SQD Portal MCP Server

An MCP server that answers blockchain questions from SQD Portal data: transactions, logs, traces, token transfers, wallets, analytics, time series, and candles across EVM, Solana, Bitcoin, Substrate, Hyperliquid, and Tron networks, with an optional in-host Explorer.

The server does not index chains itself. It validates input, plans bounded Portal queries, and returns results with coverage, freshness, pagination, and evidence metadata so an assistant can say exactly what it saw. No SQD account, API key, or client credential is required.

It speaks the stateless MCP 2026-07-28 protocol over HTTP and stdio and keeps the SDK-managed legacy negotiation path for clients still rolling out that revision. Release notes are in CHANGELOG.md. See CONTRIBUTING.md to work on the code and SECURITY.md to report a vulnerability.

Current public surface

• 28 public tools
• 3 advanced/debug tools
• public params use network
• discovery filters use vm
• no legacy tool aliases in v0.8.x

Raw query tools default to compact responses. Ask for response_format: "full" only when you need the larger payload.

Entity questions can use portal_resolve_entity first. It resolves EVM token symbols/addresses, EVM contract aliases, pool identifiers, protocol names, and Hyperliquid coin names into query-ready filters while keeping ambiguous matches explicit.

Token symbol resolution and token metadata come from open token-list data, not baked-in token address constants. Responses include explicit notices when token-list data is unavailable, stale, or unsupported for a network.

Wallet questions should start with portal_get_wallet_summary. It returns fund_flow by default, including inbound/outbound movement, asset flows, counterparties, largest observed movements, and next evidence pivots before raw-tool drill-down.

Tool groups

Discovery:

• portal_list_networks
• portal_get_network_info
• portal_get_head
• portal_resolve_entity

Cross-chain convenience:

• portal_get_recent_activity
• portal_get_wallet_summary
• portal_get_time_series

EVM:

• portal_evm_query_transactions
• portal_evm_query_logs
• portal_evm_query_traces
• portal_evm_query_token_transfers
• portal_evm_get_contract_deployment
• portal_evm_get_contract_activity
• portal_evm_get_analytics
• portal_evm_get_ohlc

Solana:

• portal_solana_query_transactions
• portal_solana_query_instructions
• portal_solana_get_analytics

Bitcoin:

• portal_bitcoin_query_transactions
• portal_bitcoin_get_analytics

Substrate:

• portal_substrate_query_events
• portal_substrate_query_calls
• portal_substrate_get_analytics

Hyperliquid:

• portal_hyperliquid_query_fills
• portal_hyperliquid_get_analytics
• portal_hyperliquid_get_ohlc

Tron:

• portal_tron_query_transactions
• portal_tron_query_logs

Advanced/debug:

• portal_debug_query_blocks
• portal_debug_resolve_time_to_block
• portal_debug_hyperliquid_query_replica_commands

These groups are also the toolsets (discovery, convenience, evm, solana, bitcoin, substrate, hyperliquid, tron, debug). A deployment can trim the catalog with MCP_TOOLSETS or MCP_TOOLS, and an HTTP connection can narrow it further with ?toolsets= or an X-MCP-Toolsets header; see the HTTP deployment notes. With nothing configured the full 31-tool surface is served, and the hosted endpoint keeps that default.

Supported data

• EVM networks indexed by Portal, including Base, Ethereum, Optimism, Arbitrum, Monad, Hyperliquid EVM, and many others
• Tron native transactions (TRX transfers, TRC-10 transfers, contract calls with inline logs and internal transactions) and TVM event logs such as TRC-20 transfers, with Base58 or hex addresses, exact TRX amounts, and the parent transaction hash on every log; the bundled SQD plugin skill documents the raw Stream API for anything beyond that
• Solana mainnet
• Bitcoin mainnet
• Hyperliquid fills and replica commands
• Substrate networks indexed by Portal

Substrate support is currently historical only. It does not have a real-time tail.

Response shape

Most tools return the same envelope in MCP structuredContent and in a compact JSON text fallback for older clients. The envelope contains a normal result body plus shared metadata such as:

• answer
• display
• next_steps
• investigation
• _freshness
• _coverage
• _pagination
• _ordering

investigation is a compact evidence guide for agents: it identifies the primary result path, bounded window, useful pivot fields such as addresses or transaction hashes, follow-up filters, and limitations before a result is treated as complete. Successful material results also carry an _evidence receipt with canonical arguments, a deterministic digest, row reconciliation, source windows, and either exact or semantic replay semantics. Exact receipts pin their evidence window. Semantic receipts disclose that rerunning a moving relative window can return a newer snapshot.

When a response uses estimated, partial, sampled, capped, or paginated data, the top-level answer and metadata disclose it. _pagination.has_more is true exactly when next_cursor is present, and _coverage states whether the requested window was read through (window_complete) and whether the response holds every matching row (result_complete). Safe pagination follow-ups include executable tool-call metadata with explicit cursor arguments; suggestions that cannot be reconstructed safely are marked non-executable.

Chart-oriented tools also return chart and table descriptors so MCP clients or LLMs can render them without reverse-engineering the payload.

SQD Explorer (beta)

SQD Explorer is an MCP App that renders tool results inside hosts that support MCP Apps. It is in beta and off by default. A default deployment answers with structuredContent and compact JSON text only, and no tool result asks a host to open a UI. There are two ways to opt in:

• One connection: add ?app=1 to the endpoint, for example https://portal.sqd.dev/mcp?app=1. Use this to try the beta without changing anything for other users.
• Whole deployment: set MCP_APP_ENABLED=true. A connection can still override it in either direction, so ?app=0 opts a single client back out.

The app resource stays registered either way, so a host can read it directly without anyone opting in.

When it is enabled, compatible hosts receive an inline card sized to its content and a full-screen workspace for 21 data tools: metrics led by the primary number, multi-series and signed-value charts, price candles with linked volume and a fixed readout, ranked and timeline panels that show ten rows with a Show all control, evidence tables that page ten rows with search across every row, explorer links for addresses, hashes, and blocks, chain logos and names from SQD network metadata, continuation controls, current-session history, and JSON or CSV export through the host. Pointer and keyboard inspection expose exact plotted values. Missing buckets remain visible as gaps, identifiers stay unshortened, and any local row cap is separate from server completeness. Failed follow-ups keep the last good result under the error. The app is self-contained and does not use persistent browser storage; its only browser-side requests are chain logo images from cdn.subsquid.io and sqd.dev, the two origins declared in the resource CSP. Hosts without MCP App support receive the same structuredContent and compact JSON text fallback, so the underlying answer never depends on the UI. docs/explorer-design.md records the design rules the app follows.

Three MCP prompts provide reproducible starting points without adding tools:

• investigate-wallet
• investigate-contract
• investigate-market

For a chart-first demo, ask: Show BTC price action and trading volume on Hyperliquid for the past hour, using five-minute candles. Explain whether the final candle is closed. The result opens the SQD Explorer with a candle chart, volume, an evidence table, requested and indexed time bounds, and a receipt. Replay the returned requested_window_start_timestamp and requested_window_end_exclusive as fixed from_timestamp and to_timestamp inputs when you need a stable verification run.

Try it in Claude

The hosted endpoint has the Explorer off, so a new user opts their own connection in:

1. In claude.ai or Claude Desktop, open Settings → Connectors → Add custom connector, enter https://portal.sqd.dev/mcp?app=1, and choose no authentication. The ?app=1 turns the beta on for this connection only. On Claude Desktop you can instead install sqd.mcpb from the latest release and switch on its "SQD Explorer (beta)" setting.
2. Start a new chat and enable the SQD connector for it.
3. Ask a data question. Any of these lands on a tool that carries the Explorer:
   • What has this wallet been doing on Base lately? with an address (wallet summary)
   • Show me recent activity on Ethereum (recent activity)
   • Chart hourly transaction counts on Base for the last day (time series)

The result renders as an inline card instead of a block of text; open the card for the full-screen workspace. Only the 21 data tools carry the Explorer. A catalogue question such as Which networks do you support? calls portal_list_networks and answers in plain text, which is expected rather than a failure.

Install

npm install
npm run build

Run

stdio:

npm start

HTTP:

npm run start:http

Developer discovery

The server exposes a structured tool-selection guide for client builders:

• MCP resource sqd://tools returns grouped tool metadata, examples, starting points, and integration notes.
• MCP resource sqd://tools/{name} returns the guide entry for one tool, for example sqd://tools/portal_get_time_series.

Tool and resource discovery stays on the MCP protocol itself. The server does not maintain a duplicate HTTP catalog.

Codex plugin

The Codex plugin wrapper lives in plugins/portal and defaults to the hosted MCP endpoint at https://portal.sqd.dev/mcp.

Install it from this repo-local marketplace:

codex plugin marketplace add .
codex plugin add portal@sqd

Open a new Codex thread after installing. First-use prompts include Hyperliquid BTC perp fills, recent Base transaction volume, and the latest USDC transfers on Base.

Claude Code plugin

The Claude Code plugin uses the same hosted MCP endpoint and the same public selector:

claude plugin marketplace add subsquid-labs/portal-mcp-server
claude plugin install portal@sqd

Open a new Claude Code session after installing so the SQD MCP tools are loaded.

Grok

Grok chat can use SQD as a custom connector:

1. Open grok.com/connectors.
2. Choose New Connector, then Custom.
3. Enter https://portal.sqd.dev/mcp as the MCP server URL.
4. Leave authentication unset.

Grok Build reads Claude Code plugins directly, so it uses the same package:

grok plugin install --trust subsquid-labs/portal-mcp-server#plugins/portal

ChatGPT

In a workspace with custom MCP apps enabled, open Settings → Apps → Create, enter https://portal.sqd.dev/mcp, choose no authentication, scan the tools, and create the draft app. The server is read-only and does not require user credentials.

Claude Desktop

Download sqd.mcpb from the latest release and open it: Claude Desktop installs the bundle with one click and lists the 31 tools. The bundle carries the server, its production dependencies, and one optional setting, "SQD Explorer (beta)", which is off by default. It needs Node 22 or newer on the machine.

Manual fallback, from a local clone after npm run build, add an entry like this to claude_desktop_config.json:

{
  "mcpServers": {
    "SQD": {
      "command": "node",
      "args": ["/absolute/path/to/sqd-portal-mcp-server/dist/index.js"]
    }
  }
}

Usage notes

• If you do not know the exact network name, start with portal_list_networks.
• If you need recent indexed state, use portal_get_network_info or portal_get_head first.
• If the question is broad, start with portal_get_recent_activity, portal_get_wallet_summary, or portal_get_time_series before dropping to raw queries.
• Time windows accept compact and natural wording such as 30m, past 30 minutes, in the past 1h, in last 38 mins, last hour, or 30 minutes ago.
• Use portal_evm_get_ohlc and portal_hyperliquid_get_ohlc only when you actually need candle-shaped output.
• For large or exploratory queries, prefer response_format: "compact" unless you need the full record shape.

HTTP Deployment Notes

HTTP mode exposes MCP at / and /mcp, liveness at /health, and readiness at /ready. The hosted service exposes the same versioned health response at https://portal.sqd.dev/mcp/health.

• MCP and health endpoints do not require authentication.
• Tool and resource discovery use the MCP protocol; retired /tools and /tools.json routes return 404.
• Set MCP_CURSOR_SECRET on any deployment running more than one process, so a cursor one instance issues is accepted by the next. Unset, each process signs with its own random key, and a cursor stops working across a restart or a load-balanced hop.
• /health reports version and commit, the git commit the image was built from, and every tool result repeats both in _server. Docker Hub tags: latest, X.Y.Z, and X.Y come only from a v* release tag; edge and sha-<commit> come from every main push. Pin a version tag in production.
• /ready is 200 only after the dataset catalog has loaded once and the latest Portal probe succeeded within MCP_READY_MAX_AGE_MS; otherwise it is 503 with a reason and Retry-After. Point orchestrator readiness checks at /ready and liveness checks at /health. The Docker image's HEALTHCHECK uses /ready.
• The server binds 127.0.0.1 unless MCP_BIND says otherwise, and every route checks the Host header (and Origin, when a browser sends one) against an allowlist, so a DNS-rebound page cannot reach a local instance. Loopback hosts and origins always pass; requests without Origin always pass the origin check. A non-loopback bind must set MCP_ALLOWED_HOSTS and MCP_ALLOWED_ORIGINS; if either is missing the server logs a startup error and serves without that check. The Docker image sets MCP_BIND=0.0.0.0, so set both variables in the deployment, or * behind a proxy that already validates them.
• Every request is bounded: headers within MCP_HEADERS_TIMEOUT_MS, the whole request within MCP_REQUEST_TIMEOUT_MS, idle keep-alive within MCP_KEEP_ALIVE_TIMEOUT_MS, and MCP bodies above MCP_MAX_BODY_BYTES are refused with 413 before parsing (411 for a chunked body with no length).

Useful environment variables:

• MCP_CURSOR_SECRET the key pagination cursors are signed with. Set it on any deployment running more than one process. Unset, each process signs with its own random key, so a cursor issued by one instance is rejected by the next and clients lose their place across a restart or a load-balanced hop. It is also what stops a caller minting a cursor for a window the tool would never have offered.
• MCP_TOOLSETS comma-separated toolsets to serve (discovery, convenience, evm, solana, bitcoin, substrate, hyperliquid, tron, debug; all or default for everything). Unknown names are ignored with a startup error. Wins over MCP_TOOLS. Default: all nine, the full 31-tool catalog.
• MCP_TOOLS comma-separated exact tool names to serve when MCP_TOOLSETS is unset.
• Per connection, ?toolsets=evm on the endpoint URL or an X-MCP-Toolsets: evm header narrows the deployment's set for that connection only; it can never add a toolset. Prompts that reference a tool outside the active set are not offered. The active set is a bounded label (all, one toolset name, or custom) on mcp_tool_client_calls_total.
• MCP_BIND interface to listen on, default 127.0.0.1 (0.0.0.0 in the Docker image)
• MCP_ALLOWED_HOSTS comma-separated hostnames accepted in Host (port ignored) on top of loopback; * disables the check. Required for a non-loopback bind.
• MCP_ALLOWED_ORIGINS comma-separated hostnames accepted in Origin on top of loopback; * disables the check. Required for a non-loopback bind.
• MCP_REQUEST_TIMEOUT_MS, MCP_HEADERS_TIMEOUT_MS, MCP_KEEP_ALIVE_TIMEOUT_MS request timing bounds, defaults 120000, 30000, 65000
• MCP_MAX_BODY_BYTES MCP request body cap, default 1048576
• MCP_READY_PROBE_INTERVAL_MS and MCP_READY_MAX_AGE_MS readiness probe cadence and freshness, defaults 30000 and 90000
• MCP_APP_ENABLED to offer the beta SQD Explorer to compatible hosts, default off. Accepts true or 1. Per-connection ?app=1 and ?app=0 override it.
• MCP_TOOL_WEIGHT_BUDGET to bound the combined cost of active tool calls, default 32. Measured profiles allow up to 32 lookups, 4 raw or summary calls, or 2 analytics calls at once while queued work remains cancellation-aware.
• MCP_TOOL_MAX_QUEUE to bound queued tool calls, default 64
• MCP_TOOL_QUEUE_TIMEOUT_MS to bound tool admission wait time, default 5000
• MCP_TOOL_CLIENT_WEIGHT_SHARE percent of the weight budget one caller (one connection, identified by a hashed address) may hold at once, default 50; never below the heaviest single tool so every tool stays schedulable. MCP_TOOL_CLIENT_MAX_QUEUE bounds one caller's queued calls, default 16. A caller over its share gets the retryable overloaded result with reason: client_share while others keep flowing.
• MCP_TRUST_PROXY set to 1 (or to the number of proxies in front of the server) to key fairness on the address those proxies observed instead of the socket address. The header is read only when the immediate peer is itself a trusted proxy, and the hop is counted from the right of X-Forwarded-For, because a caller can write anything to the left of it. The address is hashed and never stored or labelled.
• MCP_TRUSTED_PROXY_PREFIXES a comma-separated list of address prefixes that count as your proxies, matched against the start of the peer address (for example 203.0.113. or 2606:4700:). Setting it replaces the default rather than adding to it, so include loopback or your private range if a co-located proxy also reaches the server. With it unset, loopback and private ranges are trusted, which is where a co-located proxy sits; on a shared private network that means any host on that network can present a forwarded address, so name your proxies explicitly there.
• MCP_SLOW_REQUEST_MS threshold for one JSON line on stderr per slow tool call with admission wait and execution timings and the bounded client family, default 5000.

Cost guardrails

Every scan bound in the server is compiled in and set per tool: a filtered trace scan stops at 5,000 blocks, a contract-deployment search at 1,000,000. Guardrails add a second ceiling above those that a deployment sets from the environment, so an endpoint under load can be turned down without a new image.

• MCP_GUARDRAIL_MODE one of off (default), shadow, or enforce.
• MCP_GUARDRAIL_<CLASS>_<LIMIT> sets one ceiling, where <CLASS> is LOOKUP, RAW_QUERY, SUMMARY, or ANALYTICS, and <LIMIT> is MAX_SCAN_BLOCKS, MAX_WINDOW_SECONDS, or MAX_UPSTREAM_BYTES. For example MCP_GUARDRAIL_RAW_QUERY_MAX_SCAN_BLOCKS=50000.

There are no numeric defaults. A class with nothing set has no extra ceiling, so off and enforce with nothing configured are the same server. The only way a guardrail changes behaviour is if you set a number.

shadow evaluates every ceiling and records what enforcing would have done, without changing a single response. enforce acts:

• A scan over its ceiling stops at the ceiling and reports what it covered through the same partial-coverage path a scan that hits its compiled bound already uses, so _coverage.result_complete becomes false and the response names the blocks it searched. It never claims a complete answer it did not get.
• A query window over its ceiling is refused before anything is fetched, with the cap named in the error and the next steps, because there is no partial result to return for a request that was never allowed to start.

Four counters, all with bounded labels: mcp_guardrail_admitted_total{class}, mcp_guardrail_would_block_total{class,limit}, mcp_guardrail_blocked_total{class,limit}, and mcp_guardrail_fail_open_total{reason}.

Recommended rollout. Set the ceilings you are considering and run shadow for a week. Read mcp_guardrail_would_block_total: it is exactly the set of real requests the ceiling would have cut. If that set is larger than you expected, the ceiling is wrong, not the traffic. Move to enforce once it is the size you intended.

Traces

Metrics say how often and how long; a trace says where the time went inside one call. Traces are off unless you set OTEL_EXPORTER_OTLP_ENDPOINT. With it unset nothing here is imported, allocated, or sent, and the OpenTelemetry packages need not be installed at all.

The SDK is not a dependency of this package. It and its exporter pull about 74 packages against a published tarball of roughly 3.4MB, and almost nobody running this over stdio wants any of it, so they are declared as optional peers. To turn traces on, install them alongside the server and point it at your collector:

npm i @opentelemetry/sdk-node @opentelemetry/exporter-trace-otlp-http

OTEL_EXPORTER_OTLP_ENDPOINT=http://collector:4318

One tool call is one tree:

mcp.request                       (HTTP only: method, transport)
└─ tools/call portal_evm_query_logs
   ├─ mcp.admission               (the wait for a slot)
   ├─ portal.fetch                (one per attempt: dataset, status, bytes, resend count)
   ├─ portal.fetch
   └─ mcp.format_result           (hashing the answer for the evidence receipt)

OTEL_* variables are read by the OpenTelemetry SDK itself, so sampling, headers, batching, and protocol are configured the standard way. The rest:

• A traceparent on the HTTP request, or in the tool call's _meta, makes the call part of the caller's trace rather than starting a new one. The _meta value wins, because it is the more specific claim about which turn this call belongs to.
• Each Portal request carries a traceparent naming its own fetch span, so a Portal-side trace can join this one. Nothing is added when tracing is off.
• The JSON log lines carry trace_id and span_id while tracing is on, so a log event and its span can be looked up from each other.
• /health reports whether tracing is configured, whether it started, and whether argument capture is on.

Span attributes carry no arguments, addresses, hashes, cursors, or free text. They are the tool name, its work class, the dataset, bounded counts, and a bounded outcome. A span leaves the process for a collector the query itself never reaches, so the rule is the one the metric labels follow, and stricter than the logs, which at least stay on your own stderr. MCP_OTEL_INCLUDE_ARGS=1 adds the raw tool arguments to the tool span. It is off by default and unsafe for production: a tool argument is routinely a wallet address, and often a user's own words.

Tests

npm run test:offline builds, lints, typechecks, runs the unit tests, and runs every suite that needs no Portal access. npm run test:live runs the Portal-backed suites. RELEASE_ASSURANCE.md summarises what a release verifies, and scripts/README.md lists every suite.

License

MIT