SQD developer resources
Everything needed to build against SQD, each at a URL that does not move. The SQD Portal API and its OpenAPI specification, how authentication and rate limits work, what the API guarantees and how it announces a change, the SQD MCP servers, the Squid CLI, and the files an agent reads before any of it.
If you are an agent, start with llms.txt and openapi.json. Every page on sqd.dev is also served as Markdown to a request that asks for it.
SQD Portal API
The SQD Portal is a plain HTTP API serving onchain data from 110+ networks, with no node,
archive, or indexer of your own. Responses to the stream endpoints are
JSON Lines, so a client reads them
incrementally instead of buffering a whole range. Start at
GET /datasets: the dataset name every other call takes
comes from that list.
OpenAPI specification
The whole documented Portal surface is described in OpenAPI 3.1 at
sqd.dev/openapi.json. Every
operation carries a stable operationId, typed
parameters, and a response schema, which is what a function-calling adapter or a client
generator needs. It is also the document to diff in a pipeline when you want to detect a
change to the API.
Authentication
The public Portal at https://portal.sqd.dev is shared
capacity, intended for development and bounded evaluation. Authenticated and dedicated
portals take an API key in the x-api-key request
header. Paths and payloads are identical either way, so a client written against the public
Portal moves to a dedicated one by adding one header and changing the host.
Rate limits and retries
When the Portal is at capacity it answers with error type
rate_limit_error and code
overloaded. A
Retry-After header is mandatory on those responses, is
always at least one second, and is always a count of seconds rather than an HTTP date. Honor
it: that value is the throttle signal, and there is no separate quota header to poll.
A transient failure arrives as availability_error
instead, which may also carry Retry-After and is safe to
retry with your own backoff when it does not. Branch on
error.type, match on
error.code, and never parse
error.message.
Browser clients can read the throttle signal: the Portal exposes
retry-after and
x-request-id through CORS, alongside the stream
metadata headers x-sqd-data-source,
x-sqd-head-number,
x-sqd-finalized-head-number and
x-sqd-finalized-head-hash. None of those are on the
Fetch safelist, so without that exposure a browser would see a status code and nothing else.
Quote x-request-id when reporting a problem.
Versioning and deprecation policy
The Portal API carries no version in its URL and no version header. There is no
/v1/. Every client talks to the same surface at
https://portal.sqd.dev, and that works because the
surface grows by addition: new dataset types, new block and transaction fields, and new
datasets appear without changing what an existing request means.
The surface that does move is the dataset catalogue. Datasets are added as networks are
onboarded, and occasionally retired. A retirement is announced in advance on the
Announcements page and in
the developer chat, naming the
affected datasets and a date, so watch that page rather than polling for removals. After the
date, GET /datasets/{dataset}/metadata answers
404 with error.code of
unknown_dataset.
The SDKs are versioned separately, as npm packages following semantic versioning. A breaking change lands in a major release with a migration guide, and the Portal keeps serving older SDK versions: the wire format they depend on is the same additive surface described above.
MCP servers
Two SQD MCP servers, for two different jobs. The SQD Portal MCP server at
https://portal.sqd.dev/mcp exposes onchain queries as
agent tools, across EVM, Solana, Bitcoin, Substrate and Hyperliquid. The SQD Documentation
MCP server at https://docs.sqd.dev/mcp answers
questions from the documentation. Both speak the Streamable HTTP transport and negotiate
protocol version 2025-06-18. Neither needs a key.
One handshake detail decides whether a client connects at all. The Streamable HTTP transport
requires the initializing POST to accept both media types, and a request sending only
application/json is answered
406 Not Acceptable, as is a request with no
Accept header. The exact call that works:
Most MCP clients send that header pair for you. Add the server by URL and the client handles the rest; the setup guides cover Claude, Cursor, VS Code and the rest.
Squid CLI
The Squid CLI is published on npm as
@subsquid/cli and installs one command,
sqd. It scaffolds an indexer from a template, runs it
locally, and deploys and manages it on SQD Cloud, which means an agent or a script can drive
the whole lifecycle without writing an API client first.
SDKs
Two SDKs and the raw API, for three different jobs. The Pipes SDK builds streaming pipelines into ClickHouse, PostgreSQL, BigQuery or Parquet. The Squid SDK builds indexers that serve a generated GraphQL API over PostgreSQL. Calling the Portal API directly needs no dependency at all and works from any language.
Files for agents
The machine-readable surfaces sqd.dev publishes, each at its standard location. The API
catalog is RFC 9727, the MCP server card follows the
SEP-1649 draft,
and the skill index follows the
Agent Skills discovery RFC.
The homepage advertises all of them in a Link response
header, so a client that reads headers finds them without a second request.
Any page on sqd.dev is also served as Markdown to a request that asks for it. There is no
.md suffix to append: negotiation is on the
Accept header alone.
Support
Check Portal status for Portal availability and network data freshness. Technical questions go to the developer chat, where the engineers who maintain the Portal and the SDKs answer. Dedicated portals, volume, and pricing go through the contact form.