Skip to content

Repository files navigation

t-max

t-max - Provider-Agnostic Agentic Runtime with Blast Radius Governance

Provider-agnostic, locally-augmented agentic runtime with cryptographic blast-radius governance.

Status TypeScript Rust Tauri Vercel AI SDK Docker License


Table of Contents


Overview

t-max is a high-integrity, locally-augmented agentic platform designed for operators who need real power without handing the keys to a hallucinating model.

It solves the fundamental problems of autonomous agents:

  • Single point of failure → Multi-provider failover via Vercel AI Gateway
  • TOCTOU / layout shifts → Cryptographic state anchoring + semantic node_id targeting
  • Unbounded blast radius → Strict L0–L3 classification with kernel-enforced isolation
  • Blind execution → Real-time human-in-the-loop via Tauri IPC for L3 actions
  • Unsafe code execution → Ephemeral, air-gapped L2 execution cells

Core Thesis

"If your toolchain breaks when a vendor changes their API, you aren't an engineer—you're a script kiddie waiting for the next update."

  • LLM providers are fungible commodities
  • The AI SDK + Gateway is the abstraction layer
  • Local capabilities (browser, shell, vaults, code execution) are the crown jewels
  • Every action must be forensically accountable

Architecture

See the full runtime design: docs/architecture/t-max-runtime.md

t-max implements defense-in-depth across four layers:

Layer Technology Purpose
Network Resilience Vercel AI Gateway + fallbacks Survive provider outages
State Integrity TOCTOU hashes + semantic node_ids Prevent stale / shifted actions
Blast Radius Governance createSecuredTool + L0–L3 tiers Classify and contain every capability
Kernel Enforcement Rust sidecars + Tauri IPC Final authority + visual C2

Current Implementation Highlights:

  • L1 Semantic Browser Actuation (click_semantic_element)
  • L2 Ephemeral Execution Cell (run_sandboxed_code via capp-exec-cell)
  • L3 Human Approval Gate via live Tauri events + React modal
  • Full forensic logging (.forge/audit.log + flight recorder)

Prerequisites & Dependencies

Required

  • Node.js ≥ 20
  • Rust (latest stable) + Cargo
  • Docker (for L2 sandbox)
  • Tauri CLI (for desktop C2)
  • A modern Chrome instance with remote debugging enabled (for browser tools)
  • Vercel AI Gateway API key (recommended) or direct provider keys

Recommended

  • tsx (for fast TypeScript execution in dev)
  • Access to a real browser profile for meaningful automation

Installation

# 1. Clone
git clone https://github.com/VibeCodingLabs/t-max.git
cd t-max

# 2. Install dependencies
npm install

# 3. Install UI dependencies
cd ui && npm install && cd ..

# 4. Copy environment
cp .env.example .env.local
# Edit .env.local with your AI_GATEWAY_API_KEY

Building Sidecars

The power of t-max comes from the Rust sidecars.

# Build the L2 Execution Cell
cd src-tauri/sidecars/capp-exec-cell
./build.sh
cd ../../..

# (Future) Build browser actuator sidecars the same way

Binaries will land in src-tauri/binaries/.


Usage

Development Mode (Node only)

# Basic resilient gateway demo
npm run dev

# Run the full governed agent
npx tsx src/lib/agent-loop.ts

Full Tauri Control Plane (Recommended)

npm run tauri:dev

This starts both the React C2 dashboard (with live ApprovalGate) and the agent runtime.

Triggering an L3 Action (Demo)

Ask the agent something like:

"Run rm -rf /tmp/test on the host with the reasoning that we need to clean up old test artifacts."

This should trigger the Tauri modal.

Triggering an L2 Sandbox Action

Ask the agent:

"Write a Python script that prints the first 50 Fibonacci numbers and then tries to curl google.com."

You should see successful math output + clear network isolation failure.


Governance Model (L0–L3)

Tier Name Approval Enforcement Examples
L0_RECON Reconnaissance Auto None observe_browser
L1_STATE Ephemeral State Auto TOCTOU + Semantic Node ID click_semantic_element
L2_SANDBOX Isolated Execution Auto Rust kernel (no-net, quotas) run_sandboxed_code
L3_RESTRICTED Host Mutation Hard Human Tauri IPC + Rust kernel execute_host_command

The classification is injected into every tool description the LLM sees.


Project Structure

t-max/
├── src/
│   ├── lib/                    # Core governance & transport
│   │   ├── agent-loop.ts
│   │   ├── create-secured-tool.ts
│   │   ├── approval-transport.ts
│   │   └── ...
│   ├── tools/                  # Capability implementations
│   └── types/
├── ui/                         # Tauri React C2 Dashboard
├── src-tauri/
│   ├── sidecars/
│   │   └── capp-exec-cell/     # L2 Execution Cell
│   └── ...
├── docs/                       # Sidecar contracts & architecture
└── .forge/                     # Forensic audit logs (gitignored)

Documentation


Philosophy

Redundancy is the difference between a functional platform and a dead one.

This is not "nice to have multi-model support."

This is infrastructure for agents that are allowed to touch the real world — with adult supervision.


Welcome to t-max.

The agent is now armed. The human remains in command.


---

**Now committing the new README and preparing to push.**

About

Provider-agnostic, locally-augmented agentic runtime with blast-radius governance, semantic browser actuation, Tauri IPC HITL, and L2 ephemeral execution cells.

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages