Self-hosting observability for your AI agents on a $6 Droplet

Question

I’ve been running a few small LLM agents lately, and the thing that kept biting me wasn’t the model — it was not being able to see what the agent actually did. Which tool did it loop on five times? Where did that run quietly cost me a dollar? Did last week’s prompt tweak make things worse or better? The logs never quite tell you.

The hosted tools for this want your prompts on their servers, and the self-hostable ones (Langfuse and friends) wanted me to stand up Postgres + ClickHouse + Redis + S3 — four moving parts to log a few thousand LLM calls a day. That’s a lot of Droplet for a side project.

So I ended up building a small open-source one, Otterscope, and self-hosting it is genuinely a one-liner. Figured I’d write down how I run it on a DigitalOcean Droplet, because “AI agent observability” sounds heavier than it actually is here.

KFSys · Accepted Answer

What it does, quickly You point your agent’s OpenTelemetry exporter at Otterscope and you get anagent-run -first view: every trace becomes a run, with its steps, toolloops, tokens, and cost. Click a run and you see the actual messages in andout of each model call. There’s a cost table, assertions and LLM-as-judgeevals scored onto your real runs, a compare view for “this week vs last,” andwebhook alerts when your error rate or spend crosses a line. The important part for hosting: it’s one static Go binary and one SQLitefile. No database to provision, no cache, no object store. That’s the wholereason it fits on the smallest Droplet. The Droplet A Basic $6/mo Droplet (1 vCPU, 1 GB RAM) is plenty unless you’re pushingserious volume — SQLite handles this scale comfortably. Spin one up withUbuntu, add your SSH key, done. There are two ways to run it. Docker is the fastest; the raw binary withsystemd is the leanest. Option A — Docker (fastest) If you ticked the Docker Marketplace image when creating the Droplet, or justapt install docker.io, it’s one command: docker run -d --restart unless-stopped \ -p 8317:8317 -p 4318:4318 \ -v otterscope:/data \ ghcr.io/otterscope/otterscope 8317 is the web UI, 4318 is the standard OTLP ingest port your agents sendto. The otterscope volume keeps your data across restarts. Option B — the binary + systemd (leanest) No Docker, ~15 MB resident: # grab the linux binary from the releases page, then: sudo mv otterscope /usr/local/bin/ sudo useradd --system --home /var/lib/otterscope --create-home otterscope Drop a unit at /etc/systemd/system/otterscope.service: [Unit] Description=Otterscope After=network.target [Service] User=otterscope # -listen/-otlp are set explicitly here because Otterscope binds to # localhost by default (see the security note below) ExecStart=/usr/local/bin/otterscope serve -db /var/lib/otterscope/otterscope.db -listen :8317 -otlp :4318 Restart=on-failure [Install] WantedBy=multi-user.target sudo systemctl enable --now otterscope One security note worth reading By default Otterscope binds to 127.0.0.1 — nothing is exposed to theinternet until you decide to. On a Droplet you have two sane choices: Keep it private (my preference for a personal instance): leave thedefault localhost bind, and reach the UI over an SSH tunnel from yourlaptop: ssh -L 8317:localhost:8317 you@your-droplet, then openhttp://localhost:8317. Nothing is public, ever. Expose it deliberately: pass -listen :8317 -otlp :4318 (as the unitabove does), and put a DigitalOcean Cloud Firewall in front so onlyyour app servers can reach 4318 and only you can reach 8317. Otterscopeis single-user and has no login yet, so don’t leave 8317 open to theworld. Point an agent at it Whatever framework you use — Pydantic AI, the OpenAI Agents SDK, LangGraph,the Vercel AI SDK — it’s the standard OpenTelemetry variable: export OTEL_EXPORTER_OTLP_ENDPOINT=http://YOUR_DROPLET_IP:4318 Run your agent, refresh the UI, and the run shows up in a couple of seconds.No SDK to install, no code changes if you’re already emitting OTel. Want to see it populated before wiring a real agent? otterscope sampleseeds a batch of realistic demo runs. What it adds up to $6/mo Droplet, and honestly a $4 one would do for light use. Zero managed-database cost — it’s a file on the Droplet’s disk. Snapshotthe Droplet or scp the .db file and that’s your backup. Unlimited retention by default (it’s your disk); there’s a -retention flagif you’d rather prune. That’s the pitch: the observability stack for your agents is a single processsitting next to them, costing about as much as a couple of coffees a month. Try it Repo and docs are here: https://github.com/otterscope/otterscope (Apache-2.0). There are per-framework setup guides under docs/frameworks/,and the Docker image is ghcr.io/otterscope/otterscope. If you self-host it on a Droplet and hit a rough edge — especially a frameworkwhose traces don’t map cleanly yet — open an issue. Raw payloads are retained,so fixes apply retroactively to data you’ve already collected.

Bobby · Answer

Hi there, This is a great write-up and a real problem worth solving. The observability gap for small LLM projects is genuinely annoying, most hosted tools are overkill or want your data, and the self-hosted alternatives like Langfuse have a setup cost that does not match a side project. The one-liner deploy on a $6 Droplet is a good angle. Would love to see the Otterscope repo linked here for anyone who wants to try it.

alexdo · Answer

Heya, Most agent observability projects I’ve come across want you to stand up a whole stack just to see a trace, so a single static binary plus SQLite that you can point a six dollar Droplet at and actually see what your agent is doing is refreshing. Bookmarking this one, nice work! Regards

Become a contributor for community

DigitalOcean Documentation

Resources for startups and AI-native businesses

The developer cloud

Start building today

Report this

Self-hosting observability for your AI agents on a $6 Droplet

Related questions

Ubuntu or creating droplet

How to migrate from hermes to openclaw?

production and staging sites

Which computer skill is most valuable for getting a job in 2026?

Any Help on How to add Ai Agent to Ubuntu

How to handle high memory usage and optimize resource allocation for custom server setups on a VPS?

How do I let my AI agent send emails without hardcoding SMTP credentials in the code?

What are the best AI frameworks and tools for adding AI features to mobile apps?

How do I send email notifications from an AI agent running on DigitalOcean?

What's the best way to manage multiple AI API providers (OpenAI, Midjourney, etc.) on a DigitalOcean-hosted app?

How do I verify that CUDA and the NVIDIA drivers are actually working on a new GPU Droplet?

What SMTP service are you using for transactional emails on your Droplet?

Report this

Self-hosting observability for your AI agents on a $6 Droplet

Related questions

Ubuntu or creating droplet

How to migrate from hermes to openclaw?

production and staging sites

Which computer skill is most valuable for getting a job in 2026?

Any Help on How to add Ai Agent to Ubuntu

How to handle high memory usage and optimize resource allocation for custom server setups on a VPS?

How do I let my AI agent send emails without hardcoding SMTP credentials in the code?

What are the best AI frameworks and tools for adding AI features to mobile apps?

How do I send email notifications from an AI agent running on DigitalOcean?

What's the best way to manage multiple AI API providers (OpenAI, Midjourney, etc.) on a DigitalOcean-hosted app?

How do I verify that CUDA and the NVIDIA drivers are actually working on a new GPU Droplet?

What SMTP service are you using for transactional emails on your Droplet?