- Rust 93.1%
- Shell 2.9%
- PowerShell 2.9%
- Jinja 1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| .gitea/workflows | ||
| assets | ||
| ci | ||
| crates | ||
| docs | ||
| .gitattributes | ||
| .gitignore | ||
| Cargo.lock | ||
| Cargo.toml | ||
| clippy.toml | ||
| LICENSE-APACHE | ||
| LICENSE-MIT | ||
| README.md | ||
| rust-toolchain.toml | ||
MCP Agent Toolkit
Two small, native programs that speak MCP over direct stdio. Each is installed on its own, started by the MCP client, and configured on its own:
| Program | Executable | Tools |
|---|---|---|
| Toolkit | mcp-agent-toolkit |
read_frontmatter, write_frontmatter, delete_frontmatter, search_files, uuid, ulid, time |
| Log appender | mcp-agent-log-appender |
append_log |
uuid also answers to the alias guid. The toolkit serves no other tool and
the appender serves nothing but append_log. Both expose tools only: no MCP
resources, no custom instructions and no root registration.
Installation
Download a prebuilt binary
Precompiled release binaries are published at:
Build from source
git clone https://forgejo.jrtv.dk/johnny/mcp-agent-utils.git mcp-agent-toolkit
cd mcp-agent-toolkit
cargo build --release
The release build produces target/release/mcp-agent-toolkit and
target/release/mcp-agent-log-appender (.exe on Windows).
Install to a stable local path
Installing copies the build output to a stable path, so a client can launch it
without locking target/release. Each program has its own application
directory; the installer verifies each copy and reports each program
separately.
ci/install-local.sh # toolkit only (default)
ci/install-local.sh toolkit # toolkit only
ci/install-local.sh appender # log appender only
ci/install-local.sh both # both programs
powershell -File ci/install-local.ps1 # toolkit only (default)
powershell -File ci/install-local.ps1 toolkit # toolkit only
powershell -File ci/install-local.ps1 appender # log appender only
powershell -File ci/install-local.ps1 both # both programs
| Program | Linux | Windows |
|---|---|---|
| Toolkit | ~/.local/share/mcp-agent-toolkit/mcp-agent-toolkit |
%LOCALAPPDATA%\Programs\mcp-agent-toolkit\mcp-agent-toolkit.exe |
| Log appender | ~/.local/share/mcp-agent-log-appender/mcp-agent-log-appender |
%LOCALAPPDATA%\Programs\mcp-agent-log-appender\mcp-agent-log-appender.exe |
On Linux XDG_DATA_HOME replaces ~/.local/share when it is set. Rerun the
installer after each build. The installer never touches your client or
settings files.
Configure an MCP client
Point each server entry's command at the installed program by its absolute
path. There is nothing else to start: no connect step, no background service,
no listener, no hook and no browser.
{
"servers": {
"mcp-agent-toolkit": {
"type": "stdio",
"command": "/home/you/.local/share/mcp-agent-toolkit/mcp-agent-toolkit",
"args": []
}
}
}
Checked-in entries to copy, one file per server and platform (replace you
with your user name; they are examples, not MCP resources):
| Server | Linux | Windows |
|---|---|---|
| Toolkit | docs/clients/linux/toolkit.client.json |
docs/clients/windows/toolkit.client.json |
| Log appender | docs/clients/linux/appender.client.json |
docs/clients/windows/appender.client.json |
Both programs accept the same single option:
mcp-agent-toolkit [--root <dir>]
mcp-agent-log-appender [--root <dir>]
One process, one project
Every process is pinned for its whole life to one project directory: the
directory passed with --root, or else the working directory it was launched
in. Relative paths, project settings and project log destinations all resolve
there. A client's workspace roots do not retarget a running process, and
nothing reconfigures it. To work in another project, register another server
entry (or launch another process) with its own --root, for example:
"args": ["--root", "/home/you/work/other-project"]
Access boundary
Neither program keeps its own list of allowed folders and neither decides access by project membership. Whatever the operating-system permissions of the user running the process allow, the process can do; the permissions are the only boundary. Run each program as a user whose permissions you are comfortable giving the agent.
Settings
Each program has its own global settings file and its own single project file;
the two never read each other's. See the settings guide
for locations, precedence, supported keys, the append_log call and the
guard's limits. Example files for Linux and Windows are checked in under
assets/resources/.
Coming from the earlier single-binary layout? See the manual migration guide. Nothing migrates for you.
Development
cargo build
cargo test
cargo doc --no-deps
License
Licensed under either Apache License, Version 2.0 or the MIT license, at your option.