BlockRun MCP is an open-source Model Context Protocol server whose goal is to enable Claude and any MCP-compatible AI agent to obtain real-time data and perform real operations. It provides agents with 19 tools covering market quotes, research, web search, images, video, on-chain data, and Polymarket real-time trading, on a pay-per-call basis.
In terms of main capabilities, according to the official description, it provides access to 88 large models, image and video generation, prediction market data, real-time web and X/Twitter search, on-chain queries across 40 chains, and supports real betting on Polymarket settled in USDC. Among these, blockrun_polymarket is used to place orders and constitutes a real transaction requiring confirmation; most other tools are used to read data. Paid calls are completed through x402 micropayments.
In terms of usage, there are two payment modes, and both lead to the same set of 19 tools. Wallet mode uses a self-custodial wallet, supports USDC on Solana or Base, requires no account or credit card, and uses wallet signatures for identity verification. Account mode uses a BlockRun API Key; first obtain a key by signing in with Google at user.blockrun.ai, then top up credits via credit card or wire transfer. Service calls are by default routed through https://api.blockrun.ai routing. When BLOCKRUN_API_KEY is set, account mode is preferred; when it is unset, it falls back to wallet mode.
For installation, the official recommendation is to run in Claude Code: claude mcp add blockrun -s user -- npx -y @blockrun/mcp@latest. Any MCP client that can start a stdio server can connect using command: npx, args: ["-y", "@blockrun/mcp@latest"]. If using Node from nvm or Homebrew, you can use the absolute path obtained from which npx in the JSON configuration. The Grok client has a default startup timeout of 30 seconds, and the first installation requires downloading the dependency tree, so the official recommendation is to increase startup_timeout_sec or install globally first.
For tool configuration, you can optionally use --profile <name> or set BLOCKRUN_MCP_PROFILE to expose a trimmed tool set and reduce context usage. The official statement says that using --profile trading instead of the default configuration can reduce context cost by 61%. blockrun_surf was removed in 0.49.0 because the gateway returns 410 for all Surf paths; for related issues, use blockrun_price, blockrun_defi, blockrun_markets, and blockrun_dex instead.
Regarding fees and security, in wallet mode the gateway adds a fixed network fee on top of the base data price, currently $0.001 per call on Base. The Solana gateway only reports the base price, and the account channel does not add this fee. The free tier includes blockrun_chat's free mode, blockrun_dex, blockrun_price, and blockrun_models. The official example says $5 can cover approximately 525 market queries, about 500 Exa searches, about 250 image generations, or about 14 Seedance 1.5-pro video clips. When BLOCKRUN_CONFIRM_SPEND=on is enabled, the agent will pause for approval before paid calls exceeding the threshold; BLOCKRUN_BUDGET_LIMIT can serve as a budget cap. Key securityThis machine has ~/.blockrun/.session (permissions 0600). You can optionally enable BLOCKRUN_KEYCHAIN=strict to use the system keychain. BlockRun cannot move users' funds.
Applicable scenarios include: querying prediction market consensus, executing trades based on signals (reading plus placing orders), conducting fund and profit/loss analysis on Polymarket whales, research retrieval with citations, generating poster images with text, speech synthesis and voice outbound calls, multi-agent research with fee caps, and performing raw on-chain reads on 40 chains. Regarding limitations: Polymarket trading still requires local keypair signing even in API Key mode; clients that by default do not show a confirmation dialog for paid calls need to rely on budget cap protection; the official documentation also states that the tool schema consumes context every round. Although prompt caching reduces the dollar cost, the context window cost is still calculated per round.
