Context7
Feeds AI the latest docs for libraries and frameworks before it answers, so code it writes does not use outdated, removed functions.
Description from the maker (original English)
List, read, search and (opt-in) write, edit and move files inside one folder, never outside it.
These three are never merged into one score: opinions, measurements and popularity are different things.
We have not written a plain-language summary for this one yet. Below are the facts we read from the registry.
Pick the AI app you use, then copy the settings.
FOLDER_WRITABLEThe maker has not listed install details in the registry. See the maker’s README below.
Runs on your computer; the tool list appears in your AI app after you install it.
We check information anyone can see; we have not read the code. Checks that do not apply are left out of the count. · Checked Nov 6, 2026
The registry name was verified with the GitHub account yarenty, which owns the repo
We have not read this repo from GitHub yet
We have not read this repo from GitHub yet
This one runs on your computer; there is no internet endpoint
Runs on your computer and uses local settings instead
Runs on your computer; we can only read tool lists from internet endpoints
Compared with every active server in this category in the registry. Stars count only when the publisher owns the repo.
It passed 1 of 1 applicable basic checks. We have not read the code. Before adding a key or token, see what access the server asks for and use a key with only the permissions you need.
The registry does not list a required key, though some servers ask you to sign in on first use.
It runs on your computer. Your AI app starts it when needed, and data does not pass through the maker’s servers (unless it calls an online service itself).
The GitHub repo last changed –, and the latest version (0.1.3) reached the registry 4 days ago.
Apps that support MCP, such as Claude Desktop, Claude Code, Cursor and VS Code. You also need Rust on your computer.
In the same category: Context7, Codebase Memory, Exa
Servers in the same category from other makers, sorted by stars.
Reading a card:★ GitHub starsiThree pillars: reviews · checks · usageiUpdated = latest version published to the registryi
Feeds AI the latest docs for libraries and frameworks before it answers, so code it writes does not use outdated, removed functions.
Codebase knowledge graph for AI agents — 162 languages, sub-ms queries, 99% fewer tokens.
Connect AI agents to Exa for web search, content fetching, and multi-step research.
Copied from the maker’s repo as written. PopNx MCP has not reviewed or endorsed it. HTML, images and scripts are removed. Links lead off this site.
An MCP server confined to one folder. You name a directory (or a few); the model can list, read, search, and (if you allow it) write, edit and move files inside it, and cannot reach anything outside it: not with an absolute path, not with .., not through a symlink.
emperor-folder serve --root ~/notes # read-only, Streamable HTTP on 127.0.0.1:8031/mcp
emperor-folder serve --root ~/notes --writable # also write, edit and move emperor-folder serve --root ~/notes --stdio # for desktop clients emperor-folder serve --root notes=~/notes --root code=~/src/app # two roots: notes/..., code/...
Giving an agent "the filesystem" is the fastest way to give it your SSH keys. Most file tools check paths with string prefixes and then open them with the full authority of the process; a symlink, a .. the check missed, or a race between check and open, and the model is out.
emperor-folder does not check-then-open. It opens the root once as a cap-std directory handle and does every file operation *through that handle* with a relative path, so resolution happens under the operating system's (or cap-std's) beneath-the-root rules. There is no code path that opens a file by absolute path.
It is built on emperor-mcp: stateless Streamable HTTP (no session id, load-balancer friendly), output framed so file content reaches the model as data rather than instructions, and a doctor subcommand. It follows Emperor Profile P1.1: no write tool is offered unless the operator switches writing on at start-up, and tool definitions are lean (short descriptions, compact input schemas), so every request carries less of them.
Release binaries for Linux (x86_64), macOS (Apple Silicon and Intel) and Windows are attached to each GitHub release, with a SHA256SUMS file and build provenance attestations: check a download with sha256sum -c SHA256SUMS --ignore-missing and gh attestation verify --repo yarenty/emperor-folder. From source (Rust 1.88 or newer):
cargo install emperor-folder
Or as a container (stdio, the folder mounted at /data, read-only unless you add --writable):
docker build -t emperor-folder https://github.com/yarenty/emperor-folder.git
docker run -i --rm -v ~/notes:/data emperor-folder # read-only docker run -i --rm -v ~/notes:/data emperor-folder --root /data --writable # also write, edit, move docker run -i --rm -v ~/notes:/data -e FOLDER_WRITABLE=true emperor-folder # the same, from the environment
emperor-folder doctor --root ~/notes # check first: exits 1 if something is wrong
emperor-folder serve --root ~/notes
| Flag | Default | Meaning |
|---|---|---|
--root [NAME=]DIR | (required) | A folder the server can see. Repeat for several roots; NAME= names one (default: the folder's own name). See Several roots. |
--bind ADDR | 127.0.0.1:8031 | Streamable HTTP address; MCP endpoint at /mcp, liveness at /health. |
--writable | off | Offer write, edit and move. Without it they are not listed at all. Also set by the environment variable FOLDER_WRITABLE: true, 1 or yes (also on) turn it on; unset, empty, false, 0 or no (also off) leave it off; anything else is refused. |
--stdio | off | Speak MCP over stdin/stdout instead of HTTP. Logs go to stderr. |
--hidden | off | Include hidden files and directories in list and search. |
--no-ignore | off | Do not honour .gitignore, .ignore and git exclude files. |
--max-file-bytes N | 1048576 | Largest file read returns whole, search opens, stat counts lines in, and write accepts. |
--max-results N | 200 | Most entries list returns and matches search returns. |
--max-chars N | 50000 | Most characters of text one read returns (ceiling 200 000); a longer read stops at a line boundary and says where to continue. |
--allowed-host HOST | see below | Accepted Host header, repeatable. |
--allowed-origin ORIGIN | none | Accepted Origin header with scheme (https://app.example.com), repeatable. Requests without Origin (non-browser clients) still pass. |
RUST_LOG sets the log level (default info).
Claude Code
claude mcp add --transport http folder http://127.0.0.1:8031/mcp # server already running
claude mcp add folder -- emperor-folder serve --stdio --root ~/notes # or let Claude Code start it
With --writable (and several roots) in the stdio form:
claude mcp add folder -- emperor-folder serve --stdio --writable --root notes=$HOME/notes --root code=$HOME/src/app
Claude Desktop (claude_desktop_config.json, stdio):
{
"mcpServers": { "folder": { "command": "/usr/local/bin/emperor-folder", "args": ["serve", "--stdio", "--root", "/Users/me/notes"] } } }
[kowalski](https://github.com/yarenty/kowalski) (config.toml):
[[mcp.servers]]
name = "folder" transport = "http" url = "http://127.0.0.1:8031/mcp"
or over stdio:
[[mcp.servers]]
name = "folder" transport = "stdio" command = ["emperor-folder", "serve", "--stdio", "--root", "/home/me/notes"]
Any other MCP client that speaks Streamable HTTP or stdio works the same way.
Seven tools, each a verb. Every path is relative to the root (. or omitted means the root). The three that change files are offered only with --writable; a read-only server lists four.
| Tool | Arguments | Returns | Annotations | ||
|---|---|---|---|---|---|
list | path?, glob?, recursive? (false), max_entries? | Entries with path, kind (file/dir/symlink) and size; truncated when the limit cut it. | read-only, idempotent | ||
read | path, start_line?, end_line? (1-based, inclusive) | The text, with a path lines A-B of N header. PNG, JPEG, GIF and WebP images up to 5 MiB (and --max-file-bytes) come back as MCP image content; other binary files are reported (binary: true, size), never dumped. | read-only, idempotent | ||
search | pattern (regex), path?, glob?, ignore_case? (false) | Matches with path, line number and line (cut at 240 characters); counts of files searched, skipped as binary, skipped as too large. | read-only, idempotent | ||
stat | path | Kind, size, modified time (RFC 3339, UTC), read-only flag; for text files within the size limit, line count; for images, the image type. | read-only, idempotent | ||
write | path, content, mode? (create \ | overwrite \ | append, default create) | Path and bytes written. Creates parent directories. create refuses an existing file. | destructive, --writable only |
edit | path, old_text, new_text, replace_all? (false) | Path, replacements made, bytes written. old_text must occur exactly once (else the error says how many times it occurs) unless replace_all. Text files only. | destructive, --writable only | ||
move | from, to, overwrite? (false) | The new path. Renames or moves inside the root (or between roots); refuses to replace an existing target unless overwrite. | destructive, --writable only |
Every tool sets openWorldHint: false: it touches the named folders and nothing else.
list and search use ignore, ripgrep's walker: .gitignore (even outside a git checkout), .ignore and hidden files are skipped unless --no-ignore / --hidden say otherwise, and symlinks are never followed. Regexes use the regex crate: linear-time matching, no catastrophic backtracking, compiled size capped at 1 MiB. Globs match the path as the tools show it (*.rs matches src/main.rs).
Output. File content from read is framed as a document of unknown authorship (FrameKind::Document); listings, matches and metadata are JSON framed as computed results (FrameKind::Computed). The frame's per-call nonce means a file containing a fake end-of-data marker cannot break out of the block. Refusals come back as tool errors (isError: true) with a message that says what to do instead.
--root can be given more than once. With one root nothing changes: paths are relative to it. With several, every path starts with the root's name, and . is a virtual folder whose entries are the roots:
emperor-folder serve --root notes=~/notes --root code=~/src/app
list . -> notes, code
read notes/todo.md search pattern=TODO -> matches in notes/... and code/... move notes/a.md -> code/docs/a.md (with --writable)
A root without a name is named after its folder (--root ~/notes is notes); two roots with the same name are refused at start-up. In NAME=DIR the name may not contain / or \, so ./a=b is a folder called a=b. Each root is its own cap-std handle: a path, .. or symlink cannot get from one root into another except through move, which names both ends.
What the model can do: see and read files under the roots that the ignore rules do not hide; with --writable, create, edit and rename files under the roots.
What it cannot do:
/etc/passwd, C:\ on Windows) are refused, and so is any path with a .. component, even one that would stay inside (a/../b). That check gives a clear error; the enforcement underneath is cap-std, which resolves every path beneath the root handle (openat2(RESOLVE_BENEATH) on Linux, component-by-component resolution elsewhere) and refuses any resolution that leaves it.read, write, edit, move, list and search alike; a symlink that stays inside works. The walker never follows links, and every path it yields is re-opened through the root handle, so a directory swapped for an escaping link mid-walk is skipped, not read.write, edit and move are removed from the tool list, and the file operations refuse writes as well, until the operator passes --writable (or sets FOLDER_WRITABLE), as Emperor Profile P1.1 requires. Tool annotations (readOnlyHint, destructiveHint) tell clients which tools change files, so they can ask before running them.--max-file-bytes (line-range reads stream and return at most that much) and at --max-chars characters, search skips larger files, list and search stop at --max-results, binary files other than small images are never returned.Host header must be localhost, 127.0.0.1 or ::1 (DNS-rebinding protection). On a non-loopback bind the check is off, as the Emperor Profile expects a validating ingress in front; --allowed-host turns it back on for named hosts, and --allowed-origin refuses browser requests from any other origin.Not in scope: authentication (put the HTTP endpoint behind your gateway or keep it on loopback), and ha
Only the beginning is shown; read the rest in the repo