fieldnotes-farm

A phone app, a command line, records and an MCP server for reading a farm field from the photos of a walk with AI. The app is the main way to use it; the command and the import are fieldnotes.

What does the work:

• Segmenting each photo (crops, weeds, bare soil → cover): SAM 3.1 on the Meta Model API.
• Reading the walk as one field, checking hypotheses, sorting what you said: any OpenAI-compatible chat model. Default Muse Spark on the Meta Model API; or a server in your house (llama.cpp, Ollama, …) with fieldnotes reader glimmer --url http://<host>/v1 --model <name>.
• Writing down your voice recordings: Meta's transcription API.
• Using the records: any MCP client (Claude Code, Claude Desktop, Muse Code, …) through fieldnotes mcp.

Requirements

• Python 3.11 or later
• A Meta Model API key (dev.meta.ai → API keys) for SAM, the voice and, unless the reader is at home, Muse Spark; in MODEL_API_KEY or in ~/.config/fieldnotes/model_api_key
• Internet access for the models (unless the reader is at home), the weather and the land data; the records are written to a folder on your PC

Install

pip install fieldnotes-farm                  # command line, records, MCP server
pip install "fieldnotes-farm[app,voice]"     # plus the phone app (Flet) and m4a/aac voice recordings (PyAV)

Use

fieldnotes --root ~/farm/myfield init --place "<town, region, country>" --country <ISO code> --lat <lat> --lon <lon>
fieldnotes --root ~/farm/myfield lookup                       # public land data → records/lookup.md (Japan in detail, world maps elsewhere)
fieldnotes --root ~/farm/myfield observe <field> <date>       # photos/<field>/<date>/*.jpg → records/<field>/<date>.md + overlays
fieldnotes --root ~/farm/myfield show <field>                 # the latest record
fieldnotes --root ~/farm/myfield mark <field> <date> dominant wrong "clover, not vetch"
fieldnotes --root ~/farm/myfield voice <field> memo.m4a <date>    # what you said in the field → work / seen / marks
fieldnotes --root ~/farm/myfield propose <field> --cut        # where to cut first and how high → records/_proposals/
fieldnotes --root ~/farm/myfield timeline <field>             # every day's lines in order
fieldnotes --root ~/farm/myfield reader glimmer --url http://<host>:<port>/v1 --model <name> --check   # read with a server at home
fieldnotes --root ~/farm/myfield app                          # the phone app on the home network (port 8550)
fieldnotes --root ~/farm/myfield mcp                          # stdio MCP server; --http for an AI on another machine

--root is the folder that holds photos/ and records/ (default: $FIELDNOTES_ROOT, else the current folder). Photos go in photos/<field>/<date>/; the app files them from the phone. Every record is Markdown with YAML front matter; the format is in docs/records.md, so any program or AI can read a farm folder. This package is the code that runs and nothing else; the sample farm, the record format, the manuals and the design notes are on GitHub (links below).

Links

• Manual: https://github.com/aiseed-dev/fieldnotes/blob/main/docs/manual.md (日本語: https://github.com/aiseed-dev/fieldnotes/blob/main/docs/manual.ja.md)
• Record format: https://github.com/aiseed-dev/fieldnotes/blob/main/docs/records.md
• Source, issues: https://github.com/aiseed-dev/fieldnotes

Licence: AGPL-3.0-or-later.