TicketPipe turns what your computer already saw into a precise IT ticket, or fixes simple problems on the spot. Built on screenpipe.
Stack: TypeScript · Bun · screenpipe API · LLM structured output (JSON) · local secret redaction
"VPN doesn't work, thanks" → IT replies "Which error? Which version? Can you send a screenshot?" → the user has to reproduce the bug… which may not happen again.
With screenpipe, the bug has already been recorded. One click and TicketPipe reads the last minutes of screen activity, keeps what matters, masks secrets locally, and writes the ticket an IT technician actually wants: exact error messages, steps, app version, OS, what the user already tried, and a timeline.
Track: Make it work for a team (and a bit of find what to automate: every vague ticket is a support loop we remove).
| What IT gets today | What IT gets with TicketPipe |
|---|---|
| Subject: VPN marche pas "le VPN ne marche plus depuis ce matin, je ne peux pas accéder à l'intranet. Merci" |
GlobalProtect: client certificate expired — cannot connect to eu-west-gw • Error (verbatim): Required client certificate is not found or has expired. / Client certificate expired on 2026-10-15• GlobalProtect 6.2.1-87 · macOS Sequoia 15.4.1 · MacBook Pro M3 Pro • Already tried: retry ×2, quit & relaunch the app; gateway reachable (ping OK, 0% loss) • Suggested: renew/re-issue the user's client certificate |
(Fictional example — try it yourself with bun run demo.)
- new ticket: describe the problem in a few words; TicketPipe writes the ticket from what screenpipe saw: exact errors, steps, timeline (local time), a severity chosen by the AI with its reason, up to 3 possible solutions in plain language, and at most 3 questions IT may still ask.
- quick fix: for everyday problems (printer, Wi-Fi, an app that froze…), TicketPipe suggests up to 5 simple steps the user can do alone. If it's not something a user should fix (expired certificate, locked account, admin rights…) it says so. If the steps don't work, one click turns it into a ticket that already lists what was tried.
flowchart LR
A[screenpipe<br/>records screen + UI events<br/>locally] -->|GET /search<br/>last N min| B[context.ts<br/>dedupe OCR · group by app/window<br/>detect error lines]
B --> C[redact.ts<br/>mask passwords, tokens, emails,<br/>cards, IBAN, phones]
C --> D[ticket.ts<br/>LLM → structured JSON ticket]
D --> E[Review UI<br/>user edits & approves]
E -->|POST| F[GitHub Issues<br/>or Jira / Linear / ServiceNow]
style C fill:#e6f4ea,stroke:#2b8a3e
style E fill:#e7ecff,stroke:#3b5bdb
- Collect — calls screenpipe's local REST API (
/search) for OCR text and UI events over the last N minutes. - Clean — OCR repeats the same text every frame: we dedupe lines per app/window, build a chronological timeline of "app sessions", flag lines that look like errors (EN + FR), and fit everything into a token budget (most recent first).
- Redact — secrets and personal data are masked on the machine, before any LLM call: passwords, API keys, JWT/Bearer tokens, private keys, emails, phone numbers, IBANs, card numbers (Luhn-checked).
- Generate — the LLM gets a strict prompt: only facts present in the context, error messages copied verbatim, unknowns listed as "missing info" instead of invented.
- Review — the user sees the ticket and exactly what was sent to the AI, edits it, and decides to send. Nothing is published without that click.
- Publish — creates a GitHub issue with labels (
it-ticket,severity:*). No GitHub set up? It is saved as a Markdown file inDocuments/IT tickets.
Download the file for your computer from the Releases page, then:
| System | File | First launch |
|---|---|---|
| Windows | ticketpipe-windows-x64.exe |
Double-click. If SmartScreen warns you: More info › Run anyway (the app isn't code-signed yet). |
| macOS (Apple silicon / Intel) | ticketpipe-mac-arm64.zip / ticketpipe-mac-x64.zip |
Unzip, move TicketPipe to Applications, then right-click › Open the first time (unsigned app). |
| Linux | ticketpipe-linux-x64 / -arm64 |
chmod +x ticketpipe-linux-x64 && ./ticketpipe-linux-x64. Run once with --install to add it to your applications menu. |
The app opens in its own window (it uses Chrome, Edge, Brave or Chromium if installed, otherwise your default browser). Close the window to quit.
On first launch, a 3-step setup asks for the following. It only happens once: everything is saved on the computer.
- Your screenpipe API key: TicketPipe finds it automatically when screenpipe is installed (it runs
screenpipe auth tokenfor you). Otherwise: screenpipe app › Settings › Privacy. - An AI: Google Gemini or Groq (free tiers), Claude, OpenAI, a local Ollama model, or a demo AI that needs no account.
- Where tickets go: a folder on your computer (
Documents/IT tickets), or GitHub issues (repository + fine-grained token with Issues: Read and write).
Each step has a Test button that says exactly what's wrong if something doesn't connect. Everything can be changed later in Settings. Settings are stored only on your computer (~/.config/ticketpipe, ~/Library/Application Support/TicketPipe or %APPDATA%\TicketPipe).
Just want to see it? Click Try it with a sample incident on the setup screen, or launch with --demo.
Requirements: Bun ≥ 1.2.3.
git clone <this-repo> && cd ticketpipe
bun install
bun run app # server + app window
bun run demo # same, with the sample incident
bun run dev # server only → http://127.0.0.1:4242
bun run build # executables for Linux, macOS and Windows → dist/Developers can also configure everything with a .env file (see .env.example); settings saved from the app take precedence.
bun run ticket # draft from the last 15 min
bun run ticket -m 5 -n "can't open the intranet" # 5 min + a note
bun run ticket --show-context # also print what is sent to the AI
bun run ticket --publish # draft + create the GitHub issue
bun run ticket --demo # on the sample incidentbun run snapshot 10 # save the last 10 min of screenpipe data
CONTEXT_FILE=fixtures/snapshot-<date>.json bun run dev| Variable | Default | |
|---|---|---|
SCREENPIPE_URL |
http://localhost:3030 |
screenpipe API |
SCREENPIPE_API_KEY |
– | screenpipe auth token |
CONTENT_TYPES |
ocr,accessibility,input |
add audio to include transcripts |
LOOKBACK_MINUTES |
15 |
default window |
MAX_CONTEXT_CHARS |
12000 |
context budget sent to the AI |
CONTEXT_FILE |
– | replay a saved response instead of calling screenpipe |
LLM_PROVIDER |
anthropic |
anthropic · openai (any OpenAI-compatible endpoint: Gemini, Groq, Mistral, Ollama…) · mock |
ANTHROPIC_API_KEY / ANTHROPIC_MODEL |
– / claude-sonnet-5-5 |
when LLM_PROVIDER=anthropic |
OPENAI_API_KEY / OPENAI_MODEL / OPENAI_BASE_URL |
– | when LLM_PROVIDER=openai |
TICKET_LANGUAGE |
en |
e.g. fr |
GITHUB_TOKEN / GITHUB_REPO |
– | fine-grained token with Issues: read & write · owner/repo. Empty = saved to Documents/IT tickets |
- The server listens on 127.0.0.1 only. Screen data never leaves the machine except the redacted context sent to the LLM, and the ticket the user explicitly approves.
- Redaction happens before the LLM call, and again on the generated ticket.
- The review screen shows the exact text sent to the AI — no black box.
- Pick Local model (Ollama) in Settings to keep everything on the computer, AI included.
- Settings and keys are stored in a file only the current user can read.
src/
app.ts desktop entry point: starts the server, opens the app window, quits when it closes
config.ts defaults ← .env ← settings saved by the app
context.ts screenpipe /search → deduped timeline + error highlights
redact.ts local masking of secrets & personal data
ticket.ts prompt, LLM providers (Anthropic / OpenAI-compatible / mock), JSON parsing, markdown
publish.ts GitHub Issues (+ dry-run). Add Jira / Linear / ServiceNow here
pipeline.ts fetch → clean → redact → LLM → markdown
server.ts local API (/api/settings, /api/test/*, /api/draft, /api/publish)
open.ts open links / folders with the system's default app
cli.ts terminal version
public/index.html the whole interface (setup, report, review, settings)
scripts/build.ts standalone executables for Linux, macOS, Windows
scripts/snapshot.ts save live screenpipe data as a fixture
fixtures/ demo data (fictional)
- Duplicate detection — before creating a ticket, search open issues for the same error ("12 users hit this certificate error today" → one incident instead of 12 tickets).
- Self-service fix first — match the error against a knowledge base and suggest the fix before escalating.
- IT dashboard — aggregate tickets by app/error to spot failing tools across the team.
- Connectors — Jira Service Management, ServiceNow, Linear, Zendesk.
- Native trigger — package as a screenpipe pipe + global hotkey / menu-bar button.
- Screenshot attachment — attach the redacted frame of the error (screenpipe frames API).
- Smarter windowing — auto-detect when the problem started instead of a fixed look-back.
Built by a team of 5 at the screenpipe hackathon, October 2026.


