The machines can reason.
The answers have been found.
Against all odds, humanity won.
Hunger ended. War faded. The future finally opened its doors.
Proceed to the next stage of civilization?
And suddenly the most advanced technology on Earth is waiting for Steve from accounting to press Enter.
Ghost in the Loop fixes that.
Ghost automatically continues multi-step AI conversations across ChatGPT, Claude, Gemini, Perplexity, DeepSeek, Copilot, Grok, Manus, and more.
Because if the AI already knows the next step, somebody should probably let it take it.
Humanity's greatest minds spent centuries advancing science.
Ghost needs five steps.
- You give the AI a mission.
- Ghost attaches a loop protocol to your message.
- The AI works through the problem, step by step.
- Ghost automatically continues the conversation at each stage.
- You return later and pretend you were involved the whole time.
Three modes. One premise.
Loop — step-by-step auto-proceed until the task is complete or the AI admits defeat.
Think First — the AI creates a plan before acting. This is generally considered an improvement over most corporate strategy meetings.
Roadmap Autopilot — the AI generates a roadmap, then follows it, then completes it. For the first time in recorded history, a roadmap may actually be executed.
Modern AI can:
✓ Write software
✓ Analyze research
✓ Build business plans
✓ Review code
✓ Debate ideas across eight distinct personas
✓ Generate multi-stage roadmaps
✓ Export a full structured record of everything it did
Modern AI cannot:
✗ Press Enter
Tampermonkey (Chrome, Firefox, Edge, Safari)
- Install Tampermonkey.
- Click:
- Click "Install."
- Give an AI a task.
- Leave.
- Return to discover that progress occurred entirely without your supervision.
It's unsettling at first. You get used to it.
Firefox Extension (no Tampermonkey needed)
- Clone this repo.
- Go to
about:debugging→ This Firefox → Load Temporary Add-on →extension/manifest.json. - See steps 4–6 above.
The AI works.
Ghost listens.
Steve is no longer a critical dependency.
Before taking action, the AI creates a plan. The plan is then followed. This alone puts it ahead of approximately 80% of Q3 initiatives.
The AI generates a numbered roadmap. Ghost runs every step in sequence. The AI synthesizes the results.
Somewhere, a project manager is weeping and they don't know why.
Paste a list of tasks. Walk away. Ghost runs them one after another, in order, without forgetting what it was doing, without needing coffee, and without asking to push the deadline.
Like an intern. Except it shows up on time.
Need a Researcher? Ghost has one.
Need a Builder? Ghost has one.
Need a Devil's Advocate?
We regret to inform you that Ghost has several.
Choose from: Researcher · Builder · Red Team · Devil's Advocate · Tester · Customer Voice · Executive · Round Table. The Round Table convenes all of them simultaneously, which is exactly as chaotic as it sounds and twice as useful.
Pre-built pipelines for: Deep Research · R&D Lab · Shipyard · Debate · Pre-Mortem · Trollproof.
Trollproof attempts to find every possible objection to your idea before anyone else does. It finds them.
Sometimes websites change.
Sometimes buttons disappear.
Sometimes the frontend team chooses violence.
Ghost now fails closed when the page no longer matches a reviewed adapter. It can identify a composer heuristically, but only one unique, reviewed Send control has authority to click. If delivery cannot be independently confirmed, Ghost pauses in an uncertain state and never sends the prompt again behind your back.
Every transcript export reports one of three outcomes:
- Complete — a platform archive returned the expected supported turns.
- Partial — a DOM fallback, lazy-loaded history, attachment, branch, or filter means completeness cannot be proved.
- Failed — no usable messages were captured, so no misleading empty export is presented as success.
Markdown and gitl.transcript.v1 JSON include this validation result. Cancel
aborts the active archive request and promises no file.
Because eventually someone will ask "wait, how did we get this result?"
And for once, you'll have an answer.
One click. Ghost writes a compressed briefing — mission, current position, last output, open questions — formatted for immediate paste into any AI model.
The baton passes. The work continues. Nobody has to explain anything from scratch.
The separate Capsule v2 machine format is under Export → Advanced. It preserves short and repeated turns, but it is explicitly experimental and Ghost does not claim that it is resumable until an importer exists.
Browser crashes.
Tabs close.
Power goes out.
Civilizations collapse.
Ghost remembers where it left off.
A live readiness score in the panel header. Green means everything's working. Yellow means something's drifting. Red means the platform changed its selectors again, which happens more than anyone would like.
Desktop notification when the loop completes, pauses, or encounters an error. So you can actually leave the room instead of watching a progress bar like it owes you money.
For people who want to know what's actually happening under the hood:
| What | How |
|---|---|
| Boot safety | Transactional boot isolates optional subsystems, mounts a fail-loud banner for critical failures, and records stable local error codes. An independent canary distinguishes “Ghost failed” from “the userscript manager never injected it.” |
| Signal detection | Weighted scoring: custom sigils [[GITL::PROCEED]] / [[GITL::HALT]] (+4), legacy keywords (+3), fuzzy matches (+2). HALT always wins ties. |
| Private diagnostics | Network telemetry stores timing and byte counts only. Reports omit prompts, message text, full URLs, query strings, and conversation identifiers; users review locally before copying or downloading. |
| At-most-once send | A two-phase tab lease gates one preselected reviewed actuator: a unique button, or an adapter-approved single Enter fallback when no button exists. Ghost first verifies that the complete intended prompt is present. State advances only after independent delivery evidence; ambiguity pauses for human reconciliation and is never retried. |
| Anti-automation delay | Randomized 8–15s between sends (2s on the first round). |
| Truthful export | Platform API capture is checked against expected counts and unsupported parts. DOM capture is always labeled partial. Capsule hashes are integrity hints, never a reason to delete legitimate repeated turns. |
| Transactional import | Config and Workshop bundles require exact schemas, validate every field before mutation, and roll back if persistence fails. |
| Own-UI isolation | All DOM selectors exclude #gitl descendants. Ghost cannot accidentally type into its own panel. This needed to be a feature. |
| CI tested | Jest contract tests plus Playwright boot, mobile staging, and send-safety tests in Chromium and Firefox when those engines are available. Runs on configured main/agent/** pushes and pull requests. |
Layer 0: Transactional boot + two-phase tab lease
Layer 0.5: Metadata-only network telemetry
Layer 0.7: Selector diagnostics + health scoring
Layer 1: Reviewed platform adapters; read-only selector learning
Layer 2: State + persisted send transaction journal
Layer 3: Redacted diagnostics + bounded timeline
Layer 4: Signal engine (pure logic)
Layer 5: At-most-once loop engine + human reconciliation
Layer 6: Validated export + transactional import
Layer 7: Basic controls + progressively disclosed advanced UI
The userscript is the canonical source. npm run build deterministically
generates extension/content.js; CI rejects drift. There is no second dev/
copy of the product.
No runtime dependencies. One userscript source. Works anywhere Tampermonkey works.
- Pause and Stop are always text-labeled. Stop preserves the run; Reset is a separate Advanced action.
- A guessed Send candidate is diagnostic evidence only. Generic/custom sites are manual-send unless they have a reviewed adapter.
- If Ghost attempted a Send but cannot prove delivery, choose either I see it in chat or leave the prompt for manual Send. Ghost never re-clicks.
- On a failure, open Settings → Advanced → Diagnostics. Ghost automatically prepares a redacted local report with a stable error code and offers Review, Copy, and Download. Public bug reporting never includes report content automatically.
- If no Ghost UI or error banner appears at all, install the independent canary
from
diagnostics/to determine whether the userscript manager executed on that site.
First-class: ChatGPT · Claude · Perplexity · Gemini · DeepSeek · Copilot · Grok · Manus
Generic adapter: Mistral · Kimi · Qwen · Meta AI · Poe · HuggingChat · You.com · Pi · Z.ai · Genspark · MiniMax · LMArena · Duck.ai
Custom: any chat interface, via custom selectors in the settings panel.
Before touching anything, read:
- DEVLOG.md — what was tried, what failed, and why. If something seems like a good idea, there's a chance it's already in here with a postmortem.
- CHANGELOG.md — what shipped, what bugs were found, and what Replit's Playwright tests caught that the unit tests missed.
- docs/ARCHITECTURE.md — contracts, selector patterns, signal scoring, platform quirks, test harness design.
Every push must update at minimum DEVLOG.md and CHANGELOG.md. The project has been built across multiple AI sessions and needs its history documented so future sessions don't re-solve solved problems.
Release 8.3.0 was updated by MShneur. Main editor: Agent CG (ChatGPT).
The future showed up early and forgot the manual. Everything's brilliant, nothing's finished, and half of it forgets your name between Tuesday and Wednesday. These are the strips of duct tape — free, because the parts that hold your work together shouldn't cost a subscription.
They work on their own. Use one, or tape a few together:
- CTRL-FORGE — Your AI forgot everything. Again. This is the repo that didn't.
- CTRL-AI — Teaches your AI to say "I'm not sure" instead of confidently inventing a citation.
- R-Duck — Autopilot. You mumble an idea, it hands back a plan with a straight face.
- Agents of AI — A cast of specialists. No coffee, no PTO, no LinkedIn updates.
- Ghost in the Loop (this repo) — Moves work between AIs without dropping it down the stairs. Full chat export, handoffs.
Each keeps its own license — don't assume they match.
AGPL-3.0 — see LICENSE.
Humanity achieved godlike AI. The last bottleneck was a guy named Steve.
Steve has been removed from the critical path.