Klein is an always-on autonomous strategist system designed to notice, synthesize, and propose high-leverage next moves before the user explicitly asks. It combines a low-overhead VPS brain, local execution hands, narrow MCP tooling, durable memory, and human approval by default.
Deliberately narrow interface with exactly two Claude-facing tools to reduce routing noise.
Thinking canvasExecutes local actions only when available: shell, AppleScript, files, clipboard, browser control.
Local executionDynamic approval, clarification, and handoff notifications with no exposed tool syntax.
Low-friction interfaceGit-tracked markdown, SQLite logs/state, and raw source cache. Human-readable and cheap to operate.
Persistent contextBounded autonomy with approval gates, bucketed actions, graceful degradation, and staged trust.
Control layerAlways-on VPS runtime coordinating retrieval, synthesis, ventures, research, memory updates, and decision proposals. Built to do something useful even when nothing is actively asking it to act.
The system is intentionally lean: a cloud brain on a Lightsail VPS, a local daemon for device-side actions, a narrow MCP surface for Claude, and a memory layer optimized for portability over infrastructure bloat.
Klein runs continuously on a small Ubuntu VPS, keeping canonical state, orchestrating synthesis passes, and serving the MCP layer without requiring a heavyweight cluster.
Claude sees only the context fetch and write-back tools. Routing, judgment, and internal action logic stay inside Klein rather than being pushed into a noisy tool layer.
v1 uses markdown, SQLite, and raw source caching instead of a vector-heavy stack. That keeps memory transparent, portable, cheap, and easy to audit.
Klein is meant to act like a strategist, but with approval-gated side effects, confidence thresholds, quiet hours, action buckets, and graceful fallback when local hands are unavailable.
Watches for deltas throughout the day: a meeting ended, a deadline moved closer, a free block opened up. Most ticks avoid model calls; only threshold-crossing moments escalate.
Reviews recent changes, refreshes active summaries, produces a small number of nudges or memos, and runs the venture loop that lets Klein initiate strategic work on its own.
Cross-project review, opportunity detection, pruning, and narrative correction. This is where the system checks whether it is actually surfacing things that manual chat usage would have missed.
The architecture is intentionally small-footprint while still supporting persistence, initiative, control, and graceful failure handling.
Klein separates reflexive nudges, deep reasoning, and lightweight control so the interface never collapses into a single overloaded chat box.
Natural-language prompts with dynamic response actions such as approve, edit, reject, or timeline clarifications. Replies write directly to memory without requiring a fresh model roundtrip.
Used for debriefs, writing, tradeoffs, and ambiguity. Klein preloads the right context packet instead of forcing the user to rebuild state manually at the start of every session.
One string entrypoint absorbs questions, admin, drafts, approvals, synthesis triggers, and device-side actions. Short output is a feature; longer work escalates into Claude with context already loaded.
Klein is designed to feel agentic without becoming reckless. The default operating mode is propose. Side-effecting actions are approval-gated, confidence-tiered, and separated into brain-only versus device-dependent buckets.
Klein is not presented as a consumer chatbot with polished screenshots. It is an infrastructure-backed agent system with a clear control model, cost-aware operation, and a concrete answer to the hardest part of AI products: persistent judgment across time.