A collection of focused, self-contained Python tools. Most of them wrap a free or low-cost API; a few solve a problem without any network call at all.
There is no framework and no shared runtime. Each agent lives in its own folder, has its own
requirements.txt, and runs on its own.
🎨 Altamira — Visual Illustrator
Text → illustrations via Pollinations.ai, for illustrating the same texts Narrator reads.
cd Altamira && python3 altamira.py🧭 Magallanes — World Cartographer
Turns a text description of a territory into a fantasy map. Gemini designs the logical structure
(locations, types, connections) and emits a visual prompt; Pollinations draws the map. Six visual styles,
and every map is saved with its map_data.json so the structure stays editable.
This is an LLM used as a structured data emitter, not as a prose generator: strict contract, markdown fences stripped, typed schema with enums, and a fallback chain from Pollinations to Gemini Imagen.
cd Magallanes
pip install -r requirements.txt
python3 app.py # http://127.0.0.1:5003⚔️ EncounterComposer — Encounter Generator
Generates a one-page RPG encounter (scene or interlude) and exports it to Foundry VTT (Journal Entry JSON) or Roll20 (handout text).
The design choice worth reading is that this one is deliberately not an LLM. Generation is
deterministic: frozen dataclasses for the input contract, a seeded random.Random, and a cascade of
defaults. Same seed, same encounter — every time. Structured content where you want reproducibility and
no hallucinated stat blocks does not need a model. Export runs the interpolated fields through
html.escape(). Nothing touches disk; you generate and download in the browser.
cd EncounterComposer
pip install -r requirements.txt
python3 app.py # http://127.0.0.1:5010🎭 Charme — Photo → RPG Character
Turns a photo or drawing into an RPG character, 1–4 at a time, with drag-and-drop upload.
- 🔑 Premium: GPT-4 Vision + DALL-E 3 (
OPENAI_API_KEY) - 🆓 Free: BLIP captioning + Pollinations.ai (no key)
cd Charme && python3 app.py # http://127.0.0.1:5001📐 Delineante — Photo → Isometric Blueprint
Turns any image into a hand-drawn isometric technical blueprint.
- 🔑 Premium: GPT-4 Vision + DALL-E 3 (
OPENAI_API_KEY) - 🆓 Free: BLIP captioning + Pollinations.ai (no key)
cd Delineante && python3 app.py # http://127.0.0.1:5000📖 Cronista — Adventure Weaver
Builds an RPG adventure from a world description, then expands it chapter by chapter from your prompts.
Serves without GOOGLE_API_KEY — it just skips generation.
cd Cronista && python3 app.py # http://127.0.0.1:5002🎙️ Bardo — AI Dungeon Master
Runs a text adventure with a persistent session profile, and can listen to it: Vosk transcribes your microphone offline and triggers scene changes on keywords. Web UI plus a voice mode.
cd Bardo
python3 run_realtime.py # voice
python3 web_server.py # web, http://127.0.0.1:5004🔊 sfxDrama — Voice-Triggered Soundboard
A "stream deck" you trigger by talking. Vosk transcribes offline, and a keyword plays its sound effect.
The interesting part is the threading: a sounddevice callback pushes raw bytes into a queue.Queue
and the main thread drains it, because Vosk's recognizer is not thread-safe. It reads Vosk's partial
results while you are still speaking and calls recognizer.Reset() after a match — that is what buys
sub-second reaction instead of waiting for the end of a sentence.
brew install portaudio
cd sfxDrama
pip install -r requirements.txt
python3 sfx_drama.pyKeywords and sounds are configured in soundmap.json.
✂️ Salomon — Text Splitter
Splits a long book on paragraph boundaries into numbered chunks, ready for Narrator.
cd Salomon && python3 salomon.py🎙️ Narrator — Audiobook Generator
Turns text into narrated MP3s using Microsoft Edge's neural voices (free, no GPU, no model download). Reads files or whole folders, six voices across Spanish and English.
cd Narrator && python3 narrator.py🏷️ Charactor — Fantasy Name Generator
Syllable-based fantasy name generator with per-race tables (human, elf, dwarf, halfling, gnome, orc, drow) drawn from the "Marca del Este" setting.
No AI, no network, no API — it is a syllable table and random.choices, and that is the whole point.
Earlier docs described a Gemini + Pollinations character-sheet pipeline here; that code never existed and
the description has been corrected.
cd Charactor && python3 marcaCharMaker.pyThere is no shared library. What exists is a composition pattern: agents are written to hand off files and folders to each other, and any of them can be used on its own.
books/*.txt
│
├──▶ Salomon ──▶ chunks/*.txt
│ │
│ └──▶ Narrator ──▶ *.mp3
│ └──▶ Altamira ──▶ illustrations
│
└──▶ Cronista ──▶ adventures/
▲ │
│ ▼
Magallanes ──▶ locations & connections
Concretely: Salomon chunks a book and Narrator narrates the chunks; Altamira illustrates the same text Narrator reads; Cronista stores the world and its characters, which Bardo then runs as a live session.
The one place where a pattern goes beyond file handoff is Magallanes: it treats the LLM as a typed data source, with a strict JSON contract, markdown fence stripping, enum-validated schema and a fallback chain. That is the pattern to copy for any agent that needs structured output from a model.
Each web agent keeps the same split: Flask only parses the request and renders the result; the actual work lives in a plain Python module that can be imported and tested without a request context.
- Python 3 across the board, no packaging, no monorepo tooling
- Flask for the six web agents, each on its own port (see above)
- Free-tier APIs: Pollinations.ai (images), Google Gemini (text), Microsoft Edge TTS (speech)
- Optional paid APIs: OpenAI GPT-4 Vision / DALL-E 3 (premium mode in Charme and Delineante)
- Offline / local: Vosk (speech-to-text), BLIP via HuggingFace (captioning)
- Deterministic:
random.Random(seed)and frozen dataclasses where output must be reproducible
- There are no tests and no CI. This is a personal toolbox, not a maintained library. The parts that
are worth reusing were written to be readable rather than covered — if you want to build on them,
expect to add your own tests.
EncounterComposeris the exception worth reading anyway: it is the most rigorously structured code here (frozen dataclasses, seeded RNG, engine separated from Flask) precisely because it is the one that had to produce correct output every time. - Every Flask app defaults to
debug=False. SetFLASK_DEBUG=1to enable it. - API keys are optional and live in
.env.local(git-ignored). Without them, agents fall back to free tiers or refuse to start with a clear message.
MIT — see LICENSE.
The agents call third-party APIs (Pollinations.ai, Google Gemini, Microsoft Edge TTS, OpenAI, HuggingFace). This license covers the code in this repository only: use it responsibly and respect the terms of service of whichever API you call.