Nous Research

How the team behind Hermes Agent's 22 million downloads investigates production

Updated on November 19, 2026

Nous Research is the open-source AI lab behind Hermes Agent, downloaded 22.7 million times. Its small cloud services team sends about 28 TB of logs and traces a month to Better Stack and lets a Hermes agent query them from a Discord channel, where one-off reports that once took an engineer hours now come back in about 15 minutes.

Ben Barclay
Software Engineer, Nous Research
22.7M
Hermes Agent downloads
28 TB
Of logs and traces ingested every month
15 min
For a one-off report correlating requests with Stripe
Industry
AI / open-source AI lab
Products used
Telemetry, MCP, Uptime, On-call
Customer since
October 2024

Where they started

Ben works on Nous Research's cloud services. That team operates the services behind Hermes Agent and Nous's inference API, which are growing and changing fast.

Ben and several of his colleagues came to Nous with years of experience running the self-hosted, alphabet-soup school of log tooling, plus a big-name enterprise platform. They knew the cost of that model well: ingestion that takes work to keep running, storage that gets expensive as volume grows, and licensing on top.

Better Stack gave the team a ClickHouse-based store they could send all their logs and traces to, with query patterns they liked and pricing that stayed reasonable at their volume. Status pages came built in, which the team saw as a useful extra.

“I've been very happy with Better Stack, both the system and the service. The few times I've had to reach out for support, it's always been a great experience. Huge fan.”

Ben Barclay
Software Engineer, Nous Research

How they set it up: pipe in everything, then let an agent do the watching

Nous sends OpenTelemetry data from services on Railway and Vercel, plus Vercel's own logs, into Better Stack across more than 30 sources. That comes to about 28 TB of logs and traces a month, and the production API traces source passed 1.4 TB on its own in April 2026. There were dashboards too. They were fine, and nobody stared at them much.

So Ben gave the team a colleague who would. As a side project, he put a Hermes agent in their support and monitoring channel on Discord and handed it a read-only Better Stack API key. Engineers asked it questions, and it dug through logs and traces faster than any of them could. Today the agents read the data almost entirely through the Better Stack MCP, with analysis skills the team has built around it.

The agent kept getting more to read. Read-only database access let it match traces against the state the databases ended up in. Every morning it posts a report on service health and anything odd that happened overnight.

The Hermes agent's morning service-health report in Nous's Discord monitoring channel
The Hermes agent's morning report on service health, posted in Nous's Discord monitoring channel.

Eventually, one agent that could see everything, company financials included, felt like a bit much. The team is now splitting it into smaller agents scoped by role.

One question, 15 minutes: the agent does the homework

The agent earned its keep straight away. Reading through traces, it spotted services calling the database more often than they needed to and showed where those calls could be merged. Ben's verdict: a quick, easy win.

The better trick was answering questions that used to be too much effort to ask. One day the team wanted to know how many requests had gone through Nous's inference API, and how that compared with the day's Stripe purchases.

The agent pulled the request logs from Better Stack and the purchases from Stripe, lined them up, and about 15 minutes later posted an HTML report, charts included. Ben could have built it himself. It would have taken him a long time for a one-off, and now he does not have to. The dashboards, meanwhile, have mostly retired.

The agent's HTML report matching inference API requests with the day's Stripe purchases
The HTML report the agent built, matching inference API requests with the day's Stripe purchases. Figures redacted.

“Dashboards almost feel like a legacy thing, the war board you'd put up in the office to watch metrics tick up and down. Now we can have the agents watch that for us and be a lot more proactive on our behalf.”

Ben Barclay
Software Engineer, Nous Research

What it costs now: a small team and a pricing model they like

Cost is where this started. Ben has run log tooling where getting data in is work, keeping it is expensive, and the licence comes on top. Better Stack's pricing was one of the main reasons Nous picked it, and at about 28 TB of logs and traces a month, Ben still calls it very cost-effective. The same account covers uptime monitoring, heartbeats and on-call for 14 responders.

The bigger saving is in people. Three to six engineers run a growing set of cloud services, with an agent doing the searching and reporting that would otherwise eat someone's day. Ben has worked on large enterprise teams, and he says Nous gets far more done for its size.

“Better Stack is something I can put in the background. It's nicely working infrastructure that I just don't have to think about.”

Ben Barclay
Software Engineer, Nous Research
Give your agents your telemetry, like Nous Research.

Connect the Better Stack MCP and ask your logs and traces a question.