Documentation

What this is

RaiderRouter is a bring-your-own-key MCP server for market, company, filings, options, macro and portfolio data. You connect your own vendor keys and get one MCP endpoint whose tools activate from the capabilities those keys buy. There is no shared data pool: every call is made with your own key and spends your own allowance.

The tools you see are the ones your account can serve, decided per request from your credentials. Connect a vendor key and its tools appear, and the server sends notifications/tools/list_changed. A tool your account cannot serve is not listed at all rather than listed and failing.

Your first three calls

1. Your AI asks for the surface first: which questions this account can answer today, which are dark, and which source would answer them.
2. Then it asks the question itself. The answer arrives with a record of which vendor served it.
3. Before a long run it checks what is left of your daily allowance, and prices anything it cannot reach yet.

Tools

39 tools. Each line names the tool, the capability it is catalogued under, its metering class, whether it takes a batch argument, and whether it accepts a compact response_mode.

The metering class is the class of the best-ranked source for the tool’s capabilities: the word the router prefers once you hold a key for that source, and mixed when a tool’s capabilities resolve to sources of different classes. Many of these are free on a new account because a keyless source serves them; the word your account actually spends is stamped on each tool in tools/list.

Resources

The long-form contracts are MCP resources, read on demand rather than paid for on every handshake. Read a tool’s guide before acting on a field, and the provenance guide before trusting an absent one.

• raider://guide/provenance - the envelope every answer carries, key by key.
• raider://guide/raw-lane - discover, describe and call a documented vendor endpoint.
• raider://guide/budget - what each source meters, what get_usage reports, how to plan a run.
• raider://guide/coverage - which desk needs a curated tool answers, which the raw lane, which nothing.
• raider://guide/packages - three priced ladders and what each one adds.
• raider://guide/tools/{name} - one tool’s full contract; substitute the tool name for {name}.
• raider://surface - this account’s own live surface, the same JSON describe_surface returns.
• raider://watchlist - this account’s own watchlist, the symbols it asked to keep between sessions; change it with the watchlist tool.

Budgets and credits

Every call is classed by what it spends, in one of five words. The class is a fact about your routing, not about the tool.

• free - a keyless source. Free at the vendor and unmetered here, so it raises no warning and reports no balance.
• budget - a countable allowance that runs out, counted by this service against the vendor’s documented limit. The one worth watching.
• credits - a balance the vendor itself reports on each answer and draws down per call. A call count cannot see it; the vendor’s own figure can.
• rate - a throttle over a window and nothing else. It refills, so nobody runs out of it.
• mixed - a tool whose capabilities resolve to sources of different classes for this account, so one answer may spend two kinds of allowance. Read _provenance.cost to see which.

This service also meters itself: one account gets 60 calls per minute and 2000 per hour over the MCP wire, counting every tools/call, every raider://surface read, every raider://watchlist read and every playground run, while tools/list and the guides are not counted. A call over either window is refused with a sentence opening 'Rate limited:' that names the window, the limit, the count and when the next slot opens, and the live counts are published on get_usage as service_limits.

Call get_usage before a long run, and read raider://guide/budget to plan one. Prefer a batch tool over the same tool called N times, and pass the smallest limit that answers the question.

Packages

Three ladders, priced from the shipped source guides as of 2026-09-06. Each figure is one month at one user, before tax, and each rung includes the rung below it. The full breakdown is at raider://guide/packages, and plan_data_access prices exactly the capabilities you name.

The raw lane

The curated tools answer questions and normalise the answer. For ground nothing curated reaches, three tools walk the documented endpoints of the vendors you hold keys for: list_vendor_endpoints to discover, describe_vendor_endpoint to learn the arguments, and call_vendor_endpoint to call. Read raider://guide/raw-lane first. Where a curated tool covers the same ground it is the better answer, because it carries provenance and the named-absence contract this lane does not.

A raw call spends the same budget a curated call spends: it is one request against the same key. Every request is a GET, the host is pinned per vendor, redirects are not followed, and any path naming a trading, order or money-movement surface is refused by name.

• max_bytes bounds the answer: 32 KB by default and at most 256 KB. A larger value is refused, not clamped.
• fields keeps only the top-level keys or dotted paths you name, inside arrays too, so a long array of wide rows comes back as the same rows with only the columns you asked for.
• Truncation is structural: whole elements are dropped from the end of the largest array until the JSON fits, so what comes back still parses.

Provenance

Every result carries a _provenance block. Read its unavailable list: a field named there was not published, which is not the same as zero. Read upgrades_available for keyed vendors you have not connected that would come earlier in the routing order. The full contract is at raider://guide/provenance.

_provenance always carries: sources, unavailable, upgrades_available, cache, cost.

A shortened answer adds _response. On a curated tool it carries mode, cause, field, available, returned, omitted, selection, where mode is the response mode that actually applied and cause is the control that produced the slice: your compact, limit or offset, or a ceiling or default the tool applied on its own. On a raw call it carries bytes_available, bytes_returned, truncated, fields_kept, tier_note, curated_equivalent. Its absence means nothing was left out. An item or a byte counted as omitted still exists; ask again with a larger bound to see it.

Limits

• A raw call takes at most 40 query arguments, each at most 512 characters.
• A vendor response body larger than 12 MB is refused rather than buffered, which is a separate limit from max_bytes: this one protects the service’s memory, that one protects your context.
• Only vendors you hold a currently working key for are listed and callable, and the check runs again on the call rather than being inherited from the listing.