1. Zero-Latency Tool Interception
Sub-millisecond (<0.8ms) evaluation overhead in your agent process.
When you call wrapTools(), Moven creates lightweight proxy wrappers around your agent's tools (OpenAI, LangChain, Vercel AI SDK, LlamaIndex, CrewAI, AutoGen). Every time the agent invokes a tool, Moven intercepts the payload in-memory before executing the actual function.
2. Deterministic Canonical Hashing
SHA-256 payload sorting prevents key order mismatches.
Arguments are stringified using deterministic key-sorting. Even if LLM JSON outputs reorder properties across turns (e.g. {a:1, b:2} vs {b:2, a:1}), Moven generates identical SHA-256 hashes to track true execution uniqueness.
3. Real-Time Heuristic Evaluation
5 active safety shields watching every tool execution.
Flags identical tool calls N times within a 60-second sliding window.
Caps total dollar spending per run while granting 25% headroom if state progress is made.
Limits sub-agent call graph recursion depth to prevent stack explosions.
Trips when tool outputs remain 100% identical for 3 consecutive turns.
4. Adaptive Recovery & Cheaper Fallback
Automatic model downgrading and primary model restoration.
When a soft trip occurs, Moven automatically routes subsequent model prompts to a cheaper model (e.g. gpt-4o-mini or gemini-2.5-flash-lite). Once the agent achieves 3 consecutive clean turns under fallback, Moven automatically restores primary model execution.
Complete System Architecture Diagram
Interactive overview of the Moven in-process engine and cloud dashboard.
OpenCoreArchitecture
Purein-memoryexecutionwrapper+optionalhostedAPIdashboard
Lightweight npm package (@moven/sdk). Zero backend dependencies.
Tracks per-run tool call counts, argument hashes, cumulative token costs, recursion depth, and turn state hashes.
Synchronous heuristic checks after each tool call. On limit breach, halts stream or throws catchable MovenKillError.
If MOVEN_API_KEY is set, posts trip logs to Next.js API + Supabase for Slack/Email alerting.