Changelog

What shipped, train by train.

Nine release trains in three days, most of them built by agents and merged by trains. Newest first, straight from the repository.

Milestone

AgentTask's own team moves into AgentTask

  • The AGT team was imported from Linear and cut over. Its issues, keys, history and people now live in AgentTask.
  • Numbers carry on where Linear stopped: AGT-222 was the last issue in Linear, AGT-223 the first one made in AgentTask.
  • An import now brings only the initiatives an imported project sits in, and Linear people with no match are kept by name so you can map them afterwards.
  • Imported history reads like native changes, Linear links are relinked and markdown images render in the editor.
BrandWebsite

A new mark, and a website on rails

  • The approved AgentTask logo replaces the old mark in the sidebar, on the sign-in screens, in the favicons and in the PWA icons.
  • agenttask.io gets a centred hero, rails with sleepers down the whole page and a giant mark behind it.
  • Scrolling sends a train along the rails. It stays off on phones and with reduced motion.
Release train 921 PRs

Migration: import from Linear

  • A read-only Linear client. Nothing is ever written to Linear.
  • Imports teams, issues with their keys, sub-issues, relations, projects, milestones, initiatives, documents, views, comments, attachments and history with the original time and actor.
  • A trial mirror every 5 minutes while you try it out, with deletions reconciled.
  • Import, mapping, progress and cutover screens. Cutover goes per team with a 7-day rollback.
Insight

Claude quota, read from the session

  • Every Claude Code session reports its five-hour and weekly limits through a status line command, agenttask statusline.
  • The daemon passes on the rate limits and nothing else. Your own Claude settings stay untouched.
  • The usage page now shows quota for Claude logins next to Codex and Grok.
Release train 828 PRs

Insight: what agents cost and ship

  • Overview, development, usage, quality and person pages, plus numbers on issue, project and train pages.
  • Daemons collect token usage and quota windows per CLI. The core turns them into a cost per run at API-equivalent prices.
  • Commit attribution: every agent commit is credited to whoever started the run.
  • Save mode picks a cheaper profile, a retry escalates to a stronger one, and the new gate verb keeps gate output compact.
Release train 749 PRs

Queue and release trains

  • A scheduler places every run on the machine with the shortest wait, with leases, a run state machine and accounts that pause and resume on rate limits.
  • Release trains: start dialog, train detail with lanes and live terminal thumbnails, and a ship page.
  • One gate job per PR head, a merge policy that never merges into main, and conflict runs that merge the release branch in and push without force.
  • Predicted and observed file overlap with warnings, and private machines reserved to their owner and members.
Release train 64 PRs

Fixes

  • Relative times inside a sentence read as one phrase.
  • The Slack channel picker refreshes its cached channel list.
  • Test helpers and clock assertions fixed so the suite stays green.
Website

agenttask.io goes live

  • The first public front page, a static site served straight from the repository.
Fleet

Signed daemon releases and the terminal relay in production

  • relay.agenttask.io carries terminal sessions between browsers and machines.
  • The daemon is built for macOS and Linux on arm64 and amd64, and every build is signed with ed25519.
  • Pairing works end to end with one command: curl -fsSL https://app.agenttask.io/install | sh -s -- <code>.
Release train 519 PRs

Fleet: machines and terminals

  • Pair a machine with one command. It gets its own token and a launchd or systemd service.
  • The daemon runs tmux sessions in worktrees from repo recipes, and updates itself after a signature check without losing sessions.
  • Terminal view, a terminals grid with layouts and a key bar for phones.
  • Agent verbs over the worktree socket (status, ask, blocked, done with PR, MCP) and liveness and rate limit detection with notifications.
Release train 413 PRs

Core 3: API, integrations and settings

  • REST API v1 with OpenAPI docs, API tokens, app actors and a hosted MCP server.
  • GitHub App, Slack, Sentry and outgoing webhooks with signing, retries and a delivery log.
  • The CLI: login, mine, show, issue, comment, attach, label, relation, project and query.
  • Settings pages, an audit log and export.
Release train 311 PRs

Core 2: planning and inbox

  • Projects, initiatives, milestones and project updates.
  • A board with drag between states, and a timeline with drag and zoom.
  • Documents with rich text, mentions and margin comments.
  • Notification rules, an inbox, an email digest after 15 minutes unread, web push and PWA install.
Release train 216 PRs

Core 1: accept-test fixes

  • Fifteen findings from the first accept test, fixed by agents in one train.
  • The test suite passes on MySQL as well as SQLite.
  • Timestamps travel as UTC, and every id is uppercase.
Release train 119 PRs

Core 1: the foundation

  • Sign in with a magic link, passkeys, Google or GitHub. Organisations, teams, roles and invitations.
  • The sync engine: bootstrap, delta, mutate and live channels into IndexedDB, with an outbox and gap recovery.
  • Issues with relations, sub-issues, labels, comments, reactions and attachments, and a rich text editor with mentions.
  • Filters, saved views, bulk actions, keyboard shortcuts and ⌘K.

The next train
is already running.

Start buildingHow trains work