Pythia by jangles-byte: a local, keyless world-watching oracle for agents

The gap Pythia is aimed at: nobody can watch every feed at once

The README states the problem directly: the world broadcasts its future in news, conflict movements, seismographs, storms, cyber chatter and prediction markets, and the difficulty is not a lack of signal but that no one can watch all of it and reason across it. Pythia is the project's answer, described as an oracle that takes in the live state of the planet and reports what is most likely to happen, with a probability and the reasoning behind it. The intended audience is stated in a callout aimed at agent builders: point an agent at Pythia and it gains what the README calls eyes on the whole planet, one live machine-readable view covering conflict, disasters, markets, displacement, disease, unrest and cyber. That framing matters, because it means Pythia is not competing with a dashboard. It is competing with the collection of scrapers and API wrappers an agent team would otherwise assemble by hand. The pitch is that 40-plus free keyless feeds arrive already fused, with forecasts attached, and that nothing leaves the machine.

Osiris as the eyes, MiroFish as the mind, a local LLM as the voice

The architecture in the README is a three-part pipeline. Osiris supplies the live feeds and the 3D globe: breaking news, GDELT geopolitics, armed conflict, NWS storm and flood warning zones, EONET disasters, wildfires, earthquakes, cyber threats and critical infrastructure, plus Polymarket crowd probabilities as forecasting anchors. The README states that Pythia is built around MiroFish's prediction-engine model, uses MiroFish's configured model as the oracle, and is designed to drive MiroFish's full multi-agent OASIS swarm when a Zep memory key is configured. Out of the box it runs the same model locally and ships its own council of specialist personas that deliberate every prediction and surface consensus and dissent. The third part is the local LLM, which reads the assembled world state and produces the forecast. The data flow is therefore: feeds in, salience-ranked event assembly, persona deliberation, forecast out across four horizons (24h, week, month, year) with a probability and reasoning. Two details in the config show how the fusion is tuned. EVENT_CAP defaults to 500 events per sensing pass and is salience-ranked, with the comment warning that quiet domains such as space weather can be squeezed off the board. SWARM_CONCURRENCY defaults to 1, described as the safe setting for a single heavyweight model, with 4 suggested only alongside a small council model.

Installing Pythia: what the repository actually gives you

This is where the project is weakest. The README as published does not contain an install section, and there are no retrieved releases to pin. What the repository does provide is a pyproject.toml declaring requires-python >=3.11 and dependencies on FastAPI, uvicorn, httpx, python-dotenv, pydantic and mcp, plus a uv.lock, a run-all.sh, a PYTHIA.command launcher and a .env.example. The .env.example is the most reliable setup document available, and it says the defaults work out of the box if Ollama and Osiris are running. So the honest first step is to clone the repository, install the Python dependencies with your chosen tool, copy the example environment file, and start the engine. The engine listens on port 8088 by default and expects Osiris on port 3000.

git clone https://github.com/jangles-byte/Pythia
cd Pythia
cp .env.example .env

The environment file is where the oracle's behaviour is set. The three LLM_ lines are commented out by default, which the comments explain means Pythia reuses MiroFish's configured local model read from ~/MiroFish/.env. Uncomment them to point at any OpenAI-compatible API instead. Note the horizons and the auto-refresh cadence.

OSIRIS_URL=http://localhost:3000
ENGINE_HOST=0.0.0.0
ENGINE_PORT=8088
# LLM_BASE_URL=http://localhost:11434/v1
# LLM_API_KEY=ollama
# LLM_MODEL=llama3.1
HORIZONS=24h,week,month,year
PREDICTIONS_PER_HORIZON=3
LOOP_INTERVAL_SEC=900

With Ollama running and an Osiris instance reachable on port 3000, start the engine. The repository ships run-all.sh for bringing the pieces up together; the uvicorn invocation below reflects the declared dependencies and the ENGINE_HOST and ENGINE_PORT values rather than a command quoted from the README, so treat it as the shape of the start command and check run-all.sh for the project's own version.

./run-all.sh

What you should see is the engine accepting connections on port 8088 and, once the loop is on, forecasts refreshing every 900 seconds. The README describes the outputs as predictions, chat and map overlays, with the whole world-view exposed as an agent API, and the pyproject keywords list mcp, so an MCP client is the intended consumer. The README does not document the endpoint paths, so read engine/ before wiring an agent to it.

Where Pythia will disappoint you

The first limitation is that Pythia is a fusion layer, not a source of truth. Every feed it watches comes from somewhere else, and the README is explicit that Osiris provides the feed layer. If Osiris is not running, or a given upstream feed is down, Pythia has nothing to reason over, and the README does not document how it degrades in that case. Second, the forecast quality is bounded by the local model. The README's own framing is that a local LLM is the voice; a small model voting through a persona council will produce confident-sounding text regardless of whether the underlying signal supports it, which is why the drift charts and the surfaced dissent matter more than the headline probability. Third, the hardware assumption is real. The config comments discuss a single heavyweight model, a 70b persona model, and per-persona vote budgets of 2400 tokens, with a warning that reasoning models truncate their votes if the budget is too tight. Running a council of personas at concurrency 1 on a 70b model is a workstation-class workload, not a laptop one. Fourth, the name is a genuine practical problem: the related searches for this project are dominated by Greek mythology, a clothing brand and a drag performer, so searching for help will mostly return the wrong Pythia. Fifth, and most importantly for adoption, the README does not document rollback, and there are no retrieved releases, so there is no version to pin to if an upgrade changes the forecast format your agent parses.

Pythia versus building the pipeline yourself

The obvious alternative is not another forecasting product but the assembly work Pythia is trying to save you: point your agent at GDELT, USGS, EONET, NWS and Polymarket directly, keep a local model for summarisation, and write your own salience ranking. That approach gives you control over which feeds matter and how they are weighted, and it removes a dependency on Osiris being up. The difference in approach is that Pythia does the fusion and the deliberation for you, and adds the swarm layer: a council of specialist personas that vote and argue, with the Council Chamber UI showing the deliberation vote by vote and persona chat letting you question a specialist in its own voice. If your agent only needs earthquake and storm data, the direct route is less machinery. If it needs a single ranked world state plus a probability and a rationale across conflict, markets, displacement and disease at once, that is the specific thing Pythia claims to provide, and the cost of reproducing it is the cost of writing and maintaining the fusion layer yourself.

Maintenance, upgrade cost and the MIT licence

The licence is MIT, declared both in pyproject.toml and in the repository's LICENSE file, which permits commercial use and modification provided the copyright notice and permission notice are retained. That is the permissive end of the spectrum, and it means embedding Pythia in a commercial agent stack is not blocked by the licence itself. What the licence does not do is give you any support commitment. The repository is not archived, and the last push was on 2026-09-06, so there is recent activity, but there are no retrieved releases, which means there is no tagged version to pin and no changelog to read before upgrading. The upgrade cost is therefore the cost of diffing engine/ and .env.example yourself between pulls, plus the cost of re-reading runs/swarm_models.json if you have customised per-persona models through the UI, since the README states those picks win over the seeds in SWARM_MODELS. One more licence-adjacent point worth checking rather than assuming: Pythia depends on Osiris and MiroFish, each with its own licence, and the README links to both. Verify those separately if you redistribute.

Editorial conclusion

Pythia is for engineers who already run Ollama and want a machine-readable world state plus forecasts their agents can consume without sending data to a cloud service. It is not for anyone who wants a managed, always-on feed: the README gives no install steps, no release history and no rollback path, and the repository has not been pushed to since 2026-09-06. Before adopting it, read engine/ and the .env.example defaults yourself, and confirm that your OSIRIS_URL actually points at a running Osiris instance.

Frequently asked questions

Official sources

1. Issues
2. jangles-byte/Pythia on GitHub
3. License: MIT
4. README