. WasmAgent/wasmagent-js MCP server should therefore be evaluated as a package-based runtime toolkit unless additional repository documentation supplies transport details.

Tools and capabilities

Documented capabilities include:

• Execute JavaScript through a QuickJS WASM kernel.
• Wrap MCP tools with descriptor snapshots, vetting, policy evaluation, consent handling, and taint observations.
• Emit signed AEP records after agent runs.
• Score or admit trace-derived compliance records through the related packages.
• Expose reducer-backed shared-state operations through read_state and dispatch_action.
• Provide adapters intended for AI SDK, LangChain, and OpenAI Agents integrations.

Package maturity varies. Core and the QuickJS and remote kernels are marked stable; the MCP firewall and AEP packages are beta; AI SDK and Mastra sandbox adapters are alpha. The firewall is specifically documented as a first-line filter using keyword and lightweight n-gram analysis, so it should not be treated as the only security boundary.

Limitations and notes

Some components are not published or are explicitly private, including the MCP policy and MCP attestation packages. Stability also differs by package, and alpha interfaces may add fields without a breaking-version change. Choose the package and kernel deliberately, and apply defense in depth when protecting untrusted tool use.