Interlinked.

Comparisons · measured July 2026 · 7 min read

Aider vs Claude Code. They find files the same slow way.

Two of the best terminal coding agents, and they feel different in the hand: Aider hands the model a tidy map of your repo before it starts; Claude Code rummages live with the same shell tools you would type yourself. Pick either one and the opening minutes of every task go the same place: not thinking, finding. On a real 4.47-million-file machine that floor takes 93.8 seconds where one shared index takes 16 milliseconds. This is the comparison, and the layer both of them stand on.

The two, defined

One brings a map. One brings a flashlight.

Aider is an open-source, command-line AI pair-programming tool that edits code directly in your local git repository and feeds the model a “repository map” (a ranked outline of your code's symbols) for context. The map is built with tree-sitter and ranked so the most connected definitions rise to the top, giving the model a bird's-eye view of a project it never fully reads.

Claude Code is Anthropic's official command-line agent: it runs in your terminal and explores a codebase with shell-style tools (glob for names, grep for content, read for files), deciding what to open as it goes. No map is prepared in advance; it searches the moment it needs something, then re-reads the growing transcript each turn.

TL;DR: the short version

The real difference. Aider hands the model a ranked repo map of your git project up front; Claude Code explores live with glob, grep and read. One pre-summarizes the repo, one searches it on demand.

The shared floor. Underneath, both are per-repo and both fall back to grep-class scanning for content, bounded to one project, blind to the 90% of a disk that was never in git.

What it costs. On a 4.47M-file machine the same content query takes 93.8 seconds to scan and 16 milliseconds to look up. The floor, not the model, is the bottleneck.

The fix is shared. Point both agents at one live, whole-machine index over MCP and the same lookup answers Aider and Claude Code alike, carrying notes that survive a rename.

The shared floor, drawn

Different habits. The same thing underneath.

A repo map and a live grep look like opposites, but they answer the same narrow question (“where in this one project is the thing I need?”) and both go quiet the moment the answer sits in a sibling repo or a file that was never committed. The fix does not care which agent you prefer. It is one live index of the whole machine that either one can call:

TWO TERMINALS · ONE INDEXAIDERrepo map · tree-sitterCLAUDE CODEglob · grep · readMCPMCPONE SHARED LIVE INDEXnames + contents · every repo · even gitignoredWHOLE MACHINE · ALWAYS WARMTHE EXACT FILEname · 85 µs content · 16 msone lookup answers either agent: no walk, no repo boundary

Aider and Claude Code arrive with different retrieval habits. The index they both plug into is the same one.

Aider brings its tree-sitter repo map; Claude Code brings glob, grep and read. Both connect over MCP to a single index that already holds every filename and every file's contents, across every repo, including the gitignored files a repo map and a cwd grep both miss. One lookup returns the exact file: a name in 85 µs, content in 16 ms.

Feature by feature

Where they differ, and where they don't.

The top of the stack is a real choice: a prepared map versus live exploration, git-native edits versus an agent that also runs your build. The bottom of the stack is not: both are scoped to one project and both scan for content. The last column is what changes when a warm, machine-wide index sits under either one.

DimensionAiderClaude CodeWith a shared index
How it finds codeRanked repo map (tree-sitter symbols)Live glob + grep + readOne indexed lookup
Search scopeThe git repoThe working-directory treeThe whole machine, every repo
Non-git / gitignored filesOutside the mapGrepped only if in the treeIndexed like everything else
FreshnessRepo map regeneratedRe-scanned on every queryLive: searchable ~1 ms after a save
Content searchgrep-class scangrep-class scan (ripgrep)16 ms vs 93.8 s, same query
Cross-repo linksNoNoYes, derived automatically
Memory on a fileThe chat sessionCLAUDE.md you maintainA note that survives rename

Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Content figure is one representative query: 93.8 s scanned versus 16 ms looked up; across a five-query set the indexed engine averaged 58,625× faster. Aider's repo map and Claude Code's tools are as described in each project's public documentation.

The floor, measured

Whichever you type into, the bytes still get walked.

93.8 s → 16 ms
the same content query: a grep-class scan, then one indexed lookup
6m 57s → 16ms
an agent hunting one file by hand, then one call for it
~58% → ~0%
of a paid session's tokens spent finding, not building
~7,200,000×

The distance between hunting for a file and knowing where it is: 6 minutes 57 seconds of an agent grinding against one 16-millisecond call, the length of a coffee break against a single blink. It is the same gap for Aider and for Claude Code, because it belongs to the floor, not to either of them.

A better model does not move this number. The ripgrep comparison shows why: the fastest grep ever written still reads the disk on every query, and that read, not the reasoning, is what sets the wait. Measured on a Ryzen 9 9950X3D, 64 GB DDR5, NVMe, Windows 11, 4.47M files.

Why the floor is shared

Both tools were built for one project, not a whole machine.

A repo map is bounded to the git repository and must be regenerated as code moves; it never contains the sibling repo that ships alongside this one, or the config, log or document that was never committed. Live grep has no boundary problem but the opposite one: it keeps nothing warm, so it re-walks the tree on every single query. Different failure, same edge: neither holds a ready view of the whole machine.

An agent's questions do not respect folder boundaries: the answer is often one repo over, or in the 90% of the disk that git never tracked. That is why file retrieval is the backbone of the LLM economy: the layer every agent stands on, and the one nobody made fast. Aider and Claude Code inherit its speed, whatever their model.

The fix, shared by both

One index. Either agent calls it.

Interlinked keeps a live index of every file on the machine (names and contents) and exposes it over MCP, the protocol both of these agents can speak. One signed install auto-configures 19 AI clients, Claude Code among them, and any MCP-capable terminal agent points at the same server. The index sits warm at about 44 MB, updates as you save (~1 ms typical, under 30 ms worst case), takes under 1% of the drive, and every hit can carry notes and relationships the engine derived itself, keyed to the file's OS identity, so a note survives a rename.

Same query · Aider or Claude Code

On the shared floor

Content query
93.8 s
Scope
one repo
Tokens on search
~58%

On one shared index

Content query
16 ms
Scope
whole machine
Tokens on search
~0

A per-repo scan replaced by a single machine-wide lookup, and it answers whichever agent made the call.

1.7 s
day one, a fresh machine's 12 of 12 repos were webbed: 26,958 relationship pairs from 8,455 git events, no failures
435 files
two repos it found silently sharing identical files, a link no in-repo map or grep can see
19
AI clients configured by one install: Claude Code, Cursor, Codex, Copilot, Zed and the rest, over MCP

THE AGENT AT THE TOP IS YOUR CHOICE. THE INDEX AT THE BOTTOM CAN BE THE SAME ONE.

Questions people ask

Aider vs Claude Code, answered.

What is the difference between Aider and Claude Code?

Aider is an open-source, git-native pair-programmer: it builds a ranked repo map of your project's symbols and edits files directly, committing as it goes. Claude Code is Anthropic's official CLI agent that explores a codebase live with glob, grep and read. Both live in the terminal; the deepest thing they share is how they retrieve: per repo, backed by a scan.

How does Claude Code find files?

It has no persistent index, so it explores on demand: list a directory, grep for a word, read a candidate, decide, repeat. On a real 4.47-million-file drive we watched that loop run 71 times over 6 minutes 57 seconds before it opened the right file. The model is fine; the tools underneath re-walk the disk each query.

Do Aider and Claude Code use grep or ripgrep?

Effectively, yes: both reach for grep-class scanning for content. Claude Code's content tool is built on ripgrep, and Aider augments its repo map with search when the map is not enough. ripgrep is the fastest grep there is and still scans the drive every query: on one machine that was 93.8 seconds against 16 milliseconds for a prebuilt index on the same query.

Which finds files faster, Aider or Claude Code?

On whole-machine questions, neither wins by much, because they stand on the same retrieval floor: a per-repo view plus a live scan. The bottleneck is that floor, not the agent on top of it. Give both the same warm, machine-wide index and the comparison stops mattering: the lookup returns in the microsecond-to-millisecond range for either one.

Can Aider and Claude Code share the same file index?

Yes. A live index exposed over MCP answers any agent that connects to it. One signed install auto-configures 19 AI clients, Claude Code among them, and any MCP-capable terminal agent points at the same server, so the same index that answers Claude Code answers the next agent you try.

Give both agents the same floor

Keep your agent. Change the floor.

1

Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed and the rest. No JSON to edit.

2

Point any terminal agent at it. Claude Code is wired up automatically; any MCP-capable agent connects to the same index over the same protocol. Bring the tool you already like.

3

Let it index the whole machine. Filename search works immediately; content fills in behind it, across every repo and the gitignored files a repo map and a cwd grep both miss.

4

Let the map and the grep retire. Both trade a per-repo, per-query scan for one warm lookup that returns the exact path, and stop spending the token budget on directory listings.

Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.

Measured July 2026 · Ryzen 9 9950X3D · 4.47M files · Windows 11← All research