Birla Forecasting MCP

Give your LLM access to Birla Forecasting MCP: validate time-series data, discover models, run forecasts and compare results.

Connect your client

Set up Claude

Open Claude connectors

https://birlaailabs-mcp-customer-auth-test.vercel.app/mcp

Open Customize → Connectors → Add → Custom → Web (or connector settings in your client version).
Name it Birla Forecasting MCP and paste the MCP server URL.
Connect, complete Birla sign-in, then return to your conversation.

Requirements

A Birla account with forecasting access. Claude custom connectors enabled for your account or workspace.

Manual setup

In Claude, open Customize → Connectors → Add → Custom → Web. Add the server URL and select OAuth. Labels may differ by client version.

Complete OAuth and confirm tools in your client. Account limits apply.

Try an example

Try the prompt below in a client that can fetch the linked FRED data. Running a forecast uses your account’s allowance.

Use the Birla Forecasting MCP with the Yug model to backtest a forecast of US egg prices.

1. Fetch the last 10 years of monthly data from FRED: https://fred.stlouisfed.org/data/APU0000708111
2. Hold out the final 5% of months as ground truth. Forecast that same period from the remaining data, with P10, P50, and P90.
3. Plot the actual values and the P50 forecast with the P10 to P90 band shaded, plus enough recent history for context. Trim older history so the forecast period stays clearly visible.

If anything fails, say so instead of guessing.

From data checks to forecast results

Tools perform actions, prompts provide reusable instructions, and resources supply context for your LLM.

Tools 7 actions

Discover capabilitieslist_models: Find model IDs, aliases, sizes, context and horizon limits, covariate support, and platform execution limits.
Check your datavalidate_forecast_dataset: Check prepared CSV for dates, targets, groups, frequency, missing values, duplicate timestamps and usable history before forecasting.
Forecast inline dataforecast_series: Forecast date and value arrays, with optional covariates where the selected model supports them.
Forecast a datasetforecast_csv_dataset: Validate prepared CSV and forecast its series. Review inferred columns, usable rows, dropped rows and warnings.
Start background worksubmit_forecast_job: Submit a durable forecast job with an idempotency key, then retain its job ID.
Retrieve a resultget_forecast_job: Poll the original job for status, a completed forecast or a signed result URL. Polling does not submit another forecast.
Compare outputscompare_forecast_models: Run the same input across two to five model configurations. Comparing forecasts is not an accuracy benchmark without held-out actuals.

Reusable prompts 4 workflows

Choose a prompt in your MCP client to guide a task. Retrieving a prompt returns instructions; it does not run a forecast.

forecast_dataset_help: Validate data, choose an appropriate forecast and explain its uncertainty.
explain_forecast_uncertainty: Explain means and quantiles for a general, technical or executive audience.
compare_forecast_models: Compare model outputs carefully and identify where holdout evaluation is still needed.
business_forecast_summary: Summarize direction, uncertainty and warnings for operations, sales, finance or leadership.

Reference resources 5 forecasting references

Read these MCP resource URIs through your client for input formats, model limits and workflow guidance.

birla://input-format: Inline arrays and CSV structure
birla://frequency-guide: Supported frequency codes
birla://models: Live model capabilities and limits
birla://forecasting-guide: Validation, forecasting and interpretation
birla://agent-skill-guide: File preparation, model selection and job handling

Data requirements

Prepare dates and numeric historical values. Inline arrays must have equal lengths and ascending ISO dates. CSV can include targets, groups and model-supported covariates. For large files, extract the relevant groups and bounded history locally.

Validate before forecasting; review inferred columns, missing values and dropped rows. Specify the horizon as a number of future periods and the frequency when detection is uncertain.

5min · 15min · 30min · H · D · W · M · MS · Q · Y

Results and background jobs

Results contain predictions, model details, warnings and available quantiles. Quantiles express uncertainty, not guarantees. Charts depend on your client.

Any forecast may run asynchronously. Keep its job_id and idempotency_key; retrieve the same job with get_forecast_job. Never submit a new forecast to retrieve a result.

Poll no faster than retry_after, with a five-second minimum by default. Queued responses are not completed results. Large results may use an expiring signed URL.

Failed, cancelled and expired jobs are terminal. execution_may_continue means compute may still be draining; the tool set has no cancellation action.

Troubleshooting

Sign-in or access problems

For 401, refresh OAuth in your client. For 403, check the signed-in account and model or job permissions. Keep credentials out of chat and configuration links. See Authentication.

Rate limits or an uncertain submission

For 429 or 503, respect retry guidance; don’t switch models or keys to evade budgets. If submission is uncertain, keep its idempotency key and request or job ID. Avoid duplicate submissions. See Rate Limits.

Invalid data or model limits

Review dates, frequency, horizon, validation warnings and model capabilities. For support, preserve the error code and request ID without credentials or unnecessary raw data. See API error codes.