Context7 MCP server in Continue.dev current docs, not training data.
Wire Upstash's Context7 into Continue.dev with one YAML block, so the docs your model reads match the library versions in your package.json.
Context7 fixes a specific failure mode: your model writes code against the version of a library it memorized during training, not the version you actually installed. The server — built by Upstash, sitting at 57.1k GitHub stars — fetches current, version-specific documentation and feeds it into the prompt at generation time.
Continue.dev is configuration-driven, so the whole setup is one YAML block and an extension reload. Here's the exact block, both places it can live, and the failure points worth knowing about.
What Context7 adds to Continue.dev
Continue's chat and agent modes can call MCP tools mid-conversation. With Context7 connected, asking about a library triggers a documentation lookup: the server resolves the library you mean, pulls its current docs, and hands relevant snippets back to the model before it answers. As of mid-2026 the server exposes tools for resolving a library ID and fetching its docs — check the repo for the current list.
The practical effect: fewer hallucinated APIs, fewer answers built on a framework version from two majors ago. The server itself is free and open source; you run it locally over stdio via npm.
Add the YAML block
Continue reads MCP servers from YAML. You have two options: add an mcpServers block to your global config at ~/.continue/config.yaml, or drop a standalone file into .continue/mcpServers/ inside a project if you only want Context7 available there. The package name is @upstash/context7-mcp, and npx runs it without a separate install step.
mcpServers:
- name: context7
command: npx
args:
- "-y"
- "@upstash/context7-mcp"Reload and verify
Continue doesn't always pick up MCP edits live — reload the extension (or run Reload Window in VS Code) so it re-reads the configuration. Then open a chat in agent mode and ask something version-specific, like how to set up middleware in the framework release you're on; you should see a Context7 tool call appear in the conversation.
Context7's live listing on the Loomal marketplace at https://loomal.ai/marketplace/context7 shows its current tool list if you want to confirm what the server exposes before you wire it in.
Troubleshooting in Continue.dev
YAML indentation is the top offender. mcpServers entries are list items — the dash before name matters, and mixing tabs with spaces silently breaks parsing. If Continue still loads an older config.json, note that MCP support lives in the YAML config; migrate before adding the block.
If the server starts but no tools show up: confirm Node is installed and npx is on the PATH that the editor inherits (launching VS Code from a GUI on macOS can give it a slimmer environment than your terminal). And reload again after every edit — a stale extension process is the most common reason a correct config appears to do nothing.
FAQ
How do I install Context7 in Continue.dev?
Add an mcpServers entry to ~/.continue/config.yaml with command npx and args ["-y", "@upstash/context7-mcp"], then reload the Continue extension. For a project-only setup, put the same YAML in a file under .continue/mcpServers/ instead.
Where does Continue.dev keep its MCP configuration?
Globally in ~/.continue/config.yaml, and per-project in .continue/mcpServers/*.yaml. Both are YAML — Continue's older JSON config predates MCP support, so use the YAML form.
Context7 doesn't appear after I edited config.yaml. What now?
Check the YAML indentation first (the entry is a list item under mcpServers), confirm npx runs from a terminal, then fully reload the Continue extension. Continue won't re-read MCP config until it restarts.
Does Context7 need an API key?
Not to run — the server is open source and works without one. Context7 offers an optional API key for higher rate limits; check their docs for current terms.
More MCP servers for Continue.dev.
Browse the live index — every listing shows its real tool list.