MCP server · Convex
Convex MCP Server
Pick it when the agent builds on Convex and should inspect real tables, logs and function signatures; keep the production defaults and use a deploy key to confine it to one dev deployment.
Tools: 12, ~2,376 tokens of tool definitions loaded into the model's context (no Convex login, default tools; measured 2026-10-28, v1.48.0). Compare all servers →
All 12 tools, by token cost
- logs366
- runOneoffQuery334
- status307
- insights300
- data218
- run160
- functionSpec157
- envList137
- envSet118
- envGet99
- envRemove98
- tables82
Same server, other documented flags (measured 2026-10-28):
--disable-tools data,run,envSet(docs example: data, run and envSet disabled): 9 tools, ~1,880 tokens (21% less than the default)
Install
Add `npx -y convex@latest mcp start` as a stdio server (optionally `--project-dir /path/to/project`); in Claude Code or Cursor the Convex plugin starts it for you. Official install docs ↗
Credentials: Your Convex login (`npx convex login`, all projects your account can access) or a scoped CONVEX_DEPLOY_KEY to pin it to one deployment.
What it can access
Works on your Convex deployments: lists deployments, tables and schemas, pages through documents, runs sandboxed read-only queries, reads function specs, logs and 72-hour health insights, runs deployed functions, and lists/gets/sets/removes environment variables. By default production is limited to non-PII read-only tools (insights, tables, functionSpec); `--disable-tools` removes tools.
Minimal safe setup
- Without a deploy key the server uses your
npx convex logincredentials and can reach every project your account can; setCONVEX_DEPLOY_KEYto pin it to one deployment.--project-diralone is not a security boundary. source ↗ - Remove tools with
--disable-tools. Measured: 12 tools, ~2,376 tokens by default; the docs example--disable-tools data,run,envSetleaves 9 tools, ~1,880 tokens. source ↗
Common setup errors
- Tools can't find your deployment, or pick the wrong project.
- Agents call
statusfirst, which uses the current directory. Start the server with--project-dir /path/to/projectwhen the client doesn't launch it from the Convex project. source ↗ data,logsorrunOneoffQueryrefuse to run on production.- By design: production is limited to non-PII read-only tools.
--cautiously-allow-production-piiunlocks PII reads;--dangerously-enable-production-deploymentsunlocks writes and env vars. Prefer a dev deployment. source ↗ insightsis missing.- It isn't available when the server is started with
CONVEX_DEPLOY_KEY. source ↗
How it differs from alternatives
- Supabase MCP Server: Same idea for Postgres-based Supabase projects, with project scoping and a
--read-onlyswitch. Measured: 29 tools, ~4,394 tokens. - Neon MCP Server: Serverless Postgres with branch management; Convex's server instead exposes Convex functions, logs and env vars.
Notes written from the official sources linked above, checked 2026-10-28.
Details
- Type
- MCP server
- Tasks
- Databases & SQLDebugging & observability
- Compatible agents
- Claude Code, Cursor, Codex, GitHub Copilot, Conductor (source: clients named in the official Convex docs (docs.convex.dev/ai/convex-mcp-server: plugin for Claude Code and Cursor, setup pages for Codex, GitHub Copilot and Conductor))
- Maintenance
- Ships inside the official Convex CLI (`npx convex mcp start`): npm convex 1.48.0 published 2026-10-22; source get-convex/convex-backend (npm-packages/convex), last commit 2026-10-28
- License
- Apache-2.0
- Source
- https://github.com/get-convex/convex-backend/tree/main/npm-packages/convex
- Verified on
- 2026-10-28
No independent security audit: entries summarize official sources only. Review permissions and code before installing.
Compared in: Best MCP servers for Debugging & observability in Cursor · Best MCP servers for Databases & SQL in Cursor · Best MCP servers for Databases & SQL · Best MCP servers for Debugging & observability
Badge for your README (maintainers)
The badge only says it's listed; it is not a security review or an endorsement.
