ChannelOS

ChannelOS records what I do during the day. Each night, a local language model reads the day and writes a short summary. I log from my phone with one tap. The data and the model stay on my own Linux box.

Parts

Part	What it does
Capture	Records point-in-time events, start and stop activity spans, and Apple Health samples. The log is append-only.
Catalog	Holds 120+ activities in 16 categories. The catalog is data, so a new activity is a new row and needs no deploy.
Ranking	Sorts the quick-log buttons by recency, weighted by time of day. The activity I am most likely to log next is at the top.
Insight agent	Runs each night. It renders the day as text and asks the model for a DailyInsight: a summary, highlights, anomalies, habit notes, and a score. A weekly job does the same for the week.
Client	A SwiftUI app for iOS with an offline queue. Ingest is idempotent, so a retry after a dropped connection does not log an event twice.

The API is FastAPI on async SQLAlchemy 2.0. The data is in Postgres 16, with Alembic for migrations. ARQ on Redis runs the scheduled and on-demand jobs. Everything runs in Docker Compose. The API is reachable only on my private network, and every endpoint except the health check needs a bearer token.

How the insight agent gets structured output

The simple approach is one call: give the model a JSON schema and ask for the answer in that shape. Research on this topic shows that a strict output format during reasoning can make the reasoning worse. Also, function calling on gemma4:26b was not reliable enough for me to put an agent framework in the path.

So the agent makes two calls:

1. The model reads the rendered day and writes about it in free text.
2. A second call converts that text into a flat DailyInsight. Only this call uses Ollama’s structured-output format option.

The job has no tools and one output shape, so a direct async client is enough. On a synthetic test day, the two calls took about 131 seconds in total on the box’s integrated GPU. The job runs once a night, so that time is acceptable.

The agent writes each run to an audit table with its input and output. When an insight looks wrong, I can see exactly what the model received.

What came next

ChannelOS was the first version of an idea. After the logging and the insight agent worked well for me, I built the idea again for the phone as Meontor. Meontor runs its model on the device with Apple Intelligence, so it needs no server. Meontor is the version for other people.

I still use ChannelOS and I continue to develop it. The next steps are a provider-agnostic agent layer and an agent that can act on notes and jobs. ChannelOS is a personal project. I will not release it, and the source is private.