Interlinked.

One install.
Fifteen agents.

Interlinked Files exposes a single MCP server on your machine. Every agent that speaks MCP gets microsecond file search without writing a single integration.

VS Code ext.~85 hrsNot neededCursor plugin~70 hrsNot neededCLI wrapper~60 hrsNot neededMCP serverN/A~15 minOLD WAY (PER HOST)MCP (ONE SERVER)

TL;DR: Key takeaways

Five bullets, then the details.

01

One MCP server install serves every agent that speaks the protocol (Claude Code, Cursor, Codex, and roughly nineteen clients today) with no per-host plugin to build or maintain.

02

A single filename lookup returns in 85 µs (139 µs median) across a 4.47M-file machine: about 412,000× faster than VS Code and ~480,000× faster than Windows Search.

03

Whole-machine content search runs in 7 to 9 ms; a query that takes ripgrep 93.8 s comes back in 16 ms: 58,625× faster.

04

For agents the win compounds: a task that spent 6m57s across 71 tool calls collapses to one 16 ms call, roughly 7,200,000× less wall time, with search dropping from ~58% of tokens to ~0%.

05

Everything is free forever, no card. Hosting starts at $5.99/mo.

Integrations needed

1

one MCP server, all agents served

Old way

15

separate plugins, each maintained forever

The ecosystem, April 2026

Every agent that matters speaks MCP.

AgentMCPCategoryNotes
Claude CodeYesCLIAnthropic's official CLI agent
CursorYesIDEAI-native IDE, largest agent userbase
Codex CLIYesCLIOpenAI's terminal agent
ClineYesExtensionVS Code agentic extension
ContinueYesExtensionVS Code + JetBrains
WindsurfYesIDECodeium's agentic IDE
Roo CodeYesExtensionExtended Cline fork
AiderYesCLIPython CLI pair programmer
Claude DesktopYesDesktopAnthropic's native app
GooseYesCLIBlock's open-source agent
OpenCodeYesCLIGo-based coding agent
ZedYesIDEMultiplayer editor with agent
Neovim + MCPYesPluginAny Neovim MCP plugin
Emacs + mcp.elYesPluginCommunity client
Any future clientYesProtocolMCP is an open standard

Every row is a shipping product. Every one registers an MCP server the same way: a few lines of config pointing at a local binary. One binary. Fifteen agents. Zero per-host maintenance.

Why this changes everything

The host is finally abstracted.

Before MCP, adding a tool to every coding agent meant writing nineteen different plugins. VS Code had its extension API. Cursor had its own system. CLI agents expected a different binary. Each host had its own lifecycle, its own review process, its own update cadence. Supporting three was a part-time job. Supporting ten was a company.

MCP abstracts the host. You write one server. Speak one protocol over stdio. Every MCP client calls your tools. The model decides when to search. You just answer.

Before MCP

1 plugin

per host, per language, maintained forever

With MCP

1 server

one binary, all hosts, zero maintenance per agent

The multiplier effect

Infrastructure, not another IDE.

Cursor sells a better IDE and asks you to leave yours. Interlinked sells a faster local index that works with the IDE and CLI agents you already use. The sell is additive, not replacement. Nothing changes in your workflow for the speedup to start paying off.

Agents supported

~15

MCP protocol

Install required

1

single binary

Per-agent speedup

2.3M×

on file lookup

Monthly cost

$0

free forever

The speedup is infrastructural. It changes the floor of how fast any file lookup can possibly be. The slow agents get fast. The fast agents get faster. One local daemon does the work nineteen vendors would otherwise each have to replicate badly.

$0

Per month, forever. Fifteen agents. $0.00 per agent.

Everything local is free forever. Hosting starts at $5.99/mo.

One install powers every agent on your machine.

How it works

Four tools, zero config.

The MCP server exposes four tools. The AI model picks the right one automatically. You never interact with them directly.

ToolWhat it doesSpeed
iFileNamesFilename search across the entire machine~2µs
iFileTextContent search inside all code files~8ms
iFileSearchCombined filename + content search~9ms
iFileKnowledgeEverything known about one file, folder, or repo: its machine-wide connections, plus notesn/a

What the agent sees

The model calls the tool. The tool returns instantly. No context window wasted on directory traversal.

Without Interlinked, an agent finds a file by calling find, fd, or ripgrep through a shell tool. That means spawning a process, walking the filesystem tree, streaming results back, and burning context window tokens on hundreds of irrelevant paths.

With Interlinked, the model calls iFileNames and gets back a precise result list in microseconds. No process spawn. No directory traversal. No wasted tokens. The entire machine is already indexed and sitting in memory.

Agent file lookup

Without Interlinked

shell → ripgrep → 14 seconds

~200 lines of output piped to model

14s

With Interlinked

MCP tool call → 2 microseconds

precise results, no token waste

2µs

7,000,000× faster. Every agent. Every query. Every time.

Free vs paid

Free forever. Hosting from $5.99/mo.

Install Interlinked Files and everything local is free forever, no card. The desktop app is a complete file search, content search, and browse tool: it runs offline, and no file names, paths or contents are ever sent. That is the funnel.

The MCP bridge, the connector that lets external AI agents query the same index you use locally, is free too, on all your devices. The paid product is hosting: sharing versioned projects from $5.99/mo.

The product is the index, and it costs nothing. Hosting the projects you share is how we keep the lights on.

Why MCP wins

Three reasons this architecture matters.

#1

Zero per-agent maintenance

Ship one server. Every new MCP client that ships next month gets your tool automatically. No new code, no new plugin, no review.

#2

Zero network attack surface

MCP runs over stdio. No TCP port. No DNS rebinding. No accidental LAN exposure. Physically unreachable from other processes.

#3

Additive, not replacement

Keep your editor, your CLI, your workflow. Interlinked doesn't replace anything. It makes everything you already use faster.

The bigger picture

The scaffold matters more than the model.

SWE-bench Pro shows a 22+ point swing between different agent scaffolds using the exact same model. That is 3× the gap between the best open-weight model and Claude Opus. The retrieval infrastructure matters more than the weights.

Every agent on your machine is only as good as its ability to find files, search content, and understand project structure. Interlinked raises that floor for all of them simultaneously: one daemon, one index, one install.

The model ceiling is reachable. Open weights hit 80.2% SWE-bench. The next frontier is what is wrapped around the model, and that is where Interlinked Files lives.

Common questions

Questions, answered.

Which agents does one install actually cover?

Every client that speaks MCP: Claude Code, Cursor, Codex CLI, Cline, Windsurf, Zed, and roughly nineteen shipping today, plus any that ship tomorrow. You register one local server over stdio; there is no per-host plugin to write.

How fast is it, really?

A single filename lookup returns in 85 µs (139 µs median) across a 4.47M-file machine: about 412,000× faster than VS Code and ~480,000× faster than Windows Search’s 67 s. Whole-machine content search runs in 7 to 9 ms.

Why does this matter more for an agent than for me at a keyboard?

An agent that shells out to ripgrep can burn 6m57s over 71 tool calls on one search, spending ~58% of its context on file paths. The same answer through the MCP tool is a single 16 ms call with almost no token overhead: roughly 7,200,000× less wall time.

Does the index stay fresh, and what does it cost to run?

It updates within ~1 ms of a file changing and idles at ~44 MB of RAM, so it is always current without a rescan. Benchmarks were measured on a Ryzen 9 9950X3D indexing 4.47M files.

What does it cost?

Everything local is free forever, no card required. Hosting starts at $5.99/mo.

Client list compiled from published MCP support documentation, April 2026. Speed benchmarks from Interlinked Files internal testing on 4.47M file corpus.