Ephemeris MCP server: time-series foundation model (TSFM) forecasting for AI agents

Give Claude, Cursor, ChatGPT or any MCP client the ability to forecast numeric time series with prediction intervals: sales, demand, inventory, web traffic, signups, revenue, energy load, prices, sensor readings, infrastructure metrics.

Ephemeris runs a panel of open-weights, zero-shot time-series foundation models (TSFMs) behind one API key, plus an accuracy-weighted ensemble of them:

| Model | Publisher | Use by name | |---|---|---| | Chronos-2 | Amazon | chronos2 | | TimesFM 2.5 | Google Research | timesfm25 | | Toto 2 | Datadog | toto2-313m | | TiRex-2 | NXAI | tirex2 | | PatchTST-FM r2 | IBM Granite | patchtst-fm-r2 | | FlowState r1 | IBM Granite | flowstate-r1 |

Send history, get quantile forecasts back. No training, no feature engineering, no GPU. Name a model, let Ephemeris route to the best fit for your data, or use the ensemble, an accuracy-weighted blend of the panel:

• TIME: level with the top of the leaderboard (MASE 0.639 vs 0.638 for the leader), with the best average MASE rank of 31 models
• GIFT-Eval: CRPS 0.4662 against seasonal naive, better than every model it blends (a few leaderboard entries, including TimesFM-3, score better)

Scored with each benchmark's own harness. Details: ephemeris.cascade.industries/benchmarks.

Tools

| Tool | What it does | |---|---| | get_version | Public hosted server version, nullable build commit and supported artifact schemas; no key or inference | | forecast | Forecast 1 to 64 series in one call: route, ensemble or explicit mode, any quantiles, optional covariates, horizons up to 4096 sampling steps (subject to live model limits and the 120,000 output-value budget) | | describe_forecast | Deterministic interpretation of a saved artifact; no new inference | | plot_forecast | Inline PNG by default, optional SVG, from a saved artifact; no new inference | | list_models | The live panel: health, capabilities, horizon limits, ensemble weights, prices | | get_balance | Spendable credits | | get_usage | Recent requests and what each cost |

Get an API key

Quickest, from a terminal:

npx ephemeris-mcp login

It opens Ephemeris in your browser. Sign in (or sign up), check the page shows the same code as your terminal, and approve. The key is saved to ~/.config/ephemeris/credentials, readable by you only, and the stdio bridge below uses it, with no key in your MCP config. The Ephemeris CLI (ephemeris auth login) and savetokens share the same file. npx ephemeris-mcp logout deletes it; revoke the key itself on the dashboard's API keys page.

Or sign up at ephemeris.cascade.industries and create a key (pc_live_...) in the dashboard. Pay per forecast, no subscription: pricing.

Connect

Remote server (Streamable HTTP): https://ephemeris.cascade.industries/api/mcp, header Authorization: Bearer pc_live_...

Claude Code (plugin: MCP server plus a forecasting skill)

/plugin marketplace add TensorLink-AI/ephemeris-mcp
/plugin install ephemeris@ephemeris

You are asked for your API key once; it is stored in your system's secure credential store.

Claude Code (server only)

claude mcp add --transport http ephemeris https://ephemeris.cascade.industries/api/mcp \
  --header "Authorization: Bearer pc_live_your_key"

Cursor: one click with , then replace YOUR_EPHEMERIS_API_KEY with your key in Cursor's MCP settings. Or add it by hand:

Cursor (.cursor/mcp.json) and most clients

{
  "mcpServers": {
    "ephemeris": {
      "url": "https://ephemeris.cascade.industries/api/mcp",
      "headers": { "Authorization": "Bearer pc_live_your_key" }
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "ephemeris": {
      "type": "http",
      "url": "https://ephemeris.cascade.industries/api/mcp",
      "headers": { "Authorization": "Bearer pc_live_your_key" }
    }
  }
}

Claude Desktop and other clients that only run local (stdio) servers

After npx ephemeris-mcp login, no key goes in the config:

{
  "mcpServers": {
    "ephemeris": {
      "command": "npx",
      "args": ["-y", "ephemeris-mcp"]
    }
  }
}

Or pass a key explicitly with "env": { "EPHEMERIS_API_KEY": "pc_live_your_key" }; it wins over the saved one.

Claude Code, with the saved key: claude mcp add ephemeris -- npx -y ephemeris-mcp

OpenAI Responses API, Anthropic Messages API, Codex, Gemini CLI: see the docs.

Try it

Once connected, ask:

• "Here are my last 18 months of sales: … Forecast the next 6 months with an 80% interval."
• "Forecast next week's hourly traffic from this CSV and identify the highest median forecast step, without claiming the actual peak distribution."
• "Use the ensemble to project daily signups for 90 days; plot the median and the 10th to 90th percentile band."

More in examples/prompts.md. Without MCP, the same forecast is one REST call: examples/rest_forecast.py.

Guides

• How to give Claude, Cursor or any AI agent a forecasting tool: setup for each client
• Forecasting in LangChain, the OpenAI Agents SDK and the Claude API
• Why language models are bad at forecasting numbers, and the split that works
• Which time-series foundation model should I use?
• Chronos-2 vs TimesFM 2.5 vs Toto 2 vs TiRex-2
• Tutorials: store sales, electricity load and solar, ops capacity, sensors and IoT
• All guides

Reference

• Full reference for LLMs: llms-full.txt
• API docs: ephemeris.cascade.industries/docs
• OpenAPI: openapi-m1.json
• Status: ephemeris.cascade.industries/status

The code in this repository (the plugin manifest, skill and stdio bridge) is MIT-licensed. The models keep their own licences, listed on each model page.

Saved forecasts and updates

Retain an idempotency_key before a paid MCP forecast submission. If the outcome is uncertain, retry identical inputs with that key within the server's retention period; a fresh or expired key may incur another charge. Forecasts include an ephemeris.forecast v1 artifact by default. It contains the submitted input history, response, billing information and supplied local context. Save it privately in the client; embedded URNs are not permanent download links or hosted storage.

Pass the complete saved object as artifact to describe_forecast or plot_forecast; neither accepts a local path or URL. Plotting defaults to PNG; use format: "svg" for SVG. These authenticated tools do not run inference or charge for a new forecast. Correcting supplied local labels or plotting a result does not require another forecast. Artifact input is limited to 1 MiB; use the CLI for larger local artifacts. If artifact_status is unavailable, preserve the raw successful result and original request rather than paying to rerun it.

Supply units, timezone, target definition and measurement semantics in context when known. Context reaches the hosted service but is not sent as model input. Ask for missing definitions; do not guess. A horizon counts sampling steps. Quantile bands are nominal marginal intervals: they do not establish calibrated coverage, cumulative intervals, path probabilities or the actual peak distribution. Distinguish observed sales from underlying demand and known future inputs from scenario assumptions. Billing amounts ending in _mc are millicredits (1000 mc = 1 credit).

Hosted MCP updates are deployed centrally. Reconnect or start a new conversation if the client caches its tool list. Existing npm bridges forward the new tools; updating the bridge package is only necessary for changes to the bridge itself. Client transport support and image/attachment rendering vary. The hosted endpoint uses Streamable HTTP; SSE-only clients need compatible transport support.

Version diagnostics: MCP get_version with {} reports the hosted server version, nullable build commit and supported forecast schemas without an API key or inference charge. ephemeris-mcp --version reports this installed npm bridge version offline, not the remote server version. For the standalone CLI, use ephemeris --version or ephemeris version --json. These versions are independent.