The context layer for AI agents
A wide upstream response routinely runs past 30,000 tokens. Do that across five sources and the research is over before it starts. swarmsDB sits in front of the data — typed sections instead of JSON dumps, projection instead of truncation, caching and bulk fan-out instead of one call at a time.
# claude_desktop_config.json — no API key changes hands
{ "mcpServers": {
"swarmsdb": { "url": "https://app.swarmsdb.ai/mcp" }
} }{ "key_material": null,
"grant": "workspace-bound",
"revocable": true }curl https://app.swarmsdb.ai/data/openalex/works \
-H "Authorization: Bearer $SWARMSDB_KEY" \
-G --data 'search=graph neural networks' \
--data 'select=id,doi,title,cited_by_count'{ "shape": "table", "rows": 25,
"fields": ["id", "doi", "title", "cited_by_count"] }# one call, many parameter sets — results by signed webhook
POST /bulk/requests
{ "endpoint": "eurostat/data",
"params": ["nrg_bal_c", "demo_gind", "prc_hicp"],
"schedule": "0 6 * * *" }{ "bulk_request_id": "brq_9f2c…",
"webhook": "hmac-signed" }The agent starts strong, then hedges, then repeats itself. Nothing errored. It simply ran out of room to think, three sources into a five-source question.
Cutting a payload at N bytes removes whichever fields happened to sort last. The agent cannot tell a missing field from an absent one, so it fills the gap with a plausible guess.
Knowing which fields matter, letting the caller ask for only those, earning a cache TTL per endpoint, catching upstream drift before it reaches the model — that invisible layer is the product.
The proof
On requests that project a section, smaller payloads — 392 million chars served where the upstream sent 1.64 billion chars. That is the difference between an agent reading one source and reading several within the same context window.
eodhd/eod_bulk_last_day — a live upstream fetch, not a cache replay.
Bar not to scale; at true ratio the served bar is a fifth of a pixel.
The layer
Responses arrive as typed sections an agent can project. Ask for four fields and four fields cross the wire. An oversized payload degrades to a schema plus a preview rather than flooding the window.
select=id,doi,title,cited_by_count A third party connects by grant, not by credential. The grant is workspace-bound and revocable, so access ends when you end it — no key was ever handed over to leak.
"key_material": null Per-key scopes, rate limits and data residency; org budgets enforced on the serving path rather than in a nightly report. An autonomous swarm cannot quietly run up a bill you did not agree to.
402 Payment Required A long cache TTL is justified per endpoint, never assumed from a header. Upstream drift is caught by canaries that assert on content, not status codes — a 200 full of the wrong shape still fails.
ttl: earned · drift: asserted Build vs buy
The wrapper is the easy half, and you will finish it. What breaks agents is everything after it — and it stays invisible until they start stalling and hallucinating.
Knowing which fields matter for a given question, and letting the caller ask for only those. Without it every response is all-or-nothing.
A long TTL has to be justified per endpoint. Trusting an upstream Cache-Control header is how an agent quotes a price from last Tuesday.
A parameter named query that silently means "raw expert DSL" does not error — it returns confident garbage the agent will happily reason over.
Upstream schemas move without warning. Catching it needs canaries asserting on content; a 200 with the wrong shape passes every status-code check you own.
The catalogue
Every source is wrapped the same way — typed sections, projection, earned freshness — so an agent learns one interface, not 454. Anything missing can be added as a custom connector and behaves identically.
Statistics
Markets
Energy & climate
Legal & public
Science
Geospatial
A representative slice. The full 454 are searchable in the console, and any MCP discovery call returns them.
Pricing
You pay per 1M tokens swarmsDB kept out of your context window — the reduction between the upstream payload and the section you asked for. Your unit rate is set by your own 7-day rolling cache-hit rate.
You are paying the latency to populate a cache every later caller benefits from. You warm the network, so you pay the lowest rate.
You get answers in milliseconds off work someone else did. You pay the fair share for it.
Warm the network and you pay less. Ride a warm cache and you get the lowest latency, and pay the fair share for it.
Over budget returns a real 402/429, not a silent truncation.
Discounted or free for students and research groups. Ask.
Connect over MCP or REST. If your agents read more than one source, the context bill is the first thing you will see move.
Measured on our own agent-swarm workload. swarmsDB is battle-tested, not yet widely adopted — external traction is the next milestone.