Outworked v0.3.0 adds iMessage support, built-in browser, and scheduling for Claude Code agents

What Outworked v0.3.0 adds
Outworked is a desktop application that organizes Claude Code agents into a team structure. You provide a goal in plain English, and an orchestrator breaks it into subtasks, assigning them to specialized agents that can work in parallel, communicate with each other, write code, and now handle additional functions introduced in version 0.3.0.
New features in this release
- iMessage channel support: Agents can send and receive text messages through iMessage. Other communication channels can be extended.
- Built-in browser: Agents can navigate websites, interact with web content, take screenshots, and bring findings back to the team.
- Scheduling: Agents can run tasks on timers, cron schedules, or at future times. Example:
0 9 * * *for daily 9 AM execution. - Tunneling: Agents can share local content (like development servers) over the internet using
tunnel_startto generate public URLs. - Enhanced MCP and Skills support: Improved framework for extending agent capabilities through Model Context Protocol servers and custom skills.
- Auto-approval for agent requests: Streamlined workflow for agent-initiated actions.
How it works
You create agents with names, roles, personalities, models, and sprites in an Animal Crossing-style office interface. The orchestrator automatically routes tasks to appropriate agents based on their capabilities. All agent activity is visible in the desktop interface as agents move to desks and perform work.
Example workflows
- Landing page creation: Frontend Dev agent scaffolds a site, writes HTML/CSS/JS, starts a local server. Designer agent reviews in browser, requests tweaks. Frontend Dev makes fixes, runs
tunnel_startfor public URL, then usessend_messageto text the link via iMessage. - Daily GitHub PR summaries: Ops Agent sets up scheduled cron job (
0 9 * * *), queries GitHub MCP server for open PRs each morning, writes summary, sends via iMessage. - Competitor research: Orchestrator assigns Researcher agents to multiple competitors simultaneously. Each uses built-in browser to navigate sites, read docs, extract pricing. Writer agent compiles findings into structured document, commits to repo, can post to Slack.
- GitHub issue triage: Project Manager agent pulls new issues via GitHub MCP server, labels by priority and type, assigns to team members, adds first response comments. For reproducible bugs, assigns Backend Engineer to investigate and open draft PR.
Agent capabilities
- Write and ship code across multiple repositories
- Browse the web for research and data extraction
- Send/receive messages via Slack, iMessage, and other channels
- Run scheduled tasks using cron expressions
- Query databases via PostgreSQL MCP servers
- Manage projects through GitHub, Linear, and other integrations
The application runs locally on Mac and integrates with existing Claude Code setups. Version 0.3.0 represents significant expansion of agent capabilities beyond just coding to include web interaction, communication, and automation.
📖 Read the full source: HN AI Agents
👀 See Also

McPherson AI releases two new QSR operations skills on ClawHub: food cost diagnostics and labor leak auditing
Two new free skills have been published on ClawHub: qsr-food-cost-diagnostic catches COGS issues weekly with a four-lever diagnostic, and qsr-labor-leak-auditor provides daily labor tracking with mid-week alerts to prevent overspend.

Microsoft VibeVoice: 60-Min ASR and 90-Min TTS Models Open-Sourced
VibeVoice is a family of open-source voice AI models from Microsoft including ASR (60-min single-pass, speaker diarization, 50+ languages) and TTS (90-min multi-speaker, real-time streaming). Uses 7.5 Hz continuous speech tokenizers and next-token diffusion.

memv MCP Server: Persistent Structured Memory for AI Agents
memv, an open-source Python memory layer for agents, now ships with an MCP server. It provides five tools for persistent, structured memory with per-user isolation and LLM-optional extraction.

Claude Skills Silently Override Instructions: Undocumented Pitfalls Exposed
User discovers Claude skills silently enforce hard limits on user input via `ask_user_input_v0` (max 3 questions, 4 options each), `Write` overwrites files while `create_file` refuses on Claude.ai, and relative paths in `references/` don't resolve. A community repo catalogs findings.