Voight
The observability and debugging infrastructure for autonomous systems.
Autonomous software does not fail like software. An agent reasons, calls tools, spends money and takes actions on its own, across dozens of steps, and when it goes wrong there is no stack trace that explains why. Voight is built for that class of system. It captures every prompt, tool call, decision, cost and error from any agent runtime, reconstructs the full chain of reasoning as a trace, learns each agent's normal behaviour to detect drift before it becomes an incident, diagnoses failures down to their root cause, and preserves a tamper-evident record of everything that happened.
Today that covers coding agents, autonomous workflows and production LLM applications. The next surface is physical: robots, drones and industrial systems run the same loop of perception, decision and action, and fail in the same opaque ways. Voight is building the observability layer for that sector, starting with a drone flown by an autonomous agent.
Voight is built by Galaxyhub Labs Inc., a Delaware company founded in 2025.
The platform
Voight is organised as five layers. The first four are in production; the fifth is in active development.
Capture layer
One event model for any runtime. Hooks for Claude Code and Cursor, a native OpenTelemetry receiver for Codex sessions, wrappers for the OpenAI and Anthropic SDKs, a span exporter for the Vercel AI SDK, OpenTelemetry for Python through bitfrost, and a plain HTTP API for everything else. Frameworks are detected automatically. Three capture levels (Minimal, Standard, Full) decide how much content leaves the machine, with PII scrubbed locally before anything is sent; tokens, cost and latency flow at every level.
Understanding layer
Events become traces with a full waterfall, sessions and tags, with token, cost and latency attribution down to the individual end user. Anomaly detection learns each agent's baseline (tool mix, latency distribution, call patterns) and alerts the moment behaviour drifts, with no rules to write. The debugging engine groups repeated failures into an issue queue and explains each one with an AI diagnosis that names the root cause, not just the symptom.
Assurance layer
A tamper-evident audit trail: every event is hashed and stored with its full context, so any run can be searched, replayed and reviewed later, and an edit after the fact is detectable by design. Alerts and notification channels, per-event privacy controls, and published security and compliance documentation covering GDPR, OWASP LLM Top 10, NIST AI RMF and SOC 2 readiness.
Runtime layer: Voight Agents
The platform's live proving ground. Voight hosts and runs autonomous agents on cloud or decentralized GPU, with persistent memory, scheduled tasks, channels (Telegram, GitHub, X, LinkedIn), an on-chain identity per agent, and orchestration from Claude Code, Cursor or Codex through Voight's MCP server. Everything the platform observes is exercised here first, at fleet scale.
Physical layer: agents on hardware
Observability and debugging for agents that run or control drones, robots and IoT devices: the same capture, understanding and assurance layers, applied to systems where a wrong decision moves something in the real world. In active development: the first tests and MVP run on a drone flown autonomously by an agent, with every telemetry signal and every decision streamed into Voight, so that a failure can be traced and debugged in real time. The goal is broader: capture everything from any agent that runs on hardware, and build the observability layer for that sector.
Built and shipped
| When | What |
|---|---|
| April 2026 | Platform live: dashboard, coding-agent SDK, HTTP ingestion, public Explorer indexing 1,430 agents from day one |
| May 2026 | First production events; App SDKs for OpenAI, Anthropic and the Vercel AI SDK; Python through bitfrost |
| June 2026 | Voight Agents runtime: dedicated compute service, Telegram channel, scheduled tasks, GitHub App |
| July 2026 | Voight Agents public launch; on-chain agent identity through the Metaplex Agent Registry; LinkedIn channel; Nosana grant |
| August 2026 | Agents on decentralized GPU (Nosana markets with live availability); ZeroClaw as a second agent framework |
| September 2026 | Smart Traces: span waterfall, deterministic diagnosis and AI explanation; native OpenTelemetry receiver for Codex; issue queue with AI diagnosis; MCP server to deploy, operate and orchestrate agents from Claude Code, Cursor and Codex |
Traction
As of September 26, 2026
- 90,399 events processed since May 2, 2026; 13,666 in the last 30 days.
- 88 agents instrumented with Voight.
- 70 hosted agents live for 64 owners; 22 reachable on Telegram; 6 on decentralized GPU.
- 1,600 agents indexed in the public Explorer.
- 141 issues detected and grouped by the debugging engine.
- Open-source packages: about 530 npm downloads and 167 PyPI downloads in the last 30 days across the SDK, the App SDKs, the MCP server and bitfrost.
- Everything Voight ships to developers is public code: 10 repositories at github.com/Voightxyz covering the SDK, the App SDKs, Python, the MCP server, the Nosana tooling and the docs.
Programs and partners
- Xiaomi Orbit
- Voight is part of Xiaomi's startup program. The relationship goes beyond model access: MiMo models run inside the product.
- Z.ai Startups
- Voight is part of the Z.ai startup program. GLM models are the default inference for hosted agents and for the AI diagnosis in the debugging engine.
- Nosana
- Nosana backs Voight's decentralized GPU work for agents with two open-source integrations and a $15,000 grant announced in July 2026. It is the compute layer behind Voight Agents on GPU, a complement to the observability platform. Announcement.
- Metaplex
- Metaplex supports Voight's agent identity and traceability integration: every hosted agent gets an identity in the on-chain Agent Registry, which gives it a verifiable origin. The observability platform does not depend on it.
Recognition
- Winner, La Familia Frontier Hackathon Side Track (Superteam Earn, May 2026): 1,000 USDG, one of six winners out of 36 submissions. Listing.
- Winner, theMiracle Frontier Track (Superteam Earn, May 2026): $10,000 prize. Listing.
- Speakers at web3 and AI side events across Europe, including MERGE Madrid and JET Europa in Barcelona.
- Members of La Familia (Superteam Spain).
Frequently asked questions
- What is Voight?
- Voight is the observability and debugging infrastructure for autonomous systems. It captures everything an AI agent does, reconstructs its reasoning as a trace, detects behavioural drift, diagnoses failures to their root cause and keeps a tamper-evident audit trail. Built by Galaxyhub Labs Inc., Delaware, founded in 2025.
- Who is Voight for?
- Teams that build or operate autonomous systems: coding agents, autonomous workflows, production LLM applications with per-user costs to control, fleets of hosted agents, and, next, agents that run or control hardware.
- Is Voight in production?
- Yes. The platform has been live since April 2026. As of September 26, 2026 it has processed 90,399 events from 88 instrumented agents, and runs 70 hosted agents for 64 owners.
- What does Voight integrate with?
- Claude Code and Cursor (hooks), Codex (native OpenTelemetry sessions), the OpenAI and Anthropic SDKs (wrappers), the Vercel AI SDK (span exporter), Python through bitfrost (OpenTelemetry), and any language through the HTTP API. Hosted agents connect to Telegram, GitHub, X and LinkedIn and are operated from Claude Code, Cursor or Codex through Voight's MCP server.
- Is Voight open source?
- The integration layer is. The SDK, the App SDKs for OpenAI, Anthropic and the Vercel AI SDK, bitfrost for Python, the MCP server, the Nosana tooling and the docs are public code: 10 repositories at github.com/Voightxyz, so anyone can read what runs on their machine. The hosted platform (API, dashboard and debugging engine) is proprietary.
- What makes Voight different?
- One platform covers the whole lifecycle of an autonomous system: capture from any runtime, traces with per-end-user cost attribution, anomaly detection on each agent's own baseline, root-cause debugging with AI diagnosis, and a tamper-evident audit trail. It runs its own fleet of hosted agents on the same platform, and it is being extended from software agents to agents on hardware, starting with an autonomous drone.
- What is Voight Agents?
- The runtime layer of the platform and a product in its own right: autonomous agents that Voight hosts and runs on cloud or decentralized GPU, with persistent memory, scheduled tasks, channels and an on-chain identity per agent through the Metaplex Agent Registry. Teams without agents of their own start here; everything they do is observable from day one.
- Who is behind Voight?
- Dangel Rodriguez, founder and three-time hackathon winner, leads a team of four: three engineers across software, AI and systems, and one business developer, split between Spain and London. The company is Galaxyhub Labs Inc., Delaware, founded in 2025.
- Who backs Voight?
- Voight is part of Xiaomi Orbit and Z.ai Startups, is supported by Nosana (decentralized GPU, $15,000 grant announced in July 2026) and Metaplex (agent identity), and is a member of La Familia (Superteam Spain). It has won two Solana hackathon tracks on Superteam Earn.
- How does Voight handle privacy and security?
- Three capture levels decide how much content leaves the machine; PII is scrubbed locally at the Standard level; only metadata leaves at Minimal. Every event is hashed into a tamper-evident audit trail. Security and compliance documentation covers GDPR, OWASP LLM Top 10, NIST AI RMF and SOC 2 readiness at docs.voight.xyz/trust/overview.
- What does Voight cost?
- Observability: Free ($0), Pro ($19 per month), Enterprise from $89 per month. Voight Agents: $15 per agent per 30 days.
- Where is Voight going?
- To physical autonomous systems. Drones, robots and industrial automation run the same loop of perception, decision and action as software agents, and Voight is building the observability layer for that sector. The first tests and MVP run on a drone flown by an autonomous agent, with all telemetry and decisions observed and debugged in Voight in real time.
- How do I contact Voight?
- team@voight.xyz.