Summary

Write access across the whole repository surface, with the raw API behind a confirmation string.

The coverage is wide enough that most GitHub work never needs a browser — file commits, branch creation, PR review requests and merges are all here — and the two generic REST tools mean an uncovered endpoint is not a dead end, with the write one requiring an explicit confirmation before it fires. The detail worth planning around is the SHA rule: updating or deleting a file needs the current SHA, so a read always precedes a write. Lazy tool loading exists for a real reason — forty-odd schemas is a lot of context to hand a client that only needs to read a file.

What it is

A broad GitHub server covering repository exploration and the full issue and pull-request lifecycle, with write operations that commit directly rather than only reading.

What you get

Repository metadata, directory listings, a file's decoded contents with the SHA needed to update it, the README found regardless of casing, and code search within a repository
Repositories listed for a user or organization, searched globally, created, or forked
Issues listed, opened in full, created, updated and commented on, with labels created, listed, added, replaced and removed
Pull requests listed and opened with diff stats and merge status, created, updated, given reviewers and merged
Commits listed with filters for path, author, branch and date, one commit read with its complete diff, and two refs compared with ahead/behind counts and file changes
Branches listed with protection status, created from a resolved ref, and deleted
Files created, updated or deleted by direct commit — updates and deletions requiring the current SHA
Releases listed with notes and asset URLs, and new ones created; user and organization profiles read
Two generic tools for anything not covered: an arbitrary REST GET, and an arbitrary REST write guarded behind an explicit confirmation string
Optional lazy tool loading — groups listed, searched and loaded on demand, so the client is not handed every schema at startup

Requirements

A GitHub personal access token in GITHUB_TOKEN or GITHUB_PERSONAL_ACCESS_TOKEN — a classic token with repo for private repositories and write operations, public_repo for read-only public access, and read:user for profile lookups. Passing it as a command-line flag is possible but exposes it to anyone who can run ps on the machine. Run it with npx github-mcp-server-kosta over stdio, from the npm package github-mcp-server-kosta, currently 3.1.0. --transport http serves a single Streamable HTTP endpoint, bound to 127.0.0.1 by default; if you bind anywhere else, set --http-auth-token and consider the host and origin allowlists, with --http-require-auth-on-public-bind refusing startup otherwise. An idle timeout exits the process after a period without requests, adjustable through MCP_IDLE_TIMEOUT_MS with 0 disabling it. In lazy tool mode only the bootstrap tools and the preloaded groups are exposed, defaulting to core and search.

Setup effort

One command plus a key — npm install -g github-mcp-server-kosta, then supply credentials