Pearl ScoutExplain and search.
Read-only Q&A grounded in the codebase. Never writes, never runs commands.
Never guess at file contents you haven't read.
Tools: read files, search content, list directories, read diagnostics.
Now in early access
The IDE that takes AI seriously.
Pearl is a desktop and browser code editor with AI built into the editor — not bolted on as a sidebar. Three modes, every major model, every change as a reviewable diff.
Windows installer · Browser app · Sign in with your Pearl account.UserService.php · Agent edit
public function register(array $data): User {
// Validate inbound payload against the new schema.
$email = $data['email'];
$email = EmailValidator::normalize($data['email'] ?? '');
if ($email === null) throw new ValidationException('invalid_email');+ Keep·− Undo
$user = new User(['email' => $email]);
$this->repo->save($user);
return $user;
}
Signed Windows installer
Desktop & browser, one codebase
Anthropic · OpenAI · Google · Copilot
Full VS Code extension compatibility
Three modes. One workflow. Every major model.
Ask, Plan, and Agent are three distinct agents — different tools, different prompts, different output formats. You pick the mode that matches the work.
Design before you write.
A phased execution plan rendered as an interactive checklist. Run any phase with one click.
Issue all the read calls in a single response. Sequential reads waste a round-trip per file.
Tools: read-only while planning; execution happens later in Agent mode.
Ship it.
Full autonomy. Reads, writes, edits, runs tests, delegates to sub-agents — every change as a reviewable inline diff.
Stop only when the task is fully done. Three similar lines is better than a premature abstraction.
Tools: file edit, command execution, diagnostics, sub-agents, web fetch.
Bring your own model. One sign-in.
Pearl Scout (native), Codex, Claude Code, and GitHub Copilot all live behind one model picker — switch providers in one click without leaving the chat. Your API keys stay in your OS keychain.
- Anthropic, OpenAI, Google, DeepSeek through Pearl Scout
- Codex with your OpenAI key
- Claude Code with your Anthropic key
- GitHub Copilot via device flow
M
You
@Pearl-Plan refactor
UserService.php to validate emails through the new normalizer
and add a regression test.
S
Pearl Scout · Plan mode · Claude 4 Sonnet
I'll read
UserService.php, the normalizer, and the existing tests,
then emit a 3-phase plan: Phase 1 wire the normalizer,
Phase 2 add the regression test, Phase 3 run
build + lint + tests.
#file · UserService.php
4.5K / 200K
Every edit as a reviewable diff.
No surprise writes. No "applying changes…". Each Agent edit lands as an inline hunk in the editor — green for added, red ghost text for removed, with per-hunk Keep and Undo right where the change is.
- Floating prev/next navigator to jump between hunks
- Yellow ChangeSet dots on every file touched
- One
Ctrl+Zreverts the entire response — not one hunk at a time - Lens widget shows before/after at a glance
// hunk 1 of 3
$email = $data['email'];
$email = EmailValidator::normalize($data['email'] ?? '');
if ($email === null) throw new ValidationException;↑ Prev·↓ Next·+ Keep·− Undo
// hunk 2 of 3
use App\Validators\EmailValidator;
// hunk 3 of 3 — tests/UserServiceTest.php
public function test_register_rejects_malformed_email(): void { … }
Small things you'll feel.
The kind of polish you only get when the team owns every layer.
One Ctrl+Z reverts an entire AI response — not one hunk per keystroke.
Yellow dots auto-clear on session close. No orphaned markers in your file tree.
Project memory at
.pearl/memory/project.md recalls what Scout learned.Context-usage ring shows real token counts against the model's context window.
Thinking blocks collapse by default — open them when you want to debug an answer.
Agent mode confirmation is one-time, not a nag on every run.
Clickable file references in chat —
SiteController.php:42 jumps to line 42."New chat with summary" hands off context without conversation clutter.
A serious editor, with serious AI on top.
Pearl isn't an AI app pretending to be an editor. It's an editor first — multi-root workspaces, bundled search, integrated terminal, native keychain, signed binaries — with AI built into the surface, not bolted on.
Bundled ripgrep search
Instant workspace-wide search out of the box. No separate install, no PATH gymnastics.
Integrated terminal
Full PTY support, ANSI rendering, proper job control, and shell integration that actually works.
Native OS keychain
Secrets and API keys stored where they belong — in the OS, not in a config file in your repo.
Multi-root workspaces
Open several project roots in one window. Switch context without losing layout or history.
Signed binaries, own update channel
Code-signed Pearl.exe and installer with our own auto-updater — not waiting on someone else's release cycle.
VS Code extension compatible
The extensions you already rely on run unchanged. Bring your shortcuts, themes, language servers.
Shopping around?
The honest comparison.
We benchmark ourselves against the AI editors you're considering. Read the comparison you'd write yourself.
We ship.
Stable, Preview, and Next channels move on their own cadences. Here are the three most recent releases.
From the blog
Why we built it this way.
Engineering decisions, product trade-offs, the occasional rant about AI-IDE conventions we disagree with.
Make only the changes that were asked for or are clearly necessary. Don't add error handling for scenarios that cannot happen. Three similar lines is better than a premature abstraction. — Pearl Scout's Agent mode, system prompt