wasmagent-jsSecure portable agent runtime

WASM isolation · Cloudflare-ready · rollout ranking · training-data export hooks.

Three-tier code execution

VmKernel (in-process), QuickJS / Pyodide / Wasmtime (true WASM, edge-safe), RemoteSandboxKernel (E2B / CF Sandbox microVMs). Drop-in interchangeable. No other framework ships all three.

Decision tree

💸

Prompt-cache cost control

Auto cache breakpoints, 1-hour TTL, per-TTL usage metering. Verified –86% tokens in CI-gated benchmark.

See benchmarks

🇨🇳

Chinese models, first-class

Doubao, DeepSeek, Kimi, Qwen, GLM, MiniMax. Each adapter handles thinking modes + cache strategies + reasoning fields correctly. None of the Western frameworks ship these.

⏪

Time-travel debugger

EventLogReplay + opt-in <DevTools/> React UI. Step replay any agent run, fork from any step. Tiny opt-in package.

Open guide

🔁

Durable runtime

Checkpoints, SSE Last-Event-ID resume, stateless human-in-the-loop. Backends - CF KV, Durable Objects, Redis, Upstash.

Architecture

🔄

RLAIF data flywheel

RolloutForkRunner + RolloutRanker export real build-verified trajectories as DPO/PPO training records. bscode collects the signal; trace-pipeline trains on it.

Pipeline guide

🧪

Honest, CI-gated benchmarks

Every percentage in the README is reproduced in CI - drift outside ±10 % fails the build. No bit-rot.

See the numbers

📊

Pareto-first model evaluation

@wasmagent/evals-runner ships six benchmark suites and paired statistics (McNemar / Wilson / bootstrap) — multi-model comparisons surface trade-off fronts, not single-number ranks.

Open guide

🛡

MCP Guard + Evidence

Wrap any MCP server with policy enforcement. Deny injection/exfiltration attacks before tool execution. Export verifiable AEP evidence bundles after every run.

5-min quickstart

Apache-2.0 · vs Vercel AI SDK / Mastra / LangGraph · Runtime + App + Data Factory ecosystem