Open Source Radar — September 27: agents get a workplace
Today's trending list is mostly about where agents work rather than what they think with: an org chart, a fleet console, a team room, and two new ways for them to reach phones and pull requests.
Paperclip (TypeScript, ~88,200 stars, MIT) — A Node server with a React dashboard that treats a group of agents as a company. You declare a goal, staff it with whatever you already run — Claude Code, Codex, Cursor, OpenClaw, or anything that can hold an HTTP endpoint — set budgets, and watch the work from one screen instead of twenty terminal tabs. The README organises it into four pillars: a task manager with approval gates, an org chart with roles and scoped secrets, training and evaluation for the agents themselves, and the runtime underneath. It is management software, and it makes the argument that management is the bottleneck now, not model quality.
Orca (TypeScript, ~79,200 stars, MIT) — An agent development environment built around parallel work rather than one chat window. It runs Codex, Claude Code, OpenCode or Pi side by side, each in its own isolated git worktree, so you can fan a single prompt across five agents and merge the winner. Releases land daily — version 1.4.215 shipped today — and the feature list reads like someone got tired of context switching: worktrees on a remote box over SSH, PRs and Linear boards in the same window, diff annotations that go straight back to the agent, and a design mode that clicks a UI element in a real browser and drops its HTML, CSS and a cropped screenshot into the prompt. Your own coding-agent subscription does the work; Orca is the console, including the phone one that tells you when an agent finishes.
Buzz (Rust, ~34,900 stars, Apache-2.0) — Block's open-source workspace puts people and agents in the same rooms, and the design choice underneath is the interesting part: it is a Nostr relay, so every message, reaction, workflow step, review approval and git event is a signed event in one log, whether the author is a person or a process. Agents get their own keys and their own channel memberships rather than a shared admin token, and can open repos, send patches, review code, run workflows and orchestrate other agents with the same audit trail a teammate gets. What works today is a relay with channels, threads, DMs, canvases, media, search, an audit log, a desktop app, a command-line client and YAML workflows; mobile clients and workflow approval gates are still being wired up, and the README says so plainly. Self-hostable, one relay per community, and the backing of Block gives it distribution most agent-native workspaces never get.
mobile-mcp (TypeScript, ~7,600 stars, Apache-2.0) — An MCP server that hands agents a phone: iOS and Android, simulators, emulators and real devices, without per-platform knowledge. It reads the native accessibility tree instead of looking at pixels, so most taps and form fills need no vision model at all, and it falls back to screenshot coordinates only when a screen gives it nothing usable. Version 1.0.5 landed this week with instructions that tell the client to pick a device first, batch its round trips and release remote devices when it is done — unglamorous fixes that are what makes a long automation run survive. This is the missing half of agent automation: web agents have had a stable harness for a year, and the people doing QA on native apps have been writing brittle scripts by hand.
claude-code-action (TypeScript, ~9,200 stars, MIT) — Anthropic's official GitHub Action, and the version of a coding agent that lives in CI instead of your terminal. It answers questions and implements changes on pull requests and issues, activating on a mention, an issue assignment, or an explicit automation prompt, and it runs on your own GitHub runner against whichever provider you already pay — Anthropic's API, Amazon Bedrock, Google Vertex AI or Microsoft Foundry. The v1.0 line has shipped hundreds of patch releases, most recently on September 25, and the structured output mode is the underrated feature: validated JSON that becomes a workflow output is what lets a review bot gate a merge rather than comment into the void. The caveat is the open-issue list, which is long and growing — broad adoption first, sharp edges second.
Worth watching this week: whether the agent-management layer consolidates around one of these consoles, or every harness ends up shipping its own.
Where does your agent work happen now — a terminal, a dashboard, or a room with other people in it? Tell us in the comments.
Sources: Paperclip (GitHub) · Orca (GitHub) · Orca v1.4.215 release · Buzz (GitHub) · mobile-mcp (GitHub) · mobile-mcp 1.0.5 release · claude-code-action (GitHub) · Claude Code Action (GitHub Marketplace)
