Interlinked.

Guides · Setup · Measured July 2026

You installed Claude Code. Now build the stack around it.

Out of the box, the agent can read files, edit them, and run a shell. Everything else (memory across sessions, live docs, your database, and search that actually covers the machine) you add. This is the short list worth adding, the same one that pays off in Cursor, and the single tool most setups are still missing.

FINDRECALLFETCHSHIPQUERYRUN

The stack, one picture

Six jobs the agent can't do alone.

An agent is only as good as the tools you hand it. Put it at the hub and every capability it lacks is a spoke. Below is the stack we'd wrap around Claude Code or Cursor today, with the one spoke most people leave empty drawn in blue.

YOUR AGENTClaude Code · Cursorreads · edits · runs a shellFINDEvery file and its contents,across the whole machine→ an indexed search MCPMOST SETUPS SKIPRECALLWhat was decided lastsession, not re-explained→ a memory serverFETCHLive docs and the open web,past the training cutoff→ a docs / fetch serverSHIPIssues, PRs and CIwithout leaving the chat→ the GitHub MCPQUERYSchema-aware readsstraight from your DB→ a Postgres serverRUNSharpen the shell italready reaches for→ rg · fd · jq · gh

Every spoke is a real capability the raw agent lacks. The service spokes (SHIP, QUERY, FETCH, RECALL) are each a one-line install. Our bias on order: FIND first. It is the one the agent leans on hardest and the cheapest gap to close.

The short list

What to install, and why.

CapabilityWhat it addsWhat to installWhy bother
FINDEvery file + its contents, whole machineInterlinked FilesThe gap Claude Code feels most on a big drive
RECALLDecisions and context that outlive the sessionA memory serverStop re-explaining your project every session
FETCHCurrent library docs + the open webA docs / fetch serverAnswers past the model's training cutoff
SHIPIssues, PRs and CI from inside the chatThe GitHub MCP serverFewer context switches out to the browser
QUERYSchema-aware reads from your databaseA Postgres serverThe agent stops guessing your columns
RUNA sharper version of the shell it already usesrg · fd · jq · gh · fzfCheap wins it already knows how to call

These are categories, not endorsements: pick the server you trust in each row. Most speak the same open protocol, so adding one is a couple of lines of config pointing at a local binary. The exception is the top row: whole-machine search installs itself for every agent at once: 19 AI clients auto-configured in a single install, Claude Code and Cursor included. That is the one worth walking through, because it is the one your agent quietly spends most of its time doing.

The one most setups skip

Start with search.

On a real machine the agent spends most of a session not thinking but finding: which file, which function, where a string lives. We watched Claude Code find one file on a 4.47-million-file drive, then handed it a pre-indexed lookup instead.

6m 57s → 16ms
time to the file: Claude Code hunting, then one indexed call
71 → 1
tool calls spent finding it
~58% → ~0%
of session tokens burned on file hunting

That is roughly 7,200,000× less time for the same answer; a single lookup alone went from 193.5 seconds of directory archaeology to microseconds (2,276,000×). Tokens spent hunting are tokens not spent solving, and it is the one gap you can close without changing anything else about your setup.

Same machine · same queries

Your agent already searches. It just does it the slow way.

Ask Claude Code or Cursor to find something and they shell out to the same tools you would: ripgrep, find, walk the directory tree, read the bytes, hope. Inside one open project that is fine. Across every repo plus the 90% of the disk that was never in git (configs, documents, downloads, the PDF with the answer) it crawls. A live, always-warm index answers the same questions without walking anything.

93.8 s → 16 ms
the same content query: what the agent runs today, then one indexed call
412,000×
faster than VS Code Ctrl+P on the same file (85 µs vs 35 s); ~480,000× vs Windows Search
~1 ms
from saving a file to finding its new contents, under 30 ms in the worst case

Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4.47M files. The content row is the same query run through ripgrep and through the index; the file-lookup multiplier is a single indexed lookup vs the editor's own search. At rest the whole engine sits around 44 MB (less than a browser tab) and the index on disk stays under 1% of the drive (about 24 GB on a 4 TB machine).

ONE INSTALL · 19 AI CLIENTS AUTO-CONFIGURED · CLAUDE CODE AND CURSOR INCLUDED · NO JSON EDITING.

Do this today

Set it up in an afternoon.

1

Install search first. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Cline and the rest. Filename search works immediately; content search fills in behind it.

2

Add the service MCPs you use. GitHub for issues and PRs, a memory server for recall, a docs/fetch server for anything past the model's cutoff, a Postgres server if it touches your data, each a one-line add.

3

Sharpen the shell. Make sure ripgrep, fd, jq, gh and fzf are on PATH. The agent reaches for them constantly, and they are free wins it already knows how to call.

4

Scope when you know where you are. Whole-machine when you don't know where something lives; scoped to a folder when you do. Scoped queries return in microseconds.

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