Zero-shot forecasting for real systems

See the next operational move before it hits.

Precog is an agent-friendly forecasting capability for operational time-series. Backtest first. Forecast when useful. Keep the uncertainty visible.

Explore the repository See why it matters

Built for: infra, SRE, platform and operations Interfaces: REST, MCP and Python License: MIT application code

Forecasting without the AI theater.

Operational teams already have metrics, dashboards and alerts. Precog adds a different capability: test whether the behavior you care about is actually forecastable, measure the evidence, and only then produce a forecast worth using.

Start from the operational question

Capacity, growth, utilization, workload drift. Precog is a forecasting engine, not a data-source adapter or an all-purpose observability product.

Backtest before trust

Evaluate held-out history, inspect per-target error and interval coverage, and compare with simple baselines before deciding whether the model adds useful signal.

Keep the limits explicit

Stable levels and trends may be forecastable. Sudden deployments, failures and external decisions may not be. A useful answer can be “do not forecast this”.

Designed for adoption

Agent-friendly semantics

MCP exposes forecasting and backtesting as focused semantic capabilities, not infrastructure-specific workflows.

Joint forecasting when the problem is joint

Related targets can be evaluated together even when units or numeric scales differ.

Caller-owned data preparation

Precog does not silently interpolate, truncate or mutate caller data. The meaning of the series stays explicit.

Supported v1 surfaces

Use Precog from the interface that fits your environment. The execution model stays consistent across the supported surfaces.

REST Canonical synchronous execution API

MCP Agent-facing forecast and backtest workflows

Python Sync and async execution clients

A simple discipline.

Precog deliberately keeps the workflow small: obtain the right series, define what “useful” means, backtest the forecasting problem, compare evidence, then forecast only when the result deserves it.

Step 01

Bring the series

Acquire and prepare numeric, aligned, equally sampled observations that represent the problem you actually want to forecast.

Step 02

Declare the objective

Choose a horizon that matches the operational decision instead of optimizing an abstract metric.

Step 03

Backtest the hypothesis

Use held-out evidence, simple baselines and multiple historical cutoffs when the decision matters.

Step 04

Forecast where useful

Adopt the capability where it improves the decision. Keep uncertainty and failure modes visible.

Deployment

Run the real thing.

Start the full Precog capability: the real TimesFM-3 execution engine, the REST Execution API and the MCP semantic interface. Same capability. Choose the deployment path that fits your environment.

Full capability — Docker

# 1. Create network and persistent model cache
docker network create precog
docker volume create precog-models

# 2. Start the real TimesFM-3 execution API
docker run -d \
  --name precog-api \
  --network precog \
  -p 127.0.0.1:8000:8000 \
  -e PRECOG_ENGINE=timesfm3 \
  -v precog-models:/opt/precog/hf \
  ghcr.io/albe83/precog-api:latest

# 3. Start the agent-facing MCP server
docker run -d \
  --name precog-mcp \
  --network precog \
  -p 127.0.0.1:8765:8765 \
  -e PRECOG_API_URL=http://precog-api:8000 \
  -e PRECOG_MCP_TRANSPORT=http \
  ghcr.io/albe83/precog-mcp:latest

# 4. Check readiness
curl http://127.0.0.1:8000/readyz

Try-now channel: latest moves only with stable Precog releases. The first API start downloads the pinned TimesFM-3 weights into precog-models; subsequent starts reuse the cache. For reproducible deployments, pin vX.Y.Z or an immutable image digest.

Full capability — Helm / Kubernetes

# 1. Install the latest stable published chart
helm install precog \
  oci://ghcr.io/albe83/precog-charts/precog \
  --namespace precog \
  --create-namespace \
  --set image.repository=ghcr.io/albe83/precog-api \
  --set image.tag=latest \
  --set mcp.enabled=true \
  --set mcp.service.enabled=true \
  --set mcp.image.repository=ghcr.io/albe83/precog-mcp \
  --set mcp.image.tag=latest

# 2. Access the Execution API
kubectl -n precog \
  port-forward svc/precog 8000:80

# 3. Access MCP
kubectl -n precog \
  port-forward svc/precog-mcp 8765:80

# 4. Check readiness
curl http://127.0.0.1:8000/readyz

Helm resolves the newest published chart when --version is omitted; OCI chart tags themselves are SemVer, not latest. The container images use Precog's stable latest channel. For reproducible production deployments, set --version X.Y.Z and pin the matching image tags or immutable digests. View Helm options →

Adopt Precog

Forecast what matters. Ignore the rest.

If your team already has operational time-series and real decisions to make, you have enough to start. Precog gives agents and engineers a small, explicit forecasting capability instead of another vertical platform.

Open Precog on GitHub Read the MCP guide

Application code is MIT. TimesFM-3 model weights are distributed separately under the TimesFM Non-Commercial License v1.0 and are not included in Precog images.