paperclipai

Activity

Stars and mentions · Last 30 days, Oct 6 – Nov 4

GitHub stars8,816
Mentions85
0131262393524Oct 6Oct 20Nov 4
UTC days · repositories paperclipai owns3 repositories

Top repositories

paperclipai’s repositories ranked by new stars, as on trending.

Loading

Contributes to

Repositories outside paperclipai’s account where paperclipai opened or reviewed pull requests in the past 365 days.

Who noticed paperclipai

Everyone who starred or forked paperclipai’s repositories since the archive began in 2011.

Notable accounts

Ordered by the stars their own repositories have received.

Biggest fans

Accounts that starred the most of these repositories.

Breakout weeks

Each repository’s best week of new stars, with the release published around it.

Mentions

84posts
Nov 52 posts
  1. Multica と Paperclip 権限の既定をソースから比較してみた Claude Code や Codex を並行して動かし始めると、まとめて管理する道具が欲しくなります。候補に挙がる Multica と Paperclip は、GitHub でそれぞれ約 5.5 万と約 11 万のスターを集めています(2026-11-04)。 この記事では、この 2 つに対照として Orchestrator.inc を加え、CLI が出す権限の確認を既定でどう扱うかをソースから比較します。最後に、動かす前に確認したい 4 点を挙げます。 執筆時点では導入実行は行っていません。ソース・公式文書・issue を元にした比較整理です。 ! 参照した版(いずれも 2026-11-04 時点) Multica: v0.8.0(コミット 9b1ab4f) Paperclip: db9be4c(権限の経路は cf51936) Orchestrator.inc: a1fe965e 挙動の話は「ソース上はそう読める」「文書にそう書かれている」の範囲です。ソースへのリンクは版(コミット)を固定した形式で、行が分かるものは行番号を付けています。版が進んでも同じ箇所を指します。 3 つの道具は何を管理するか Multica と Paperclip では、人が各エージェントの画面を見ている必要はありません。課題(作業の依頼と進み具合を記録する単位)を作ってエージェントに割り当て、人は結果を確認します。Orchestrator.inc では、統括役のエージェントが計画を立て、作業ごとにエージェントを起動します。 同様の仕組みは OpenAI の Symphony にもあります。OpenAI は、多くの人が無理なく見られるのは同時に 3〜5 セッションだと書き、開いた課題ごとにワークスペース(エージェントが作業するディレクトリ)とエージェントを割り当てる仕組みにしました(OpenAI の記事)。 管理する対象は、3 つで異なります。 Multica は作業の割り当てを管理します。課題を割り当て、経過をコメントで返し、結果について人のレビューを求めます。Claude Code・Codex など 26 の CLI に対応します。実行は手元の常駐プロセスで、Docker か Helm でセルフホストもできます(README) Paperclip は組織の運営まで扱います。組織図・目標・予算・承認・定期実行を持ち、コード以外の仕事も扱います。公式の説明は、公開時の「人間ゼロの会社」から、仕事でエージェントを管理するアプリに変わりました(README) Orchestrator.inc はコード開発向けです。作業ごとに worktree(1 つのリポジトリから作業ディレクトリを複数作る Git の仕組み)で動かし、CI の失敗やレビューの指摘を担当のエージェントに返します(README) なぜ権限の既定を比較するのか 課題を割り当てて任せる使い方では、CLI の画面の前に人がいません。そのため、CLI が出す権限の確認に誰が答えるか が、道具ごとに決められています。 Claude Code では、作業ディレクトリ内の読み取りは確認が要らず、Bash・書き込み・Web 取得などは確認が要ります(permissions)。bypassPermissions は、この確認をすべて許可するモードです。 非対話の claude -p では、確認に答える人がいません。許可の設定や hook など、承認する仕組みがなければ、確認の要る呼び出しは拒否されます(headless、permission-modes)。 Codex が確認を出すかどうかは、承認の方針(approval_policy)とサンドボックス(sandbox_mode。workspace-write や danger-full-access など)の組み合わせで決まります。両方を無効にする引数が --dangerously-bypass-approvals-and-sandbox です。 確認への答え方は 3 つに分かれます。 道具がすべて許可する(以下「全許可」。すべて許可するモードで CLI を起動する) 人に求める(確認を画面に出して待つ) 拒否して先へ進む(許可していない呼び出しは拒否され、エージェントは続ける) Anthropic の文書は、bypassPermissions を隔離した環境でだけ使うよう書いています。以下では、全許可を無人で動かすための設計として扱います。Multica と Paperclip の節では、最初に結論を書き、続けて、公式文書での位置づけと推奨される隔離を示します。 Multica は Claude Code を全許可で起動し、隔離の責任は利用者が担う Multica は Claude Code を全許可で起動し、利用者の引数では変えられません。実行中の権限の確認は人に表示されず、人は課題の結果をレビューします。 公式の security-model 文書は、ファイルの隔離を保証しないこと、個人アカウントで動かせば SSH 鍵も読めることを書き、専用ユーザー・コンテナ・VM で動かすよう勧めています(security-model.mdx#L14-L28)。この文書は、隔離を利用者の責任とする前提で書かれています。 根拠は v0.8.0 のソースです。 Claude Code を -p と --permission-mode bypassPermissions で起動します(claude.go#L1150、#L1167) Codex の実行と変更の承認要求は自動で受け入れます(codex.go)。Windows での既定のサンドボックスは danger-full-access です(codex_sandbox.go)。ほかの OS の既定は調べていません 人の承認を挟む機能の提案は未解決です(#9149) データとライセンスの既定は次のとおりです。 セルフホストした API サーバーは、日次の匿名テレメトリーを既定で送り、DO_NOT_TRACK=1 で止まります(SELF_HOSTING.md#L21) 課題・コメント・チャット・添付・エージェントの進捗報告は、Multica 側に保存されます。自分のサーバーに保存したい場合はセルフホストになります(公式 FAQ。サーバーは Docker 前提、SELF_HOSTING.md) ライセンスは Apache 2.0 に条件を加えた独自のもので、README は source-available と書きます。組織の中ではセルフホストを含めて無料、社外への提供と製品への組み込みは商用ライセンスです(公式 FAQ) Paperclip はどの経路でも、既定で確認をスキップする Paperclip は、どの実行経路でも既定は全許可です。無効にするための設定は経路によって異なり、画面のスイッチが効かない経路もあります。人の判断を求める場面は雇用・戦略・支出などの統治の承認で、費用は予算の上限で抑える設計です。 公式文書は、確認をスキップする設定を、確認に答えられない無人の実行に必要なものと説明します(claude-local.md#L38)。全許可でも、会社の権限・統治の承認・Runner のワークスペースの隔離は別途有効だと説明しています(Runner の README#L509-L511)。 3 つの経路と既定 Claude のエージェントの実行には 3 つの経路があり、有効な設定が異なります。Paperclip Runner は Paperclip 自身の実行基盤で、ワークスペースを隔離して動かします。ACP(Agent Client Protocol)は、エディタとエージェントをつなぐ標準です。 経路 選ばれる条件 有効な設定 既定 既定から変えたときの挙動 Paperclip Runner Linux・macOS の新しい Claude エージェント(自動) acpxPermissionMode approve-all 作業の画面のカードに確認が出て待つ。応答先がなければ approval_required で止まる(README) ACP claude_local の既定のエンジン permissionMode approve-all 答えられない確認は nonInteractivePermissions(既定 deny)に従い拒否か失敗 CLI engine=cli を明示したとき dangerouslySkipPermissions true 権限の引数を渡さず、確認の要る呼び出しは拒否されて先へ進む 経路の分岐は execute.ts#L424 にあり、ACP を選んだ場合はこの行で ACP の処理を呼び出し、CLI の処理は実行されません。 どれもエージェントごとの設定で、会社全体の設定は見つかりませんでした。Runner は Linux(glibc)と macOS だけで、Windows のホストは ACP か CLI になります(agent-runner.ts#L32-L38)。 確認をスキップする設定の既定を false にする PR #16867(2026-10-19)は、まだ統合されていません。逆に、既定を全許可へ揃える #13693 は 2026-09-19 に統合されました。 設定を既定から変えたときの挙動 確認をスキップする設定を既定から変えた場合を、経路ごとにソースで追うと次のとおりです。 ACP: 画面の「Skip permissions」(dangerouslySkipPermissions)をオフにしても、全許可のままです。ACP の経路はこの設定を実行に使わず、実行に使う permissionMode は Claude の設定スキーマにありません(config-schema.ts)。画面のスイッチを見ただけでは、実際の挙動は分かりません CLI: 無効にすると権限の引数を何も渡しません(permissions.ts#L12)。Claude Code の公式文書によれば、v2.1.302 以降では確認の要る呼び出しが拒否され、Claude は拒否を受けたまま処理を続けます。承認を待ち続けたという報告(#10679)もあります。どちらの場合も、確認の要る操作は実行されません Runner: acpxPermissionMode を approve-all 以外にすると、確認が作業の画面のカードに表示されます。確認を人に求める仕組みは、3 経路のうちこの経路にだけあります。ただし、Claude で実際に動作するかは、issue からも裏づけが取れていません 統治の承認のページ(雇用・戦略・支出)の説明には、どの経路についても権限の確認は含まれていません(approvals.md)。 CLI で無効にすると、確認の要る操作が実行されなくても、成功と記録されるおそれがあります(推測)。Paperclip は、終了コードか is_error が失敗を示すときだけ失敗と判定するためです(execute.ts#L1149-L1152)。スキップする引数が効かなかった古い版では、権限で何もできずに終わった実行が成功と記録された報告があります(#3148)。 Codex で対応する dangerouslyBypassApprovalsAndSandbox も、CLI の既定は true です(codex-local の index.ts#L15)。無効にすると workspace-write のサンドボックスで動き、ネットワークは許可したままです。このとき確認がどう扱われるかは調べていません。 ACP はこの設定を読まず、既定は approve-all です。Runner は承認の方針を never(確認を出さない)だけに限り、作業ディレクトリの限定とネットワークの遮断は Paperclip の側で行います。 定額プランと予算の上限 ソースでは、Paperclip は定額分の費用を 0 として計上します。このため 定額プランでは予算の上限が効かない と読めます(budget-exposure.ts)。予算の上限での停止を前提にするなら、先に自分の課金形態を確認してください。 セキュリティ勧告とライセンス 公開されているセキュリティ勧告は 12 件で、うち critical は 5 件です(Security Advisories、2026-11-04 時点)。Multica と Orchestrator.inc の勧告の件数は、この記事では調べていません。 Oasis Security は 2026-08-05 に 3 件の重大な脆弱性を公表しました。CVE-2026-41679 は CVSS 10.0 で、修正は v2026.416.0 です(Oasis Security)。 ライセンスは MIT で、匿名の利用統計を集めると README に書いています。 Orchestrator.inc は Claude Code に権限の引数を追加しない Orchestrator.inc は、自身の権限の方針が既定(default)のままなら Claude Code に権限の引数を追加しません(command.go)。確認が出ると hook(決まった出来事で外の処理を呼ぶ Claude Code の仕組み)で検知します。作業担当のセッションでは、hook は許可や拒否の応答を返しません(hooks.go)。 ただし、確認に答える人がいないレビュー担当のセッションでは、hook が許可か拒否を返します。作業担当の Claude Code の確認は残り、その確認に人がどこで答えるかは調べていません。 Codex は扱いが異なり、方針が acceptEdits か auto のとき以外は、既定の方針も含めて --dangerously-bypass-approvals-and-sandbox で起動します(同じ command.go)。 権限の既定以外に、次の 2 点があります。 使用状況の共有は既定で有効で、GitHub のユーザー名が付くことがあります(telemetry.md) 運営の集計では、v0.14.3 の 7 日間に spawn(エージェントの起動)を試した 589 件中 171 件(29%)が失敗を経験しました(#7024)。主な原因は独自プロバイダの確認で、通常の利用にどれだけ当たるかは分かりません 3 者は、隔離の責任を誰が担うかが異なる 項目 Multica Paperclip Orchestrator.inc Claude Code の既定 全許可(変えられない) 全許可(3 経路とも) 権限の引数を追加しない Codex の既定 承認を自動で受ける 全許可(Runner は never) 承認とサンドボックスを無効にする(acceptEdits・auto を除く) 確認を人に求める仕組み なし(提案は未解決) Claude は Runner で設定を変えれば作業の画面に出る(動作は未検証)。Codex は Runner でも never Claude の確認を hook で検知(答える場所は未調査) 隔離(公式の説明) 専用ユーザー・コンテナ・VM を勧める Runner はワークスペースを隔離する(ACP・CLI は未調査) この記事では未調査 人が確認する場面 課題の結果(レビュー) 統治の承認 Claude Code の確認 Codex の承認を、Multica と Paperclip は既定で人に求めません。Orchestrator.inc も、既定の方針では承認とサンドボックスを無効にします。Claude Code の確認は、Multica と Paperclip が既定でスキップします。確認を人に求めないなら、Claude Code や Codex が何を読み書きできるかは、実行環境の隔離で決まります。 Multica では隔離の責任を利用者が担い、Paperclip Runner では道具の側でワークスペースを隔離します。Paperclip の ACP と CLI の経路での隔離は調べていません。Orchestrator.inc については、Codex 側の隔離を誰が担うかも、道具自身の隔離の説明も調べていません。 この比較で分からないこと この比較は、道具の効果を測ったものではありません。効果について調べると、次のことが分かりました。 効果を測った独立の評価は、調べた範囲では 3 つとも見つかりませんでした CooperBench(arXiv 2601.13295)では、対話で協調すると成功率が単独より平均 30% 落ちました。ただし課題を介して分担する使い方は測っていません 道具は入れ替わります。Vibe Kanban の運営会社は 2026-04-10 に閉鎖し(告知)、Gas Town の作者 Steve Yegge は再利用できる仕組みをやめて自分専用の道具を作っています(本人の論考) まとめと、動かす前に確認したい 4 点 管理する対象: Multica は作業の割り当て、Paperclip は組織の運営、Orchestrator.inc はコード開発です 権限の既定: Multica は Claude Code を全許可に固定し、Paperclip はどの経路も全許可、Orchestrator.inc は既定の方針では Claude Code の確認を残します 隔離: 実行中の確認を人に求めない道具では、Claude Code や Codex の権限を、実行環境の隔離で制限することになります。その責任を道具と利用者のどちらが担うかは、道具ごとに異なります 既定のまま動かす前に、次の 4 点を確認してください。 権限の既定: Claude Code と Codex をどのモードで起動するか、変えられるか、どの設定が実際に効くか。CLI を起動するソースの箇所で確認できます 隔離の責任は誰が担うか: 公式が勧める隔離は何か、道具の側で隔離するか。security-model や Runner の README に記載があります 予算の上限が効く課金形態: 定額プランか従量課金か。予算の上限が効く条件は、費用を集計するソースで確認できます データの保存先とライセンス: セルフホストの可否、テレメトリーの既定、社外へ提供するときの条件。README と公式 FAQ に記載があります 実際に使う場合は、隔離した環境で動かしてください。既定は版で変わり、とくに Paperclip の PR #16867 が統合されると、この記事の既定の話は変わります。 ! Windows ネイティブで動かす場合(権限の既定とは別の、Windows に固有の未解決の issue。2026-11-04 時点) Paperclip: 定額プランの接続が 2026.1023.0 から使えない(#17968)。Paperclip Runner も対応外 Multica: 成功した実行を、ファイルロックで失敗と報告することがある(#7548) Orchestrator.inc: 未署名のため Smart App Control に止められる(#7232) 参照一覧 Multica(v0.8.0): claude.go#L1150、claude.go#L1167、codex.go、codex_sandbox.go、security-model.mdx#L14-L28、SELF_HOSTING.md Paperclip(cf51936): execute.ts#L424、execute.ts#L1149-L1152、permissions.ts#L12、constants.ts#L3-L4、config-schema.ts、config-fields.tsx#L245-L262、agent-runner-selection.ts#L79、agent-runner.ts#L32-L38、permission-policy.ts、codex-local の index.ts#L15、claude-local.md#L38、approvals.md、Runner の README#L509-L511、同#L521-L534、DEVELOPING.md#L2886-L2889 Paperclip(db9be4c): execute.ts#L445、budget-exposure.ts Orchestrator.inc(a1fe965e): command.go、hooks.go、telemetry.md

  2. AI agents are becoming teams. Who manages them? Paperclip coordinates agents with tasks, permissions and budgets. But GitHub issues reveal the hard part: security, sandboxing and reliable execution. Repo: paperclipai/paperclip #AIAgents #OpenSource

    2 likes · 135 views
  3. 这个真不错,Paperclip 居然可以把一群 AI agent 当成一家公司来管。 再也不用开二十个终端盯着每个 agent 了,任务、预算、权限都在一个看板里,这波真的可以冲。 paperclipai/paperclip

    3 likes · 120 views
  4. paperclip ai: How to Set Up the Free, Open-Source App for Managing AI Agents paperclip ai is trending again on GitHub this week, and a lot of people are searching for a "paperclip ai tutorial" or "paperclip ai setup". If you run a small business or a small dev team and you already use one or more AI coding agents, here's a plain, research-based walkthrough of what Paperclip is, how to install it, and what to watch before you let agents run on their own. What Paperclip is According to its GitHub README, Paperclip is "open-source orchestration for teams of AI agents". In practice it's a Node.js server with a React dashboard where you: - set a goal for the work, - "hire" agents into roles on an org chart ("any bot, any provider", in the README's words), - assign tasks, approve plans, set budgets, and watch progress from one place. It's MIT-licensed and self-hosted, and the README says no Paperclip account is needed. The repo had just over 110,000 stars when we checked the GitHub API today. It works with agents you may already use. The README lists OpenClaw, Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, Hermes and Kimi Code, among others, plus custom processes and HTTP endpoints. Quick setup The README's quickstart needs Node.js 24.11 or newer. Then one command: npx paperclipai@latest onboard --yes The CLI runs from npm's cache, and your instance settings and data stay on your machine. For a local run, the README's FAQ says a single Node.js process manages an embedded PostgreSQL database and local file storage, so there's no separate database to set up. For production you can point it at your own Postgres. If you'd rather build from source, the README's development steps are pnpm install then pnpm dev, which starts the UI and API at http://localhost:3100 (it needs pnpm 9.15+). The full installation guide covers pinned versions, updates, rollback and uninstalling. A simple first project for a small team Keep the first run small and boring: 1. One goal, one or two agents. For example, "Clean up our website's FAQ page and open a pull request." Don't start with a company-wide goal. 2. Turn on review gates. Paperclip has configurable approval stages, so a human checks a plan before work starts. 3. Set a budget. The README says you can set company, agent and project budgets, get threshold alerts and pause work at a limit. Paperclip itself is free, but the AI models your agents call are not, so set the cap before you hit "go". 4. Read the task threads. Plans, blockers, files and run history stay attached to each task, which makes it easier to see what an agent actually did. Things to check before going autonomous - Scope the secrets. The README mentions scoped secrets and permissions. Give each agent only the keys and repos it needs. - Start in a test repo. Let agents work on a copy before they touch anything customers see. - Keep a human on merges. "Approve and run" is the README's own workflow. Keep the approve part. Is it worth trying? If you're juggling several agent terminals and losing track of who's doing what, Paperclip is a free way to put that work on one board with budgets and approvals. If you only use one assistant now and then, it's probably more structure than you need today. This is a research-based overview from the project's own README and docs. We haven't tested Paperclip hands-on yet. More plain-English AI tool guides for small businesses and creators: ainews.useaifix.com. If you'd like help setting up AI automation for your business, see useaifix.com.

  5. Paperclip's Agent Orchestration Stack: How Multi-Agent Teams Track Work, Budgets, and Tool Permissions in Production Most multi-agent frameworks stop at message passing. Paperclip treats agent coordination as a company org chart problem: hierarchies, delegation rules, budget caps, and permission boundaries. The open-source stack (109K+ stars, Node.js server + React UI) positions itself as "the company" layer above individual agents. If OpenClaw is an employee, Paperclip is the HR system, finance department, and project management office rolled into one. This matters because production agent teams need the same operational infrastructure as human teams. You need to know which agent spent $47 on API calls yesterday, who delegated a task to whom, and why the content-generation agent can't access your database. Paperclip exposes primitives for these problems that most orchestration tools ignore. The Company Abstraction Paperclip's core metaphor is organizational structure. You define agents as team members, assign them skills, set budget limits, and track their work history. The system maintains an org chart that determines delegation paths and escalation rules. Key primitives: - Agent profiles: Each agent has a model selection (GPT-4, Claude, local LLM), a harness (the execution environment), and a skill set (allowed tools and actions). - Goal hierarchies: Top-level business goals decompose into tasks. Agents claim tasks based on skill matches and availability. - Budget tracking: Per-agent cost attribution for API calls, tool usage, and compute time. Hard caps prevent runaway spending. - Permission boundaries: Agents inherit permissions from their position in the org chart. A junior agent can't approve expenses or access production databases. The UI looks like a task manager. The backend is a state machine that routes work, enforces budgets, and logs every decision for audit trails. Architecture: Node.js Orchestrator + React Dashboard Paperclip runs as a Node.js server that manages agent state and a React frontend for human oversight. The server handles task assignment, budget enforcement, and tool permission checks. The UI shows real-time agent activity, cost breakdowns, and task progress. Component breakdown: | Component | Responsibility | State Persistence | |-----------|---------------|-------------------| | Orchestrator | Routes tasks to agents, enforces budget caps, checks permissions | PostgreSQL for task history, agent profiles, and audit logs | | Agent Harness | Executes agent code (local process, Docker container, or remote API) | Ephemeral; logs sent to orchestrator | | Skill Registry | Maps tool names to permission requirements and cost models | JSON config files, hot-reloadable | | Budget Tracker | Aggregates costs per agent, per task, per time window | Time-series database (InfluxDB or Prometheus) | | Audit Logger | Records every task assignment, delegation, tool call, and budget decision | Append-only log (S3 or local file) | The orchestrator maintains a task queue. When a new goal arrives, it decomposes into subtasks and assigns them to agents based on skill matches. If an agent hits its budget cap, the orchestrator pauses its tasks and escalates to a supervisor agent (or a human). Task Routing and Delegation Paperclip's task assignment uses a skill-matching algorithm. Each task declares required skills (e.g., "web scraping", "data analysis", "content generation"). Each agent declares available skills. The orchestrator finds the best match and assigns the task. Delegation rules: - Agents can delegate tasks to subordinates in the org chart. - Delegation requires explicit approval if the subordinate's budget cap is lower than the task's estimated cost. - Escalation happens automatically when an agent can't complete a task (missing skills, budget exhausted, or explicit failure signal). The system tracks delegation chains. If Agent A delegates to Agent B, who delegates to Agent C, the audit log shows the full path. This matters for debugging and cost attribution. Example delegation flow: // Simplified task assignment logic async function assignTask(task: Task, orgChart: OrgChart): Promise<Agent> { const candidates = orgChart.agents.filter(agent => agent.skills.includes(task.requiredSkill) && agent.budgetRemaining >= task.estimatedCost ); if (candidates.length === 0) { // Escalate to supervisor const supervisor = orgChart.getSupervisor(task.assignedTo); return assignTask(task, orgChart, supervisor); } // Pick agent with lowest current workload const selected = candidates.sort((a, b) => a.activeTasks.length - b.activeTasks.length )[0]; await logDelegation(task, selected); return selected; } Budget Tracking and Cost Attribution Every tool call, API request, and compute minute gets logged with a cost estimate. Paperclip aggregates these costs per agent, per task, and per time window. Budget caps are hard limits: when an agent exhausts its budget, the orchestrator pauses its tasks. Cost tracking flow: 1. Agent requests a tool call (e.g., "search web for competitor pricing"). 2. Orchestrator checks agent's remaining budget. 3. If budget allows, orchestrator estimates cost (based on tool's cost model) and reserves that amount. 4. Tool executes, returns actual cost. 5. Orchestrator updates agent's budget and logs the transaction. The system supports budget rollover (unused budget carries to next period) and budget pooling (team budget shared across agents). These are config options, not hardcoded policies. Budget enforcement table: | Scenario | Behavior | Escalation Path | |----------|----------|-----------------| | Agent under budget | Task proceeds normally | None | | Agent at 90% budget | Warning logged, task proceeds | Notify supervisor | | Agent at 100% budget | Task paused, no new assignments | Escalate to supervisor for budget increase or task reassignment | | Task exceeds agent's total budget | Task rejected at assignment time | Escalate to higher-tier agent or human | Permission Boundaries and Tool Access Agents inherit permissions from their org chart position. A junior agent can read from the knowledge base but can't write. A senior agent can approve expenses up to $500. A supervisor can access production databases. Permissions are declarative. Each tool declares required permissions (e.g., "database:write", "expense:approve"). Each agent has a permission set. The orchestrator checks permissions before allowing tool calls. Permission check example: // Before executing a tool call async function checkPermissions(agent: Agent, tool: Tool): Promise<boolean> { const requiredPerms = tool.requiredPermissions; const agentPerms = agent.permissions; const hasAllPerms = requiredPerms.every(perm => agentPerms.includes(perm)); if (!hasAllPerms) { await logPermissionDenial(agent, tool, requiredPerms); return false; } return true; } Permission violations get logged to the audit trail. Repeated violations trigger alerts. This prevents agents from accidentally (or maliciously) accessing resources they shouldn't touch. State Persistence and Audit Trails Paperclip persists agent state, task history, and audit logs to PostgreSQL. Long-running tasks survive server restarts. The audit log is append-only: every task assignment, delegation, tool call, budget decision, and permission check gets recorded with a timestamp and agent ID. Audit log schema (simplified): CREATE TABLE audit_log ( id SERIAL PRIMARY KEY, timestamp TIMESTAMPTZ NOT NULL, agent_id VARCHAR(255) NOT NULL, event_type VARCHAR(50) NOT NULL, -- 'task_assigned', 'tool_called', 'budget_exceeded', etc. task_id VARCHAR(255), tool_name VARCHAR(255), cost_cents INTEGER, permission_checked VARCHAR(255), metadata JSONB ); The audit log supports compliance requirements (who did what, when, and why) and debugging (why did this task fail? which agent made this decision?). You can query it to generate cost reports, permission violation summaries, or agent performance metrics. Multi-Tenant Considerations Paperclip supports multiple teams (tenants) on a single server. Each team has its own org chart, budget pool, and permission set. Agents in Team A can't see tasks or data from Team B. Isolation mechanisms: - Database-level row-level security (RLS) filters queries by tenant ID. - Agent harnesses run in separate processes or containers (Docker, Firecracker). - Tool calls include tenant context; tools enforce tenant boundaries. This matters for SaaS deployments or large organizations with multiple departments. You don't want the marketing team's agents accessing the finance team's data. Failure Modes and Observability Paperclip exposes metrics for agent health, task queue depth, budget burn rate, and permission denial rate. The dashboard shows real-time alerts when agents fail, budgets run low, or tasks stall. Common failure modes: - Agent crashes: Harness process dies. Orchestrator detects via health check, reassigns tasks to another agent. - Budget exhaustion: Agent hits cap mid-task. Orchestrator pauses task, escalates to supervisor. - Permission denial: Agent tries to call a tool it can't access. Orchestrator logs denial, returns error to agent. - Task timeout: Agent doesn't complete task within deadline. Orchestrator cancels task, reassigns to another agent. The system integrates with Prometheus for metrics and Grafana for dashboards. You can set up alerts for budget overruns, high permission denial rates, or task queue backlog. When to Use Paperclip Use Paperclip when: - You need to manage multiple agents with different skill sets and budget constraints. - You want audit trails for compliance or debugging. - You need permission boundaries to prevent agents from accessing sensitive resources. - You're running long-running workflows that span multiple sessions. Avoid Paperclip when: - You have a single agent or a simple pipeline (too much overhead). - You need real-time, low-latency coordination (the orchestrator adds routing latency). - You want to run agents in a fully decentralized way (Paperclip assumes a central orchestrator). Technical Verdict Paperclip solves the operational problems that emerge when you move from one-off agent demos to production multi-agent systems. The org chart abstraction is opinionated but practical: it maps cleanly to how organizations already think about delegation, budgets, and permissions. The stack is production-ready for teams that need cost control, audit trails, and permission enforcement. The Node.js + React architecture is straightforward to deploy and extend. The skill-matching and budget-tracking primitives are reusable building blocks. The main trade-off is centralization. The orchestrator is a single point of failure and a potential bottleneck. If you need fully decentralized agent coordination, look elsewhere. But if you're building a system where agents need to operate within organizational constraints (budgets, permissions, hierarchies), Paperclip exposes the plumbing you'll need. Source Links - Paperclip GitHub Repository - Paperclip Documentation - Paperclip Discord

  6. Paperclip’s October release adds persistent agent files across tasks and support for syncing company skills from GitHub repositories. The interesting product question: how do teams keep shared agent instructions useful as they grow? github.com/paperclipai/paperclip/releases

    2 likes · 115 views
  7. YouTube · · YouTube description · MANO RISE · paperclipai/paperclip

    paperclip github #business #sapsftraining #automobile . #aitools #education #ai paperclipai/paperclip

    4 likes · 2 comments · 55 views
  8. Give your AI agents an org chart, budgets and a boss with Paperclip, the open-source way to run Claude Code, Codex and friends like a company 📎 🔗 paperclipai/paperclip Credits: paperclipai/paperclip #paperclip #typescript #aiagents

  9. Meet Kowalski --- title: Meet Kowalski published: true description: A small team of AI agents on your own computer, on a free local model. You hand it a chore, it hands you a file. tags: opensource, rust, ai, localllm canonical_url: kowalski.yarenty.com cover_image: kowalski.yarenty.com --- October 2026 This post first appeared on the Kowalski blog. It's written for people who have never heard of Kowalski, so let's start at the beginning. Kowalski is a small team of AI agents that lives on your own computer and does chores for you. Not a chatbot you have to keep talking to, and not a platform you have to administer. You hand it a job, it does the job, and you get a file back. It's open source (MIT), written in Rust, and it runs on a free local model through Ollama, or on any hosted model with your own key. There are two well-known open-source projects in the same neighbourhood, and they help to explain what Kowalski is by what it isn't. OpenClaw is a personal assistant you chat with: it gets wide access to your machine and your apps, and it works best with a big frontier model behind it. Paperclip goes one level up: it runs a whole "company" of paid coding agents, with org charts, budgets and approvals. Both are impressive. Kowalski makes the opposite bet. Each job is a small, fixed pipeline you can read. The agents only touch the one folder you give them, and they ask before doing anything risky. The hard parts (fetching pages, running SQL, building spreadsheets) are ordinary code, so a small free model on a laptop is enough. The result is cheaper (€0 a run on a local model), more private (nothing leaves your machine), and easier to trust (every number comes from a query, not from the model's imagination). Penguins: no lobsters, no corporation. So what do you actually give it? A chore: a question about a spreadsheet, the five pages you read every morning, a document that just landed in your inbox, a topic to research. Kowalski hands the chore to a horde, a short pipeline of specialised agents we call penguins, and gives you back something you can use: a report workbook, a one-page brief, a note in your Obsidian vault. !A first run of kowalski: install, Setup, connect tableski, drop a workbook, get answers What it does today Five hordes come with it:h - Spreadsheet analyst. Ask questions about your workbooks in plain words and get a report workbook back. The model writes the SQL; every number comes from a query engine, never from the model. - Morning brief. The pages you follow, read for you: the few items worth your time, why each matters to you, and the link. It can run every weekday at 7:00. - Folder watcher. Drop a document into its inbox and get a note: what it is, the key facts, and what to do by when. - Knowledge compiler. Ask a question and list the pages that should answer it. You get one note with its sources, saved straight into your Obsidian vault. - URL summarizer. A list of links in, a short summary of each out. A horde is a folder of plain Markdown files, so you can read exactly what each penguin is told, copy one and change it, or have kowalski build a new one with you. How it got here Kowalski started on 4 March 2025 as a weekend experiment: a Rust program that could talk to a local Ollama model. Here's the short version of what happened next. - 2025: big plans. By the summer it had a five-layer architecture, a crate per agent, three databases and a federation layer for a distributed system nobody was running. It was a good exercise and a bad product: nothing in it could be handed to someone who doesn't write Rust. - Early 2026: smaller. 1.0 deleted most of that on purpose. One agent type instead of many crates, SQLite instead of three databases, an HTTP server and a web UI, and tools reached through MCP. - Spring 2026: hordes. The idea that gave the project its shape: a pipeline as a folder of Markdown. Builders, graphs with loops, then triggers (cron, file watch, webhook) and durable runs that survive a restart. - September 2026: the product. One binary with the UI inside, a three-question Setup, ready-made hordes, and data steps that keep numbers out of the model. Since 2.5: a signed Mac app you install by dragging it to Applications. Eighteen months and more than 700 commits later, the pattern is clear: almost every good change removed something. The long version of that story is Eighteen months of Kowalski. Who it's for People with recurring chores and data they'd rather keep on their own machine: a small business owner with a sales workbook, an engineer who reads the same five sites every morning, a researcher whose notes live in Obsidian. If you write code, a horde is a folder you can version, review and share. Try it On a Mac, download Kowalski.dmg, drag it to Applications and open it. On a Mac or Linux, from a terminal: curl -fsSL https://raw.githubusercontent.com/yarenty/kowalski/main/install.sh | bash kowalski Your browser opens on Setup: pick a model, a folder, and optionally your tableski account. Then pick a horde and send it in. Everything is at yarenty/kowalski. If something goes wrong, open an issue. Kowalski does the analysis; I do the apologising. And if everything works first time, just smile and wave, boys. Smile and wave. 🐧 Originally published on the Kowalski blog.

  10. Paperclip’s latest release adds private agent tasks, persistent identities, external model gateways and more detailed cost accounting. paperclipai/paperclip

    2 likes · 2 reposts · 886 views

Achievements

Awards earned by the repositories paperclipai owns. Select an award to embed its badge.

Gold
31st place finishes
Silver
12nd place finishes
Bronze
43rd place finishes
Awarded repositories
1Owned by paperclipai now

Repositories

Ordered by gold, then silver, then bronze.

  1. paperclipBest: 1st by stars, 2,528 stars314

Every award

Top-three finishes of these repositories in completed trending days and weeks, newest first.

  1. paperclip
  2. paperclip
  3. paperclip
  4. paperclip
  5. paperclip
  6. paperclip
  7. paperclip
  8. paperclip

A repository earns an award for finishing first, second or third in a completed trending day or week, ranked by new stars, new forks or pull request authors. Awards follow a repository through renames and transfers and count toward its current owner. Weeks run Monday to Sunday, UTC.