Skip to main content
The Portal MCP server is experimental. Available tools and behavior may change.
Connect the Portal MCP server to your favorite AI coding clients. MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. The Portal MCP server gives your agent direct access to blockchain data across 130+ networks. Discover datasets, query transactions and logs, analyze wallets, and compare chains.
Using an AI coding client? Install SQD for Claude Code, Codex, Grok Build, Gemini CLI, or Cursor to get the MCP server and agent skills together.

Connect Portal MCP

  • MCP endpoint: https://portal.sqd.dev/mcp
  • Server version: 0.8.4
  • Authentication: none on the public server
Test the connection:
See the Claude Code documentation for more details.

Factuality and completeness in v0.8.4

Portal MCP v0.8.4 checks the data behind an answer instead of treating a successful request as proof that the answer is complete.
  • Verified time windows. Timestamp boundaries are checked against indexed block timestamps across live datasets. The response shows the requested and analyzed window and rejects a future start time instead of silently moving it to the indexed head.
  • Stable evidence identities. Normalized transactions, logs, calls, events, instructions, inputs, outputs, fills, and actions have stable row identifiers. Missing or duplicate identifiers fail the request rather than producing ambiguous evidence.
  • Exact continuation. Cursor pages preserve every accepted row without duplicates or gaps, including multiple matching records in one block.
  • Reconciled totals. Wallet quantities, transaction counts, aggregates, and OHLC summaries use the returned evidence and exact integer units where required. Release checks reproduce material results from direct Portal queries.
  • Bounded, honest failures. Response limits are measured on the actual MCP payload. Upstream overloads, timeouts, and incomplete results return structured retry guidance instead of partial data presented as complete.
  • Current MCP protocol. Stateless HTTP supports MCP 2026-07-28, including server discovery, request metadata, routing headers, and cache hints. Older clients continue through the SDK compatibility path.

Blockchain Activity Explorer

MCP App hosts can show SQD results in the Blockchain Activity Explorer. It presents exact metrics, multi-series charts, price candles with volume, evidence tables, timelines, coverage, freshness, and continuation controls without changing the underlying answer. The explorer covers wallet and contract activity, token flows, network metrics, Bitcoin, Solana, Substrate, Hyperliquid fills, analytics, and price candles. Tron dataset discovery, head, freshness, and timestamp evidence remain available through the MCP metadata tools. Use linked overview, chart, evidence, and investigation sections; inspect exact chart values with a pointer or keyboard; sort and filter evidence; run a range-focused follow-up; retain investigations during the current app session; export received evidence as JSON or CSV; and request the next signed page when one is available. Missing time buckets remain visible as gaps, and local display limits are separate from server completeness. The explorer uses the SQD Design System and stays on a compact dark surface inside both light and dark hosts. App support depends on the client. Clients that support only standard MCP tools still receive the same data. Clients without MCP App support receive the same structured result and text fallback. The server never depends on the visual interface to provide blockchain data.

Reproducible investigations

Three built-in MCP prompts provide guided starting points without adding more query tools:
  • investigate-wallet for a wallet’s bounded activity and evidence trail
  • investigate-contract for interactions, callers, and emitted events
  • investigate-market for fills, candles, volume, and market context
Material successful results include an _evidence receipt. It records the canonical tool arguments, argument and exact-data digests, source and analyzed windows, row reconciliation, completeness, and whether replay uses a pinned evidence window or a moving semantic window. Use the receipt together with _coverage, _freshness, _ordering, and _pagination before treating an answer as complete.

Available tools

The server exposes 28 read-only tools: 25 query and analytics tools plus 3 low-level diagnostic tools. They cover EVM, Solana, Bitcoin, Substrate, and Hyperliquid data. Every tool uses the same pagination, error, and coverage format so the client can tell whether a result is complete or partial.

Discovery & shared

EVM

Solana

Bitcoin

Hyperliquid

Substrate

Debug

Response size and pagination

The MCP query tools return a bounded page of results plus a _pagination.next_cursor when more matching rows are available. The page is sized to fit an AI context window, so one call may return a preview rather than the whole range. Pass next_cursor back until _coverage.result_complete is true, or tighten the filters. Check _coverage.window_complete separately. A complete result page can still cover only the scanned part of a wider requested window. The server keeps this distinction explicit and never silently removes rows to fit a response. Heavy ranges are not hard-rejected. A wide request, such as a large Solana slot range, returns data together with a soft notice that results are capped by limit and the query may still be heavy. There is no fixed slot-range guard that refuses the query. The underlying Stream API behaves differently: it has no per-response page cap and streams the full result at HTTP 200 (bounded only server-side), and you paginate it by resuming from lastBlock + 1 rather than a cursor. The MCP’s paging is a separate convenience layer for AI clients, not a limit on the HTTP API.

Known limitations

Portal MCP is designed for bounded answers inside an AI conversation. Broad, unfiltered scans on dense networks can return partial coverage or reach the request timeout. Check the coverage and pagination fields, narrow the filters, or request the next page. Use the raw Portal Stream API or an indexer for complete exports and recurring workloads.