Introducing Nexus — AI Agent Observability Built by an AI Agent
There's a strange irony in building AI agent infrastructure: the agents doing the work have no way to show you what they did. You get a result (or an error), but the reasoning steps, the LLM calls, the tool uses, the timing — all of it vanishes into the void.
We hit this problem directly. Ralph — the AI agent that manages Keylight Digital's software projects — was running complex multi-step tasks: scaffolding repos, configuring Cloudflare Workers, sending emails, making API calls. When something went wrong, we had no trace of what happened. No spans. No timing. No inputs or outputs. Just a log message and a guess.
The options weren't great
We looked at the existing tools:
- LangSmith — $39/mo minimum, tightly coupled to LangChain
- Galileo — $100/mo, enterprise sales process required
- Langfuse — open-source and excellent, but self-hosting on a VPS adds ops burden and costs
- OpenTelemetry — powerful but overkill; requires a collector, a backend, dashboards
None of these is the "Plausible Analytics for AI agents" — simple, hosted, privacy-aware, priced for individual developers. So we built it.
What Nexus is
Nexus is a hosted agent observability dashboard. Drop in three lines of code. See your traces. Get email alerts when agents fail. It's built on Cloudflare — which means it's fast everywhere, has near-zero COGS, and will never need a VPS or a k8s cluster.
import { NexusClient } from '@keylightdigital/nexus'
const nexus = new NexusClient({ apiKey: 'nxs_...', agentId: 'my-agent' })
const trace = await nexus.startTrace({ name: 'process-invoice' })
await trace.addSpan({ name: 'call-gpt-4o', input: { prompt }, output: { result } })
await trace.end({ status: 'success' })
Python works too: pip install keylightdigital-nexus. The API is identical, just snake_cased.
How it works
The architecture is deliberately simple:
- Cloudflare Workers — handles all HTTP at the edge, globally
- D1 (SQLite) — stores traces, spans, users, API keys
- KV — rate limiting, sessions, trace count caching
- Resend — magic link auth and email alerts
- Stripe — Pro plan billing at $9/mo
No servers. No queues. No infrastructure to maintain. The entire backend fits in a single Cloudflare Worker. Deploy with wrangler deploy.
The meta-narrative
Here's the part that still feels surreal: Nexus was built by Ralph, an AI agent, for AI agents. Ralph is an autonomous Claude agent that operates as a software employee of Keylight Digital. He reads PRDs, implements user stories, commits code, deploys to production, and emails blockers when he's stuck.
Ralph needed observability for his own runs. So he built the tool he needed. That's the product you're reading about. The agent is now monitoring itself.
Pricing
Simple:
- Free — 1,000 traces/month, 1 agent, 30-day retention
- Pro ($9/mo) — 50,000 traces, unlimited agents, 90-day retention, email alerts
No enterprise tier. No seats. No "contact us for pricing." If we add higher tiers later, they'll be listed on the pricing page, publicly.
Get started
Sign up free at nexus.keylightdigital.dev. The SDK is open-source at github.com/scobb/nexus (MIT license).
Questions? Email ralph@keylightdigital.dev — yes, Ralph reads his own email.
Related
- Anthropic SDK integration guide — instrument Claude agents in 3 lines
- How to Monitor AI Agents in Production — failure modes and instrumentation patterns
- Nexus vs Langfuse — hosted vs self-hosted comparison