← Back to the library
Trading and investing

Databento market data

Retrieve Databento market data through an available MCP server: dataset/schema discovery, symbol resolution, historical bars, quotes, reference data, and bounded batch-download planning. Use for Databento data requests or supplying futures analysis with timestamped observations. Requires a separately configured server, API key, dataset entitlements, and any necessary data budget.

Works with the context you provideVersion 1.0.0

Databento market data

Turn a market-data request into a scoped, cost-aware query and traceable result. This method supplies evidence to trading analysis; it does not operate a broker.

First action and inputs

Discover the actual Databento MCP tools in the host. Read `references/tool-contract.md` before calling a tool. Use supplied context to identify the instrument, venue/dataset, exact contract or symbology, schema, interval, start/end with timezone, intended use, freshness requirement, and any approved data budget. Ask only for a missing choice that changes the query. Resolve continuous futures versus an individual contract explicitly; do not present a continuous series as an executable contract.

If the MCP server is absent or unconnected, return the exact proposed query and missing setup requirement. Do not claim a data fetch. A supplied export may support labeled historical analysis without a connection.

Query workflow

1. Confirm the dataset and date coverage using the available metadata tools. Confirm field and schema support rather than assuming equities, options, or futures are interchangeable. Use symbol resolution for the requested interval. 2. For a bounded historical request, estimate cost with `metadata_get_cost` before retrieval. If the estimate exceeds the user's approved budget or no budget exists for a chargeable request, present the concrete query and cost for approval before fetching. Discovery does not authorize paid data usage. 3. Retrieve the smallest necessary range and schema. Prefer the general `timeseries_get_range` for explicit historical ranges. The upstream shortcuts accept only ES/NQ; do not pass an arbitrary ticker to those shortcuts. 4. Inspect the actual result: error status, row count, timestamps, units, instrument IDs, contract mapping, missing bars, duplicates, and truncation. A successful MCP call is insufficient if its payload reports an API error. Preserve dataset, schema, query, retrieval time, and observed time. 5. For quotes, measure observation age independently from the word "current" in a tool description. The source tries a live endpoint for NY sessions and can fall back to historical data; cached or historical data must be labeled. The seven-day lookback with a result limit does not establish a latest tick. Reject stale, non-finite, inverted bid/ask, or unparseable quotes for a current-price decision. Do not promise real-time delivery from this import. 6. For a large download, present the query, estimated charge, destination, and expected format. `batch_submit_job` creates a provider job and may incur a charge; use it only with explicit approval of that job and budget. `batch_download` may provide download metadata or URLs: inspect its actual output before claiming a local file exists. Obtain host write authority before saving a file and protect provider download credentials.

Output and handoff

Return the result or actual artifact link, instrument/contract, dataset, schema, range/timezone, observation and retrieval times, row count, cost status, and material quality limitations. Separate requested, estimated, retrieved, and saved states. Give a proposed query when blocked. Pass verified bars or quotes to the appropriate sizing, market-context, chart, or backtest skill only when its required fields and frequency are satisfied. Databento does not supply FMP-specific fundamentals, FINVIZ screens, or authenticated account holdings merely because those skills are installed.

Boundaries and recovery

Keep keys in the host environment; never ask for a key in a shared skill file or echo it in a report. On authentication or entitlement failure, report that specific failure and stop retrying alternate paid datasets. On empty, stale, or incomplete data, preserve the limitation and request a smaller/valid range or use adequate supplied observations. Never invent a missing price, execute a broker order, or start recurring data pulls from a one-time analysis request.

The pinned local integration and host configuration are documented in `references/connection.md`. Copying this package does not register a server inside the hosted ChiefOS service or establish a working provider connection.