# Relay: AI auto-ticketing for software teams

## What is Relay?

Relay is an open-source (Apache-2.0), self-hosted tool that turns meeting notes, chat, email and transcripts into reviewed tickets for software teams. It drafts each ticket with its evidence quote, waits for a person to approve it, then pushes it to GitHub, Linear, Jira or Notion. Relay automates routing, not decisions.

## How does Relay turn meeting notes into tickets?

Relay captures a note from chat, email, a document or a transcript, then uses an LLM to extract candidate tickets. Each draft links to the exact quote it came from. Drafts wait in a review queue until a teammate approves, edits or rejects them; only approved drafts are pushed to the configured tracker.

## Which trackers does Relay support?

Relay supports four trackers: GitHub, Linear, Jira (Cloud and Server) and Notion. It also ships first-class Slack and Discord apps for capture and review. A REST API and an MCP server connect other sources. Approved tickets are pushed through an idempotent outbox, so a retry never creates a duplicate issue.

## Is Relay self-hosted?

Yes. Relay is Apache-2.0 and self-hosting is the working path today, with every feature and no artificial limits. You run the API, worker, MCP server and database yourself using Docker Compose or the CLI. Hosted cloud plans are a design only: they are planned, not live.

## What does Relay do with the work it captures?

- **Capture:** text, Markdown, VTT/SRT transcripts, DOCX, PDF and transcript JSON; a deterministic demo mode needs no credentials.
- **Extract:** LLM ticket drafts, each linked to a source quote or segment.
- **Review:** approve, edit, reject, delegate, merge, split and bulk-approve; low-risk drafts may auto-approve per policy.
- **Push:** idempotent, exactly-once pushes via a database-backed outbox.
- **Assign:** up to three ranked owners with evidence, reasons and confidence scores.
- **Memory:** searchable decisions, commitments and tickets with provenance, plus digests.
- **Audit:** an append-only, hash-chained log attributing every write.

## How does Relay assign an owner to a ticket?

Smart Assignment ranks candidates using expertise, recent context, availability and fairness. It returns up to three suggestions, each with a confidence score and the signals behind it, so a reviewer can check the ranking before assigning. Every manual owner change is stored as feedback for the learning mode.

## How does Relay remember decisions and commitments?

Decisions, commitments and their sources stay linked to the ticket they came from. Standup and weekly digests are built from the same record, so context survives the sprint. Provenance travels with each item: a memory entry points back to the note or transcript it came from.

## How it works

1. **Capture.** A note, chat thread, email or transcript enters Relay through chat, the REST API or the MCP server.
2. **Extract.** Relay drafts tickets with type, priority, assignee, due date and an evidence quote.
3. **Review.** A person approves, edits, rejects or delegates each draft; nothing reaches a tracker without approval.
4. **Push.** Approved drafts enter an outbox and are pushed exactly once to GitHub, Linear, Jira or Notion.
5. **Remember.** The decision, commitment and source are stored in team memory and summarised in digests.

## What are the plans?

Hosted cloud plans are **planned, not live**. Self-hosting is the working path today.

| | Self-hosted | Cloud Free (planned) | Cloud Pro / Team (planned) |
|---|---|---|---|
| Cost | Free; you pay infrastructure | Free | Per-seat / usage |
| Who runs it | You | Relay | Relay |
| Source | Always available | Same code | Same code |
| Trackers | All four | One (GitHub or Linear) | All four |
| Chat apps | Slack + Discord | One | Both |
| Captures | Unlimited | 250 per month | Unlimited |
| Team size | Unlimited | Up to 3 | Unlimited |
| Smart Assignment | Full | Rule-based | Full + learning mode |
| Memory / audit retention | Your database | 30 days | 1 year, exportable |
| SSO / SCIM | Bring your own | — | Team |
| Status | Available now | Not live | Not live |

## Frequently asked questions

### What does Relay automate, and what does it leave to people?

Relay automates transcription, extraction and routing. It leaves the decision to a person: nothing reaches a tracker without approval. A reviewer sees the source and the evidence quote before the ticket exists, and can batch-approve low-risk drafts while reviewing higher-risk ones individually.

### How do I connect Jira to Relay?

An admin connects a tracker once with `relay connect jira --payload '{...}'`, supplying the Jira base URL, email, API token and project key. Credentials are encrypted per workspace with AES-256-GCM. Demo environment fallbacks are disabled when `DEMO_MODE=false`. Integration health is reported through `relay integrations`.

### Which roles can approve a draft?

The roles are `admin`, `reviewer`, `member` and `viewer`. Admins connect integrations and manage keys; viewers can read but not approve. Reviewers work a focused queue with out-of-office-aware delegation, snooze and escalation.

### Can an AI agent push tickets without confirmation?

No. Every AI tool goes through the MCP server or the REST API. MCP external writes require an explicit `confirm: true` parameter, so an agent cannot push a ticket without it. Agent keys are ordinary scoped keys, and a `viewer` key gives read-only access. Hosted OAuth apps for agents are not built.

### Is there a hosted version I can sign up for today?

No. Hosted signup and SSO, device login, OAuth apps, and billing, metering and entitlements are not built. The plan matrix exists as a tested quota layer in core, but it is not wired to live requests, no hosted limit is active, and no payment is accepted. Hosted plans are planned, not live.

### What limits should a team expect?

On messy multilingual fixtures, extraction measures about 0.78 precision and 0.75 recall, so expect to edit some drafts. Audio capture needs a transcription endpoint or a supplied transcript. Dedupe and digest keys are in-process; multi-instance deployments need shared state. Live chat and tracker smoke tests need real credentials.

## Where to read more

- [Relay on GitHub](https://github.com/vanshcodeworks/relay) — source, README and the self-hosting quick start.
- [Workflow guide](https://github.com/vanshcodeworks/relay/blob/main/docs/WORKFLOW.md) — the team loop and known limits.
- [Authentication guide](https://github.com/vanshcodeworks/relay/blob/main/docs/AUTH.md) — identities, keys and agent access.
- [Hosting guide](https://github.com/vanshcodeworks/relay/blob/main/docs/HOSTING.md) — plans and deployment topology.
