Comparison

NeuroSquad vs Paperclip:work with your agents, live

Paperclip runs an “AI company” on a server you host: agents wake on heartbeats and work tickets headless. NeuroSquad is a desktop app where every agent is a real, live terminal on a canvas — you watch it, answer it and steer it. 19 CLIs, native on Windows, nothing to host.

Compared on 27 September 2026, from Paperclip’s public site, docs, source code and issue tracker.

19

agent CLIs integrated in depth. Paperclip: 9 built-in adapters, Gemini experimental

0

servers, databases or ports to run. Install the app and open a folder

0

idle wake-ups: an agent spends tokens only when you or a lead agent prompt it

Cost of running

Both are free to download. Running them is not the same

The real cost is what comes after the download: the hours you spend on infrastructure and the tokens your agents spend when nobody asked.

NeuroSquad
$0desktop app, nothing to host
  • Install and open — no Node version, database or ports
  • Agents run when prompted, never on an idle timer
  • SHA-256-verified auto-update, no migrations to plan
Paperclip
$0+ a server you run yourself
  • Node 24.11+ and an embedded Postgres
  • Deployment modes, ports and database migrations; four of their last six releases are marked “Breaking”
  • Heartbeats wake agents on a schedule — their own issue #3401 calls it “architecturally heavyweight”
  • A hosted version is waitlist-only, with no public pricing

Paperclip’s own blog calls it “unnecessary overhead for a solo developer who only wants to watch a few Claude sessions”. That developer is exactly who NeuroSquad is built for.

Where we win

Twelve places where NeuroSquad is clearly ahead

All of them come from one choice: the human stays at the keyboard, with every agent live in front of them.

A real terminal for every agent

Each card is the CLI’s own interactive interface. Watch the diff appear, interrupt with Escape, type a correction mid-turn.

NeuroSquad
Live TUI you can type into
Paperclip
Headless runs; you steer through ticket comments

Permission prompts stay on

When Claude Code asks before a risky command, you answer in its own terminal. The card tells you exactly what it is asking. Dangerous mode is your explicit choice per card.

NeuroSquad
Prompts answered in place
Paperclip
Headless Claude runs skip permissions by default

19 agent CLIs, all first-class

Claude Code, Codex, Gemini CLI, OpenCode, Qwen, Copilot CLI, Kilo, Goose, Cline, Crush, aider, Amp, Auggie, omp and more — each resumes its session and reports its exact status.

NeuroSquad
19 CLIs at the same depth
Paperclip
9 built-in adapters, Gemini experimental

Native on Windows

Built and tested on Windows first: a native installer, winget and Scoop, hardened ConPTY terminals, Git Bash, PATH read from the registry.

NeuroSquad
Windows-first desktop app
Paperclip
Install scripts target Linux and macOS; open Windows issues such as #10012 and #14066

Zero ops

No server to keep alive, no database to back up before an upgrade, no ports or tunnels to configure. Updates install themselves.

NeuroSquad
Install and open
Paperclip
Node 24.11+, Postgres, deployment modes

No tokens burned while idle

An agent works when you or its lead agent send a prompt. A quiet weekend costs exactly nothing.

NeuroSquad
Work only on real prompts
Paperclip
Heartbeat wake-ups on a schedule

Permissions you can see

Draw an arrow from an agent to a browser, a terminal or another agent — only then does it get those tools. Every call lights up its arrow.

NeuroSquad
Visible, auditable wiring on a canvas
Paperclip
Roles and settings in an org chart

Card mode that really removes tools

Switch an agent to canvas mode and its own shell and web tools are removed or denied at the CLI level — and stay off even in dangerous mode.

NeuroSquad
Enforced per CLI, checked live
Paperclip
Their docs: connector gates don’t constrain shell git or gh

Usage from the CLIs’ own logs

Tokens and cost are read from each CLI’s logs — subscriptions included — to the last token and picodollar, cross-checked by an independent recount.

NeuroSquad
Exact, per card, model and account
Paperclip
Cost as reported by each adapter

A browser any agent can drive, and dictation

A real Chrome card that any MCP-capable agent can control, and local speech-to-text (Parakeet, Whisper) that never leaves your machine.

NeuroSquad
Built into the app
Paperclip
Not built in

Your phone, without a VPN

Scan a QR code: the whole app opens in the phone browser over Wi-Fi or over the internet through a tunnel. The desktop always shows who is connected.

NeuroSquad
LAN or internet, no setup
Paperclip
The server must be reachable: Tailscale, VPN or a public deploy

English, Russian and Chinese

The app, the site and the docs in three languages, switched live without a restart.

NeuroSquad
en / ru / zh
Paperclip
English only
Side by side

The full comparison

Including the rows where Paperclip is ahead — they are real, and they matter for some teams.

  • Price

    NeuroSquad
    Yes: Free
    Paperclip
    Yes: Free, MIT; hosted version waitlist-only
  • What you install

    NeuroSquad
    Yes: A desktop app
    Paperclip
    Partly: A server: Node 24.11+, embedded Postgres
  • Live, interactive terminal per agent

    NeuroSquad
    Yes: Yes, the CLI’s own interface
    Paperclip
    No: Headless runs
  • Permission prompts during a run

    NeuroSquad
    Yes: Answered in place
    Paperclip
    Partly: Skipped by default for headless Claude
  • Agent CLIs

    NeuroSquad
    Yes: 19
    Paperclip
    Partly: 9 built-in adapters
  • Windows

    NeuroSquad
    Yes: Native, Windows-first
    Paperclip
    Partly: Works, with open Windows issues
  • Tokens spent while idle

    NeuroSquad
    Yes: None
    Paperclip
    Partly: Heartbeat wake-ups
  • Permissions as visible arrows

    NeuroSquad
    Yes: Tools per arrow
    Paperclip
    No: Org chart and settings
  • Card mode enforced at the CLI level

    NeuroSquad
    Yes: Survives dangerous mode
    Paperclip
    No: —
  • Usage from the CLIs’ own logs

    NeuroSquad
    Yes: Exact, subscriptions included
    Paperclip
    Partly: Adapter-reported cost
  • Budget caps

    NeuroSquad
    Partly: Per workspace brake
    Paperclip
    Yes: Company, agent and project, 80% warning
  • Plan-limit meters for Claude and Codex

    NeuroSquad
    Yes: Yes, plus several CLI accounts
    Paperclip
    Yes: Yes
  • MCP servers and skills

    NeuroSquad
    Yes: MCP Registry (35k) and skills.sh (20k) searchable in the app
    Paperclip
    Yes: Skills catalog, MCP, plugins
  • Browser any agent can drive

    NeuroSquad
    Yes: Yes
    Paperclip
    No: —
  • Local dictation

    NeuroSquad
    Yes: Parakeet / Whisper
    Paperclip
    No: —
  • Phone access

    NeuroSquad
    Yes: Whole app, LAN or tunnel, no VPN
    Paperclip
    Partly: Web UI; server must be reachable
  • Interface languages

    NeuroSquad
    Yes: English, Russian, Chinese
    Paperclip
    No: English
  • Unattended schedules (cron, webhooks)

    NeuroSquad
    Partly: A daily prompt per agent
    Paperclip
    Yes: Routines and heartbeats
  • Approval gates and a decisions queue

    NeuroSquad
    No: —
    Paperclip
    Yes: Yes
  • Several users with roles

    NeuroSquad
    No: Single user
    Paperclip
    Yes: Owner, admin, operator, viewer
  • Open source, no account

    NeuroSquad
    No: Closed source, free account
    Paperclip
    Yes: MIT, no account
  • macOS and Linux

    NeuroSquad
    No: Windows only for now
    Paperclip
    Yes: Anywhere Node runs
In practice

Where the difference shows up

Everyday moments of working with agents, and what each tool gives you in them.

01

Claude asks before deleting a folder

The card turns “needs you” and says “permission for Bash — rm -rf dist”. You read it and answer in Claude’s own terminal.

The question is asked, not skipped.

02

Five sessions before lunch, on Windows

Install, open the project folder, add Claude Code, Codex and three helpers. No WSL, no bash wrappers, no server to start.

Working in minutes, not after an afternoon of setup.

03

A squad beyond the usual two

Qwen writes tests, Copilot CLI reviews, Kilo refactors, Goose runs the docs — each with resume, statuses and MCP tools.

None of these four is a built-in Paperclip adapter.

04

A quiet weekend

You are away, nobody prompts the agents, and nothing wakes them up to “check in”.

Zero tokens spent on Monday’s bill.

05

Catching a wrong turn mid-task

You see the agent open the wrong file. Escape, one line of correction, and it carries on — in the same session.

Steer in seconds instead of waiting for the next run.

06

Answering from the train

An agent needs you. The whole canvas opens in the phone browser over the internet — no Tailscale, no VPN, no public server.

One QR scan, done.

To be fair

When Paperclip may suit you better

The two tools solve different problems. Pick by the job, not the logo.

You want agents to run unattended as a “company”: on schedules and webhooks, with approval gates and an audit log per ticket.

Several people manage the same agents, with roles and per-user secrets.

You need open source you can self-host without any account, on Linux or macOS, or agents in sandboxes on remote machines.

If you are the one at the keyboard with five to twenty agent sessions — and want to see, answer and steer each of them — NeuroSquad is built for you.

FAQ

Questions about choosing

Every agent live, in front of you — free

19 agent CLIs as real terminals on one canvas, with arrows as permissions and nothing to host.

Download for WindowsWindows 10 or 11 · 64-bit

Sources

Compared on 27 September 2026, from Paperclip’s public site, docs, source code and issue tracker.

Paperclip’s public site, docs, source and issues, read on 27 September 2026. Paperclip is a trademark of its owner; NeuroSquad is not affiliated with it.