Official Solana MCP Server

Try it out at https://mcp.solana.com !

The official Solana Developer MCP. Purpose: serve up-to-date documentation across the Solana ecosystem to AI agents and developer tooling.

Architecture

• Ingestion (ingestion/): Databricks notebook crawls the sources listed in ingestion/sources.yaml, chunks markdown, and writes embeddings into a Delta-backed Vector Search index.

• Retrieval (lib/services/databricks/): MCP tools query the index via the Databricks managed AI Search MCP server, which retrieves and reranks results server-side. get_documentation falls back to a SQL read of the docs_chunks Delta table when a source has no published llms.txt.

• Server (lib/index.ts, server/cloudrun.ts): Exposes five tools over MCP — Solana_Expert__Ask_For_Help and Solana_Documentation_Search (semantic RAG), list_sections and get_documentation (canonical-spec retrieval modelled after the Svelte AI server), and program_autofixer for Anchor and Pinocchio Solana program Rust checks. Deployed on Cloud Run as a containerised Node service fronting mcp.solana.com; calls the Databricks workspace REST API directly for retrieval.

• Section catalogue (ingestion/sources.yaml → lib/sources.generated.ts): pnpm gen:sources emits a typed catalogue of every source, its tags from a closed 21-section taxonomy, and use_cases keywords used by list_sections to route the agent.

• Analytics (lib/services/s3/analytics.ts): Tool calls + initializations are buffered in memory and uploaded as JSONL objects to the S3 prefix configured by ANALYTICS_S3_URI.

Related MCP server: solagent-mcp

CLI

cli/ ships @solana/mcp, a dependency-free command-line client for agents and scripts without MCP support. Each command is one tools/call against mcp.solana.com, so it needs no credentials.

npx @solana/mcp search "derive a PDA with Anchor"
npx @solana/mcp ask "why does my CPI fail with PrivilegeEscalation?"
npx @solana/mcp sections
npx @solana/mcp docs anchor-docs frameworks
npx @solana/mcp check programs/vault/src/lib.rs   # or: cat lib.rs | npx @solana/mcp check -

Add --json for the raw tool result, and --framework anchor|pinocchio to check when the framework is not detected from imports. Exit codes: 0 ok, 1 check found a syntax error or an undismissed critical/high issue, 2 usage, network, or server error. Set SOLANA_MCP_URL to point at another server, e.g. http://localhost:8080/mcp.

To release, bump version in cli/package.json, merge, then run the Publish CLI workflow from main (dry run first). It publishes to npm through Trusted Publishing and creates a cli-v<version> GitHub release.

Local Development

pnpm install
cp .env.example .env  # set DATABRICKS_HOST + DATABRICKS_TOKEN + DATABRICKS_VS_INDEX
pnpm dev:local
pnpm inspector  # connects MCP Inspector at http://127.0.0.1:6274

Deploy

Production runs on Cloud Run (mcp.solana.com → server/cloudrun.ts, built via the root Dockerfile). Push to main triggers .github/workflows/deploy-cloudrun.yml, which submits a Cloud Build and rolls the new revision. Runtime env vars (DATABRICKS_HOST, DATABRICKS_TOKEN, DATABRICKS_VS_INDEX, DATABRICKS_WAREHOUSE_ID, ANALYTICS_S3_*, AWS_*) and deploy config (GCP_*, VPC_CONNECTOR) are loaded from the Doppler prd_github config at deploy time.

The Databricks side (databricks.yml) deploys two resources via just deploy:

• the daily ingestion job (crawl_and_index.py notebook) — crawls sources, MERGEs into Delta, syncs the Vector Search index;

• the Lakeview dashboard. Analytics source files land in S3 and require downstream ingestion before dashboard consumption.

Per-environment values (catalog, warehouse, index) live in the gitignored prod.yml (see template inline in databricks.yml).

just deploy   # builds, pushes ingestion job + dashboard

Evals

Per-environment values (catalog, warehouse, index) live in the gitignored prod.yml; supply each variable listed under variables: in databricks.yml.