Content Agent

A local web app that generates on-brand content — Bluesky posts/threads and email newsletters — for one or more brands. Each brand gets its own background, voice, and audience context; each content type has its own generation agent, its own configurable instructions, and its own history so it doesn't repeat itself. Runs against a local Ollama model by default, no API key required.

What it does

• Multiple brands, each with its own tab: name, background, voice, audience, and SkyPilot account ID.
• Bluesky agent: generates a single post or a thread (each post capped at 300 characters), using per-brand hashtags and formatting instructions. Review and edit before you Regenerate, Send to SkyPilot (immediately or scheduled), or Mark as Used.
• Newsletter agent: generates an HTML body plus 5 subject/description pairs for A/B testing, optionally following a brand-specific HTML template. Review and edit before you Regenerate, Copy to Clipboard, or Mark as Used.
• Anti-repetition: the last 20 pieces of content per brand, per content type are kept and fed back into generation so new posts/newsletters don't reuse the same jokes or angles.
• Everything in one SQLite file — brand config, per-agent settings, and content history. Never committed (see .gitignore).

Quickstart

Prerequisites: Python 3.13, uv, and a local Ollama install with the default model pulled:

ollama pull gemma4

Then:

uv sync --group dev

cp .env.example .env
# Edit .env: OLLAMA_BASE_URL is already set for a local Ollama install.
# Add SKYPILOT_API_KEY if you want "Send to SkyPilot" to work
# (https://skypilot.social/static/html/api.html).

uv run uvicorn web.main:app --reload --port 8000

Open http://localhost:8000 — with no brands yet, it lands on the "add a brand" form. The app's SQLite file (content_agent.db by default) is created automatically on first run.

Using it

• Add a brand: click the + tab, fill in name/background/voice/ audience/SkyPilot ID.
• Generate content: pick a content agent in the left column (Bluesky or Newsletter), write a prompt, hit Generate. Newsletter generation runs in two stages (body, then subject lines) and shows progress for each.
• Review: the draft is fully editable before you act on it. Regenerating discards edits and reruns from the original prompt without touching history; the other actions save the (possibly edited) content.
• Brand settings: click the gear icon on a brand's tab to edit its core fields, configure Bluesky hashtags/instructions and newsletter instructions/HTML template, browse or purge past content, or delete the brand entirely.

Configuration

All settings are read from .env (copy .env.example to start). Agent settings use an AGENT_ prefix; provider/integration keys keep their standard names.

Variable	Default	Notes
AGENT_MODEL	ollama:gemma4	Any pydantic-ai model string. Switching to a hosted provider (anthropic:..., openai:..., google:...) needs that provider's API key set too.
OLLAMA_BASE_URL	http://localhost:11434/v1	Required for the default local model — must include the /v1 suffix.
AGENT_DB_PATH	content_agent.db	The single SQLite file holding all brand config and content history.
SKYPILOT_API_KEY	—	Required only for "Send to SkyPilot".
AGENT_LOG_LEVEL	INFO	Standard Python logging level.
LOGFIRE_TOKEN	unset	If set, traces go to Logfire cloud; otherwise they print to the console.
AGENT_JUDGE_MODEL	anthropic:claude-opus-4-8	Used only by evals that don't ship yet — reserved for future LLM-as-judge content-quality checks.

Development

uv run pytest              # unit tests — TestModel, no real model calls, no API key
uv run pytest -m eval      # reliability evals against the real model, see below
uv run ruff check .        # lint
uv run ruff format .       # format

Reliability evals (evals/test_reliability.py): each agent generates 5 times against the real configured model and every attempt must succeed and respect its hard constraints (300-char cap, exactly 5 distinct newsletter subject lines). This exists because a small local model occasionally produces output a much larger hosted model wouldn't — these evals catch a reliability regression (from a model swap, prompt edit, or schema change) before a user hits it. No API key needed against the default local model, but it does need a real, running Ollama with the model pulled, and takes a few minutes to run.

See CLAUDE.md for architecture notes, including how to add a third content agent.

Project structure

db/                    SQLite layer — brands, per-agent settings, content history
  schema.sql / connection.py / models.py / repository.py

agent/agents/
  bluesky.py            Bluesky post/thread generation
  newsletter.py          Newsletter generation (body, then subject lines)

integrations/
  skypilot.py            SkyPilot API client (posting/scheduling)

web/                    FastAPI + Jinja2 + HTMX UI
  main.py, routers/, templates/, static/, jobs.py (background-job polling)

tests/                  Unit tests (TestModel, no real model calls)
evals/                  Reliability evals against the real model (-m eval)

License

BSD 3-Clause — see LICENSE. Built from agent-template.