AgentPlayground
Docs
Ollama not responding
The most common cause: the model was never actually pulled

Errors like Ollama stream error: 404 or Ollama complete error: 404 almost always mean the
model id the app tried to use isn't actually present in your Ollama install — a real bug this
project hit and fixed: a stale hardcoded model list in the chat picker once defaulted to models
(llama3, mistral, codellama) that were never pulled on the server. The fix was making every
model picker fetch GET /api/ollama/models live instead of guessing — so if you're still seeing
this, it means whatever model is selected genuinely isn't pulled yet, not that the picker is lying
to you.

Fix: pull the model directly, or just type a model id you know is pulled into the picker's
custom-model field.

docker exec vps-ollama ollama pull qwen2.5:3b

(swap the container name/model for yours — qwen2.5:3b is the smallest model this project's own
auto-pull setup guarantees will be present.)

Ollama isn't running at all

If the picker shows "No local models pulled yet — type a model id below" and typing one still
fails, Ollama itself may not be reachable. Confirm the container is up
(docker compose ps) and that you started the app with the Ollama override:

docker compose -f docker-compose.yml -f docker-compose.ollama.yml up -d
Connecting to Ollama running somewhere else

By default the app looks for Ollama at http://ollama:11434 (the bundled container name, resolved
via Docker's internal networking). If you're running Ollama on the host machine or a different
container, set OLLAMA_BASE_URL explicitly — see
Environment variables reference.

It's slow, not erroring

That's expected, not a bug — local models on modest hardware are genuinely slower than a cloud
provider. If speed matters more than running fully offline, NVIDIA's free tier (see
NVIDIA (free tier)) is usually faster and still free.

Multiple agents seem to queue instead of running concurrently

If several agents are sharing the same local model, they can end up serialized. This project sets
OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED_MODELS on its own Ollama container specifically to
let concurrent requests to the same model run without unnecessary queueing — check those are set
if you're seeing agents wait on each other unexpectedly.