ui ed20 · api ed20

    MCP Server

    Forecasting MCP Server

    Yeer's MCP server turns demand forecasting into a tool your AI agent can call. Connect Claude, Cursor, or your own agent to ask about your forecasts in plain language — or hand it a raw time series and get a forecast back, with no data onboarding at all.

    Ask your forecast a question

    Planners already know what they want to ask; the dashboard is just where the answer happens to live. With the MCP server the question goes wherever the planner already works — a chat window, an IDE, a scheduled agent — and the answer comes back with the same numbers the dashboard would show.

    “What is our September forecast for plant HH50, and which SKUs drive it?”

    “Which ten material-plant combinations have the worst accuracy right now?”

    “Are we over- or under-forecasting by plant this quarter?”

    “Why is the forecast for this SKU up in Q4?”

    “How does the toothpaste forecast compare with the sales target?”

    “Here is 24 months of history for a new listing — forecast the next six.”

    Two ways to use it

    On your own forecasts

    Your planning data, in your agent

    Thirteen read tools cover the questions a planner actually asks: totals, accuracy, bias, drivers, targets, promotions, and an ad-hoc query over history. Every number comes from the best model per entity, selected by backtested accuracy, so an agent can never quote a weaker model than the team uses.

    Requires a Yeer project with data in it. Keys are scoped to one organization.

    On any time series

    Forecasting with nothing to set up

    The forecast tool takes a series inline — ten points or more, daily through quarterly — picks a zero-shot model, and returns the forecast with prediction intervals. Turn backtesting on and it evaluates candidate models on a holdout and reports the accuracy it measured, so the agent knows how much to trust the number.

    No project, no upload, no model choice. Useful for one-off questions and for agents that generate their own data.

    The toolset

    Fourteen tools, all read-only over your data. There is deliberately no model-selection tool: choosing a model is Yeer's job, not the agent's.

    Discovery

    Start here — every other tool takes a config_id.
    catalog_list_configsLists your forecast projects with granularity, frequency, horizon and status.

    Forecast values

    Always the AI-selected best model per entity — the same numbers as the dashboard.
    forecast_sumTotal forecast for a date range, with top and bottom contributors and override-adjusted totals.
    forecast_by_groupTotal, accuracy and signed bias grouped by a dimension such as plant or material.
    forecast_get_periodsPeriod-by-period forecast for one entity, with p10 / p50 / p90 intervals.
    sales_get_actualActual historical sales, broken down by entity and period.

    Accuracy

    DPA is the accuracy measure the demand planning team already reports.
    accuracy_overallHeadline accuracy across all entities.
    accuracy_rank_entitiesBest or worst material-plant combinations by accuracy.
    dpa_by_categoryAccuracy grouped by a product attribute such as category.

    Context

    The 'why' behind a number, plus the commercial context around it.
    explain_forecastDecomposition drivers: base, seasonality, recent demand and the rest.
    attributes_getProduct metadata for one entity: category, price tier, lifecycle.
    targets_getForecast revenue against sales targets, as a gap analysis.
    promotions_getPlanned promotions for entities and date ranges.
    data_queryAd-hoc natural-language question over the project's history; returns the SQL and the rows.

    Compute

    No stored project needed — works for anyone, on any series.
    forecastForecast a raw time series passed inline. Yeer picks the model, and can backtest it on a holdout to return a reliability estimate plus a plain-language summary.

    Connect in two minutes

    Create a key in the developer console, drop the endpoint into your MCP client, and ask the agent to list your forecast projects. The server speaks Streamable HTTP and authenticates with a bearer token, so there is nothing to install locally.

    Endpoint
    https://yeer.ai/mcp
    MCP client config (Claude Desktop, Claude Code, Cursor)
    {
      "mcpServers": {
        "yeer-forecasting": {
          "url": "https://yeer.ai/mcp",
          "headers": { "Authorization": "Bearer yeer_sk_your_key_here" }
        }
      }
    }
    Verify from the terminal
    curl -X POST https://yeer.ai/mcp \
      -H "Authorization: Bearer yeer_sk_your_key_here" \
      -H "Content-Type: application/json" \
      -H "Accept: application/json, text/event-stream" \
      -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

    Signed in already? Keys, live rates and a per-tool usage log live in the developer console.

    What a call costs, and what it can reach

    Metered per call

    • Read and insight calls cost one credit each.
    • Natural-language queries cost ten, because they run a language model over your data.
    • An inline forecast is priced from the series length and horizon, plus a surcharge when it runs a model bakeoff. The resolved price comes back in the response, so an agent can see what it just spent.
    • Every new organization starts with free credits. The console shows live rates, balance and usage.

    Scoped and read-only

    • Each key resolves to exactly one organization; tools refuse a project that organization does not own.
    • No tool trains, edits, overrides or deletes anything.
    • Rate limits are per organization: 120 calls a minute for read tools, 30 for inline forecasting.
    • Customer data is never used to train our models. See security.

    Already running Yeer?

    The MCP server reads the same projects your team uses on the enterprise platform — the ones behind demand forecasting and demand sensing. Nothing to re-onboard: create a key and the forecasts your planners already trust are available to your agents.

    Frequently asked questions

    What is the Yeer forecasting MCP server?

    It is a Model Context Protocol server that exposes Yeer's demand forecasting as tools an AI agent can call. Point Claude, Cursor, or any MCP client at https://yeer.ai/mcp with an API key, and the agent can read your forecasts, accuracy and drivers in plain language — or forecast a raw time series that has never been uploaded to Yeer.

    Which MCP clients does it work with?

    Any client that speaks Streamable HTTP with a bearer token, including Claude Desktop, Claude Code, Cursor, and custom agents built on the MCP SDKs. The server is stateless, so it sits behind a normal load balancer and needs no session affinity.

    Do I need to be a Yeer customer to use it?

    Not for the forecast tool. It takes a time series inline — at least ten points, daily through quarterly — picks a model, and returns the forecast with prediction intervals, so any agent can forecast without onboarding data. The other thirteen tools read forecasts that already live in your Yeer projects, so those need an account with data in it.

    Which model does it use?

    You never pick one. Every tool returns Yeer's AI-selected best model per entity, chosen by backtested accuracy — exactly what the dashboard shows — so an agent cannot accidentally quote a weaker model than the planning team uses. The inline forecast tool selects a zero-shot model the same way, and with backtesting on it reports the accuracy it measured on a holdout.

    Can an agent change my forecast through MCP?

    No. Every tool that touches your data is read-only. There is no tool to train, edit, override, or delete, and each API key is scoped to one organization, so an agent can only ever see that organization's projects.

    How is it billed?

    Usage-based credits, drawn per call: a read or insight call costs one credit, a natural-language query costs ten, and an inline forecast is priced from the size of the series and the horizon, with a surcharge when it runs a model bakeoff. Every new organization starts with free credits, and the developer console shows the live rates, your balance and a per-tool usage log.

    What are the rate limits?

    Read and insight tools allow 120 calls per minute per organization; the compute-heavy inline forecast tool allows 30. Both are per organization rather than per key, so adding keys for different agents does not change the ceiling.

    Is my data used to train Yeer's models?

    No. Customer data is never used to train our models, and the MCP server adds no exception: it reads forecasts and history that already exist in your tenant and returns them to your agent.

    Put your forecast where your agents are

    Talk to us about access, or create a key and run the first call today.

    Talk to Us
    Live pilot on your data Weeks to value, not quarters Your data never trains our models