Predictalot

One HTTP service, two model families, zero ceremony.

• Foundation time-series — 5 zero-shot forecasters (chronos-2, timesfm-2.5, moirai-2, toto-1, sundial-base-128m). Hand them a context window, get quantile or sample-path forecasts. No training step. Six modality-specific endpoints under /v1/timeseries/<type>/.
• Model lifecycle — POST /v1/models/unload releases all resident foundation-model weights and runtime caches. A forecast body with "unload": true schedules its model for teardown after concurrent forecasts using that model finish.
• Tabular ML — 9 supervised learners (lightgbm, xgboost, hist-gbt, random-forest, logistic, mlp, svm-rbf, knn, naive-bayes) + 3 meta-learners (calibrated, stacking, diversified). Train on YOUR engineered features, persist server-side by modelId, forecast on the latest snapshot. Under /v1/tabular/.
• MCP — streamable-HTTP tools at /mcp. One named tool per (FM type, model) cell plus per-type ensemble + listing. Tabular endpoints are HTTP-only for now.

Quick start

docker run -d --name predictalot \
  -v $HOME/predictalot-models:/models \
  -e PREDICTALOT_AUTH_TOKENS=changeme \
  -p 8080:8080 \
  psyb0t/predictalot:latest

# Zero-shot FM forecast
curl -s http://localhost:8080/v1/timeseries/univariate/forecast \
  -H "Authorization: Bearer changeme" -H "Content-Type: application/json" \
  -d '{"model":"chronos-2","context":[[10,11,12,13,14,15,16,17,18,19,20]],"config":{"horizon":5}}' | jq

# Train + persist a tabular model on your own features
curl -s http://localhost:8080/v1/tabular/train \
  -H "Authorization: Bearer changeme" -H "Content-Type: application/json" \
  -d '{"modelId":"my-model","backend":"lightgbm","target":[[100,101,99,...]],
       "features":[{"rsi":[55,58,...],"macd":[0.3,0.4,...]}],
       "config":{"mode":"direction","horizon":3,"nEstimators":400}}' | jq

# Then forecast on the latest snapshot
curl -s http://localhost:8080/v1/tabular/forecast \
  -H "Authorization: Bearer changeme" -H "Content-Type: application/json" \
  -d '{"modelId":"my-model","features":[{"rsi":[58],"macd":[0.4]}]}' | jq

# Release all resident foundation-model memory when it is no longer needed
curl -s -X POST http://localhost:8080/v1/models/unload \
  -H "Authorization: Bearer changeme" | jq

Documentation

CHANGELOG.md tracks per-version changes.

Agent integrations

The skill works in any agent that reads .agents/skills/ (such as peen), and installs natively in the clients below.

Claude Code

claude plugin marketplace add psyb0t/agents
claude plugin install predictalot@psyb0t

Claude Code prompts for the predictalot URL and, if auth is enabled, the token — the token is stored in your OS keychain.

Codex

codex plugin marketplace add psyb0t/agents
codex plugin add predictalot@psyb0t

Installed via the marketplace, the skill invokes as $predictalot:predictalot. Codex also picks the skill up automatically, with no install, in any repo containing .agents/skills/ — there it invokes as plain $predictalot.

OpenClaw

The skill is published to ClawHub on every release:

openclaw skills install @psyb0t/predictalot

For MCP clients that speak local stdio, the @psyb0t/predictalot plugin bridges to predictalot's /mcp endpoint:

openclaw plugins install clawhub:@psyb0t/predictalot

Then set PREDICTALOT_URL (and PREDICTALOT_AUTH_TOKENS if the server requires one).

License

Code: WTFPL (see LICENSE). The MCP plugin under .agents/plugins/predictalot/ is MIT (its own LICENSE). Foundation models retain their upstream licenses — chronos-2 / timesfm-2.5 / toto-1 / sundial-base-128m: Apache 2.0; moirai-2: CC-BY-NC-4.0 (non-commercial). Tabular backends use their upstream licenses — lightgbm / xgboost / scikit-learn: permissive. Review each before commercial use.