skipper
MCP server for working in a git repository: local git in-process, and GitHub, GitLab, Gitea, and Forgejo behind one small set of tools.
Why
The existing forge MCP servers expose the API, not a workflow. Counted from their own sources: GitHub's has ~125 tools, GitLab's ~88, Gitea's ~54. One tool per endpoint, three servers, three vocabularies, and an agent that has to work out which one it is talking to.
Skipper registers 20, and shows fewer than that: only the ones this repository can actually use, and the four that write only when you turn them on.
An agent does not need to be emperor of your forge. It does not need to create or destroy an organization on any given turn. It needs to know what this repository is, whether the build passed, and what broke. Skipper exposes that and skips the rest.
By default, no skipper tool can write to a forge. There is no create, delete, close, or fork, and out of the box no merge, push, fetch, or pull either: the only state any tool changes is local (staging, committing, local branches), so an agent cannot move a commit to a remote through skipper.
Merging and syncing are opt-in, in your global config only
(~/.config/skipper/config.toml); a repository's own skipper.toml cannot
turn them on:
[writes]
enabled = true # show pr_merge, git_push, git_pull, git_fetch
confirm = true # default: ask before each merge or push
With confirm, every merge and push is put to you through the client (MCP
elicitation) with exactly what will happen, such as push master to github, origin or rebase #6 "chore: release 0.6.0" (release → master), and runs only
if you approve. A client that cannot ask is refused; confirm = false lets
writes run unattended. Even then nothing forces or deletes: +ref, a:b
and -option refspecs are refused, and a merge is pinned to the head sha it
was checked at, so a push in between fails it.
The destructive operations still exist in the CLIs. To delete a release, an
agent calls gh directly, which your harness can allow or deny with one
rule. That is one binary to gate instead of 267 tools to audit.
How it talks to things
Authentication
Local git needs no credentials: it reads and writes the repository directly, never the network.
For forges, skipper reuses the credentials the CLIs already hold. It stores no tokens of its own. Log in as usual:
gh auth login
glab auth login
tea login add
For GitHub, skipper asks gh auth token for the token once per host (it
honours GH_TOKEN), keeps it in memory only, and calls the API itself: one
HTTP request instead of a gh process per query. A 401 fetches the token
again, so gh auth refresh or a rotated token needs no restart.
For Gitea and Forgejo, skipper reads the token from tea's config instead of
running tea, which can block on an interactive prompt that never resolves
under a non-interactive server. tea login add is still how you set it up.
Requirements
Building from source requires Rust 1.99 or newer. The repository pins Rust 1.99.0 for local and release builds.
Local git needs no binary: it runs in-process on gitoxide.
A missing, outdated, or unauthenticated CLI just hides that provider's tools.
An auth check that fails on the network (timeout, connection reset) is not
treated as logged out: skipper retries it with backoff for a few minutes and
announces the tools with tools/list_changed once it answers. Detection
results are logged to stderr, which MCP clients keep in their server logs.
Tool visibility
Tools follow the repository, not the toolbox. A forge's tools appear only when all of the following hold:
- the workspace is a git repository,
- one of its remotes points at that forge, and
- the forge's CLI is installed, new enough, and authenticated.
Git tools need only a repository; they do not depend on any CLI.
So an installed gh does not put GitHub tools in front of a checkout hosted on
Forgejo. github.com, gitlab.com, and codeberg.org are recognized out of
the box; self-hosted instances are mapped in config:
# ~/.config/skipper/config.toml (global) or ./skipper.toml (per repo)
[hosts]
"github.enterprise.com" = "github"
"gitlab.internal" = "gitlab"
"192.168.3.20" = "forgejo"
Forge names accept aliases: github/gh, gitlab/glab, and
gitea/forgejo/tea (Gitea and Forgejo are both driven by tea). Remotes
whose host isn't mapped simply expose no forge tools. The unmapped hosts are
logged at startup so you know what to add.
Visibility is re-evaluated while the server runs. The remotes are re-read
before every tools/list and tools/call, so git init, git remote add,
re-pointing a remote, or removing one applies to the next call; a background
watcher also emits notifications/tools/list_changed within a few seconds so
clients re-list. No restart needed.
Installation
curl -fsSL https://raw.githubusercontent.com/seuros/skipper/master/scripts/install.sh | bash
It picks the latest release and the right build for your machine, and
installs to ~/.local/bin. Pass --prefix for somewhere else or --version
for a specific tag. Prebuilt for Linux and macOS, x86_64 and arm64.
Or build it yourself:
cargo install skipper
Either way you get the skipper-mcp binary. Skipper is a server, not a
library: the crate builds the binary alone. Depend on the MCP tools and
resources.
MCP setup
Skipper speaks MCP over stdio. Add it to your client config:
{
"mcpServers": {
"skipper": { "command": "skipper-mcp" }
}
}
Tools
Resources
Questions about this repository are resources, not tools, and they are scoped by the workspace's remotes rather than your account.
Current remote
Issue resources read from one remote, picked fresh on every read:
- the remote the checked-out branch tracks (
branch.<name>.remote); - else the only remote;
- else
origin.
With several remotes and none of these, the read fails and says so rather
than guess. Switching branches or running git branch -u moves the next read
to the other forge; no restart. Every issue payload names the forge and
repo it came from. A Gitea/Forgejo token needs the read:issue
scope.
GitHub-only reads (PR checks and discussion, CI runs, pr_watch) take the
current remote when it is on GitHub, else the one GitHub repo among the
remotes, so they keep working while the branch tracks a mirror on another
forge. Two different GitHub repos (a fork and its upstream) with neither
current is an error naming both.
Any PR or issue read can name another repo: ?repo= on the resources
(skipper://pr/7289/comments?repo=upstream, skipper://issues/open?repo=matz/spinel)
and repo on pr_watch, pr_merge and pr_build_wait, as a remote name or
owner/name. Skipper never follows GitHub's fork link on its own: it is
history, not intent (trailblazer/representable is a fork of roxml). A PR
missing from a fork says what the parent is, and the agent picks. A branch's
PR read on another repo is still the one from this checkout's branch. Every
read names the repo it came from, so an empty list is never mistaken for the
wrong repo's answer.
Claude Code never lists MCP resources to its model, so for claude-code
clients skipper appends the resources visible to that session to its server
instructions. Other clients get the plain instructions. Claude Code also
replaces any failed resource read with its own text, so for claude-code the
failure comes back as content naming the reason (#5 is a pull request; read skipper://pr/5/comments). Both quirks are client specs in
src/mcp/instructions.rs; add a version constraint (claude-code < 2.93)
once a release fixes one.
Build monitoring
build_watch blocks until the run's status changes or wait_secs elapses
(default 60), then returns the status with a terminal flag. Call it again
while terminal is false. It polls; it does not rely on the client
surfacing server notifications.
PR check monitoring (GitHub)
pr_build_wait blocks until every check on a PR concludes, or until the
first failure with fail_fast (the default), and returns the overall
conclusion, per-bucket counts, and the failing checks with links. It declares
optional MCP task support. On timeout_secs (default 30 min) it returns the
pending snapshot rather than an error.
The skipper://pr/{number}/checks resource template returns the check
matrix for any PR (checks grouped by workflow with bucket and links), and skipper://pr/current/checks resolves the current branch's PR.
PR watching (GitHub)
pr_watch keeps one watch per server. The first call blocks until
something happens on any watched PR (up to wait_secs), then returns the
events. A pr_watch call made while another one blocks only adds its PR and
returns at once; the blocking call reports that PR's events too. Merged,
closed, and unreadable PRs always wake the call and leave the watch; until
picks which of comment, review, checks, and push also do. Read
skipper://watch for the current state without blocking or consuming events.
It is built not to hammer the API. Every read carries the last ETag, so an
unchanged PR costs a free 304; comments and reviews are fetched only after
the PR itself changed. Each PR polls every 10s, backing off to 2 minutes while
quiet (5 when no call waits) and back to 10s on any change; running checks
cap it at 30s. Retry-After and a low X-RateLimit-Remaining stretch it
further, and a process-wide token bucket caps the request rate. A watch with
no calls or resource reads for 30 minutes stops.
Configuration
Optional: skipper.toml in the repo root or ~/.config/skipper/config.toml
(local overrides global). See skipper.example.toml.
Testing
cargo test # unit tests; git fixtures need `git`, nothing reaches a forge
Contributing
Contributions are welcome for tools that come up in real workflows. The bar is one question: did you need this while actually working, or does it just exist in the API? "Check whether my PR is green" passes. "Transfer repository ownership" does not.
A few things that help a proposal land:
- say which workflow it serves, and how often you hit it
- prefer read-only; if it writes, explain why the CLI isn't good enough
- prefer one tool that answers a question over three that expose endpoints
- scope it to the current repository where that makes sense; skipper's tools follow the workspace's remotes, not your whole account
Creating an organization is banned until the heat death of the universe. Fifteen-plus years of GitHub and that has never once needed to be done programmatically, let alone by an agent mid-task. The same goes for deleting organizations, transferring ownership, and managing billing. If you genuinely need those, the CLIs still have them, and that is exactly the point: they live behind one binary your harness can deny, instead of inside a tool surface you have to audit.
