Filesystem MCP server in Continue.dev stdio server, YAML config.
Run mcp-server-filesystem as a local stdio process under Continue.dev — a five-line YAML block that gives your Continue agent fenced read/write/search access to chosen directories.
Where GitHub-style remote servers give Continue.dev a URL to call, mcp-server-filesystem (@agent-infra/mcp-server-filesystem, maintained in ByteDance's 36.3k-star ui-tars-desktop repo) is the other kind of MCP server: a local process Continue spawns over stdio, granting file operations on directories you whitelist.
It's a useful pairing for agent workflows that need to read or produce files beyond the open project — build artifacts, notes, datasets — with a hard fence around everything else.
What the server adds to a Continue agent
The tools cover reading and writing files, creating and listing directories, and searching by name or content — all gated by an allowed-directories check on every call. In Continue's agent mode that turns prompts like "read the three newest reports in my data folder and produce a summary file" into pure tool execution.
The complete probed tool list lives on the server's Loomal listing: https://loomal.ai/marketplace/mcp-server-filesystem.
The YAML block
Stdio servers in Continue use command and args instead of a url. Add this to the mcpServers list in ~/.continue/config.yaml (Node 18+ required, since npx fetches and runs the package):
mcpServers:
- name: filesystem
command: npx
args:
- "-y"
- "@agent-infra/mcp-server-filesystem@latest"
- "--allowed-directories"
- "/Users/you/agent-workspace"Project-scoped variant
If the file access only makes sense for one repo — say, an agent that writes into the repo's docs/ folder — put the same block in a standalone file under .continue/mcpServers/ in that project instead of the global config. Continue loads every YAML file in that directory, and since this entry contains no secrets (unlike API-backed servers), it's safe to commit.
One subtlety: each args item is a separate YAML list element. Collapsing the flag and the path into one string ("--allowed-directories /path") works with some arg parsing but is fragile; keep them as two items.
Choose the fence before the first run
Whatever paths follow --allowed-directories are the entire world this server can see. A dedicated agent-workspace directory is a clean pattern with Continue: the agent gets a sandbox to read from and write into, and your dotfiles, keys, and unrelated projects stay structurally off-limits. Add multiple paths as additional list items when one folder isn't enough.
Use absolute paths. The server is spawned by the Continue extension, and a relative path resolves against the extension's working directory, not your project.
Reload and debug
Continue reads MCP config at startup, so after saving, reload the extension or run Developer: Reload Window. The filesystem server should then appear in the agent tool picker with its tools listed.
If it doesn't: indentation first — the dash on name and the two-space nesting under it are where hand-edited YAML breaks. If the entry parses but the server errors, the spawn failed: confirm npx is on the PATH that VS Code sees (launch from a terminal if not) and that the package name carries its @agent-infra/ scope. If tools run but report paths as not allowed, your fence list doesn't contain the path the agent tried — check for ~ that never got expanded.
FAQ
How do I add the filesystem MCP server to Continue.dev?
Add a list item under mcpServers in ~/.continue/config.yaml with command npx and args running @agent-infra/mcp-server-filesystem@latest plus --allowed-directories and an absolute path. Reload the Continue extension and the tools show up in agent mode.
Can I commit this config to a repo for my team?
Yes — drop the block as its own YAML file in .continue/mcpServers/ inside the project. It contains no secrets, so committing is fine; teammates just need Node 18+ for npx to resolve the package on their machines.
What directories should I allow?
As few as the workflow needs. A dedicated agent-workspace folder is the cleanest setup: the agent reads and writes there, and everything else on disk is rejected by the server's allowed-directories check. Avoid allowing your home directory wholesale.
The server errors immediately on load — why?
It's a spawn problem: either VS Code's environment can't find npx (start VS Code from your shell so it inherits PATH) or the package name is mistyped — the @agent-infra/ scope is part of the name. The error detail appears when Continue surfaces the server's stderr.
More MCP servers for Continue.dev.
Find stdio and remote servers with live tool lists on Loomal.