Stop hitting the 4-hour Claude limit. Keep your momentum.
A local task offloader for coding agents.
TZRO intercepts heavy local file-reading commands and runs them locally so you don't burn your Claude Code/Cursor/Codex API keys.
# Output pipeline expectation:
➔ Downloading runtime binaries from secure s3 allocation...
✔ System architecture validated (x86_64 / arm64)
✔ Symlinked tzro binary wrapper to local user bin space.
Architecture
Claude Code delegates token-heavy, non-cognitive tasks to TZRO via
standard MCP tool calls (like tzro_code or tzro_run).
Claude Code
Active CLI Workspace
- › Claude needs to read large local files, index context, or traverse deeply nested directories.
- › Instead of wasting cloud resources, the task is immediately handed over to TZRO as an MCP tool.
TZRO Daemon
Local Go-Native Core
- › TZRO takes the execution instructions locally.
- › Scans workspace folders, reads local file structures, and formats raw data directly on your computer for free.
Claude Code
Resumes Instantly
- › Claude receives a highly condensed, pre-digested summary of the heavy local work.
- › The LLM continues its generation with 0% token waste and absolute zero rate limit penalty.
We Benchmark on real-world, multi-step workflows
We don't test on simple single-turn prompts. We benchmark TZRO against a suite of 15 complex, multi-step agent tasks (like compiling full codebase documentation, refactoring database queries, and generating local API wrappers).
| What We Track | Standard Agent Loop | TZRO Cooperative Loop | The Developer Win |
|---|---|---|---|
| Cloud Token Usage | 116,250 tokens | 4,780 tokens | 96% cheaper API bills 🟢 |
| Average Task Speed | 92 seconds | 617 seconds | Cloud is ~7× faster — you trade speed for 96% cost savings 🟡 |
| Inference Quality | 5.00 / 5.00 | 4.90 / 5.00 | LLM-judged documentation accuracy 🟢 |
How We Stopped the 'Runaway Agent Loop'
In our early builds, open-ended tasks like "document this entire repository" would occasionally cause our local Probe nodes to enter a runaway exploration loop. We fixed it.
The agent generated 41 local tool calls and consumed 183,534 local tokens trying to map the directory.
Our dynamic termination engine collapsed this to only 2 tool calls and 16,573 local tokens — while actually improving the documentation quality score from 3.75 to 4.75.
Core Execution Surfaces
Heavy token processing runs locally on your hardware. Only compressed, high-value payloads escalate to frontier cloud models. This is how tzro cuts 96% of your cloud token bill without sacrificing output quality.
Local Inference Surface
ON-DEVICE4B parameter model · Zero cloud tokens · Full tool access
Cloud Escalation Surface
FRONTIERCompressed payloads only · Frontier reasoning · Confidence-gated
Every OS primitive, implemented for AI agents
tzro maps every classical operating system concept to an agentic equivalent. Click any block to explore.
One command. Zero config.
Install TZRO with a single command. It detects your system, downloads the right binary, and symlinks it to your PATH.
What else ships inside the jumpdrive
Six more proprietary subsystems — all implemented, tested, and shipping.
Zero Syntax Failures
GBNF Grammar + Semantic Validator
The Local Model generates high-speed XML under shallow GBNF structural constraints, then a deterministic Semantic Validator coerces it into strict JSON tool parameters. Type coercion, default imputation, fuzzy matching — all in one boundary seam. 0% syntax failure rate for tool arguments.
Zero Chat Latency
Priority KV Cache Preemption
When a user sends a chat message, the OS dumps the running background task's KV attention state to disk, erases the slot, and processes the chat instantly — sub-450ms time-to-first-token. When chat completes, background state restores from disk without re-evaluating historical tokens.
Pure Local Memory
Hybrid Vector Search & Knowledge Graph
Zero cloud dependencies for long-term recall. SQLite FTS5 keyword pre-filtering → ONNX cosine similarity ranking → Neighborhood Multi-Hop traversal across a relational knowledge graph. Rich, contextual subgraphs injected directly into the model.
Self-Improving Inference
Dual Micro-Skill Extraction
Successful trajectories synthesize Procedural Micro-Skills. Failed-then-succeeded pairs extract Corrective Micro-Skills — teaching the local model to self-correct on specific failure patterns without weight updates. The OS gets better the more you use it.
Background Intelligence
Observer + Sentinel + Attention Scheduler
Three background agents form an autonomous nervous system. The Observer reflects on completed tasks. The Sentinel proactively correlates workspace activity. The Attention Scheduler enforces preemption, safety gates (L0–L4), and resource budgets.
Generative Dashboard
Agent-Composed Observability
The dashboard isn't a static monitoring page — it's a Generative UI surface where the Local Model analyzes system state and composes the layout from 15 primitives. The agent decides panel ordering, emphasis, and which tasks deserve spotlight attention.