admin-valentin/agent-ecosystem
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
agent-ecosystem/docs/research/gastown-paperclip-comparison-2026-05-16.md
777genius bfebdff3cf chore: save local dev updates
2026-05-16 20:03:30 +03:00

7.9 KiB
Raw Permalink Blame History

Gastown и Paperclip comparison для лендинга и README

Дата проверки: 2026-05-16 Цель: публичная таблица Agent Teams | Gastown | Paperclip | Cursor | Claude Code CLI без угадываний по конкурентам. Метод: gh repo view, gh api по первичным GitHub-файлам, официальные docs Cursor и Claude Code, страница Claude pricing.

Snapshot

Проект Позиционирование Статус на 2026-05-16 Лицензия
Gastown multi-agent workspace manager для coding agents 15,228★, latest v1.1.0 от 2026-05-07, push 2026-05-15 MIT
Paperclip control plane для autonomous AI companies 65,796★, latest v2026.513.0 от 2026-05-13, push 2026-05-16 MIT

Что изменилось после проверки 2026-05-05

  • Gastown: свежий GitHub snapshot изменился с v1.0.1 на v1.1.0; README/provider/scheduler факты для публичной таблицы остались валидными.
  • Paperclip: свежий GitHub snapshot изменился с v2026.428.0 на v2026.513.0; README/adapters/budget/runtime/Kanban facts остались валидными.
  • Claude Code costs: официальный cost guide теперь называет /usage как команду для session token/cost tracking. Поэтому публичная строка Budget controls для Claude Code CLI обновлена с /cost + workspace limits на /usage + workspace limits.
  • Claude pricing: Team pricing page явно включает Claude Code в Team seats; публичная строка Claude plan or API usage остаётся корректной.
  • Cursor: official docs по Background Agents, Diffs & Review, Bugbot и usage/pricing по-прежнему поддерживают текущие формулировки таблицы. Background Agents остаются remote/async agents on separate branches/VMs with auto-run terminal commands; Bugbot остаётся PR-review product with its own pricing.

Проверенные публичные формулировки

Gastown

  • README по-прежнему позиционирует Gas Town как workspace manager для Claude Code, GitHub Copilot, Codex, Gemini и других coding agents.
  • Provider guide по-прежнему описывает tmux/provider contract для Claude, Gemini, Codex, Cursor, AMP, OpenCode, Copilot и других.
  • Scheduler docs по-прежнему подтверждают scheduler.max_polecats, deferred dispatch, capacity governor, pause/resume и daemon dispatch cycle.
  • Dashboard остаётся monitoring view for agents, convoys, hooks, queues, issues and escalations, а не Kanban product.
  • Refinery merge queue есть, но это не hunk-level diff review UI.

Публичная оценка не меняется:

  • Task dependencies - ✅ Beads DAG waves
  • Kanban board - ❌ Dashboard, not Kanban
  • Per-task code review - ⚠️ Merge queue, no diff UI
  • Budget controls - ⚠️ Cost tiers + digest, no hard caps

Paperclip

  • README по-прежнему описывает org charts, budgets, governance, goal alignment and agent coordination.
  • Adapter overview подтверждает Claude Local, Codex Local, Gemini experimental, OpenCode Local, Cursor, OpenClaw Gateway, Process and HTTP adapters.
  • Budget docs подтверждают per-agent monthly budgets, warning threshold at 80%, hard stop at 100%, auto-pause and no more heartbeats.
  • Runtime services docs подтверждают manual UI-managed services/jobs and execution workspaces with isolated checkout/branch/runtime state.
  • Kanban source по-прежнему содержит backlog, todo, in_progress, in_review, blocked, done, cancelled and @dnd-kit.
  • Work product validators подтверждают preview_url, runtime_service, pull_request, branch, commit, artifact, document and review statuses.

Публичная оценка не меняется:

  • Kanban board - ✅ 7 columns, drag-and-drop
  • Per-task code review - ⚠️ PR/work products, no inline diff
  • Hunk-level review - ❌ Bring your own review
  • Budget controls - ✅ Per-agent budgets + hard stops

Cursor

  • Background Agents are asynchronous remote agents that edit/run code in isolated machines, use GitHub branches, support follow-ups and can auto-run terminal commands.
  • Diffs & Review still supports diff review with accept/reject flows and selective acceptance.
  • Bugbot still focuses on PR review and comments with explanations/fix suggestions; pricing remains separate from normal Cursor subscriptions.
  • Usage/pricing docs still describe Free + paid usage, included agent usage by plan, dashboard token/usage breakdowns, background-agent access and spend limits.

Публичная оценка не меняется:

  • Full autonomy - ⚠️ Background agents, not teams
  • Hunk-level review - ✅
  • Review workflow - ⚠️ PR/BugBot only
  • Flexible autonomy - ⚠️ BG agents auto-run commands
  • Price - Free + paid usage

Claude Code CLI

  • Agent teams are still experimental and disabled by default through CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS.
  • Official docs still confirm shared task list, mailbox, direct teammate messaging, task dependencies, plan approval requests, quality-gate hooks and local team/task storage.
  • Agent teams require Claude Code v2.1.32 or later according to the docs.
  • Worktrees remain an official workflow for isolated sessions, but this is not the same as a product-level worktree strategy UI.
  • Cost docs now use /usage for detailed token usage statistics, plus workspace spend limits and usage reporting in Console for API users.

Публичная оценка после свежей проверки:

  • Agent-to-agent messaging - ✅ Team mailbox, no UI
  • Linked tasks - ✅ Shared task list
  • Task dependencies - ✅ Team task deps, no UI
  • Budget controls - ⚠️ /usage + workspace limits
  • Multi-agent backend - ⚠️ Claude-only experimental teams

Источники