Read a codebase you did not write

The problem this is for

An agent dropped into an unfamiliar repository fails in several distinct ways, and they need different fixes. Most people reach for one retrieval tool, fix one of them, and conclude that agents are bad at large codebases.

The failure	What it looks like	What answers it
Cannot locate things	greps for a string, finds the changelog	semantic retrieval
No intent, only code	"simplifies" a workaround back into the bug	history and pull requests
Wrong library version	correct-looking call to an API that was removed	live documentation

The pairing worth understanding

Serena and the GitHub server look like they overlap and do not. Serena reads the code as it stands now. It cannot tell you that the odd retry loop on line 340 was added by a pull request titled "fix flaky deploys under load," because that sentence exists nowhere in the working tree. An agent that can only read the current state treats every workaround as an accident.

Context7 is the narrowest tool here and catches the failure that produces confidently wrong output rather than visibly stuck output, which is the more expensive kind.

What is missing from this stack, and why

The obvious fifth item is a bulk repository packer, something that flattens the whole tree into one file for the first ten minutes before you know what to ask Serena about. Repomix is the tool people use for this. It is not here because it has no row in this catalog: it is a CLI tool rather than a skill or an MCP server, and neither of our two ingestion adapters produces it.

That is a gap in our shelves, not a judgment about the software, and saying so is cheaper than quietly shipping a stack with an unexplained hole in it.

What the search changed

Checked 2026-08-13.

This page used to have a fourth server on it. hustcc/mcp-mermaid was here to render what the agent had worked out, on the argument that everything else in the stack is private scratch work that dies with the session.

We then put every stack in this directory through one grounded question: does any public source describe somebody running this combination, rather than these tools individually. The answer for the three above was yes, repeatedly. The answer for the fourth was that it appears with Serena and Context7 in one place and never alongside the GitHub server, and it is the weakest-maintained item of the four.

• Serena + Context7 + GitHub MCP Server, named together in a curated stack.
• Serena + Context7 + GitHub MCP Server again, in a second, independent one.
• Serena + Context7 described as a complementary pair, with neither of the other two present.
• Serena + Context7 + mcp-mermaid, without the GitHub server.

So the trio is the part that other people converge on, and the diagram server was ours alone. It has been cut. The argument for it was good and the evidence was not there, and this catalog's whole position is that those are different things.

One caveat this page owes you: no source found describes all four running as one stack, and none describes anybody running the three as a deliberate set either. What the sources show is the same three servers recommended together in curated lists. That is weaker than a deployment report and stronger than nothing, and it is the level of evidence behind this page.

claude mcp add context7 -- npx -y @upstash/context7-mcp