An open gallery for interesting agent-native experiments.

Claude Code • Codex • MCP • Agent UI • Multi-Agent Workflows • Evaluation

We are watching an unusual acceleration curve: frontier models are iterating faster, and the release cadence itself is becoming a signal.

_{Source: Tencent News / OpenAI flagship model release timeline}

Our bet is simple: if agent capability keeps scaling, the surrounding infrastructure should scale with it too.

OpenClaudex is a place to build and collect the small pieces that make that future usable: execution lanes, control surfaces, traces, evaluation harnesses, and multi-agent workflows.

• 🛤️ Agent lanes: agents should work beside humans, not steal their mouse and keyboard.
• 🧑‍✈️ Less human intervention: make AI operate longer with clearer state, recovery, and safety surfaces.
• 🐝 AI swarms: scale from one assistant to many cooperating agents that divide work and report back.
• 🧪 Evaluation traces: collect real rollouts, failures, and compatibility notes instead of only demos.
• 🔌 MCP-first tooling: small, inspectable services that Claude, Codex, and other harnesses can reuse.

Agent-native context recovery across local coding-agent harnesses.

• Cross-harness recall for local Claude Code, Codex, CodeBuddy, and other agent traces
• Async compact jobs for current-session context
• Reviewable context candidates before anything enters the active prompt
• Agent-added source discovery for Cursor, Windsurf, Trae, Cline, Continue, Aider, and internal tools
• VCC-inspired trace compilation and retrieval workflow

Agent skill and CLI toolkit for safe OpenReview submission workflows.

• Submission copilot for authors and research agents
• Author profile matching with affiliation evidence
• Cross-venue transfer preflight with dry-run-first safeguards
• Batch submission scaffolding with explicit confirmation
• Schema-aware OpenReview edits through openreview-py

Open-source background computer use for Claude Code, Codex, and MCP agents on macOS.

• Native Swift MCP server
• App-aware virtual cursor
• Accessibility + screenshot context
• Background execution without stealing the user's mouse and keyboard
• Codex-style behavior, open implementation

Native Apps	Background Work	Feishu / Lark

• Human-AI coexistence: agents should cooperate with human workflows, not hijack them.
• Operational scale: more agents should mean more throughput, not more babysitting.
• Visible autonomy: tool execution should be observable, recoverable, and inspectable.
• Composable infrastructure: Claude, Codex, MCP, and local-first tools should plug together cleanly.

• 🧰 More MCP execution services
• 🧾 Agent behavior trace kits
• 🧩 Compatibility notes for real-world apps
• 🐝 Small harnesses and demos for agent-human coworking

• Follow the org for Claude, Codex, MCP, and agent-native experiments.
• Start with the recent projects: agent-context-manager, openreview-agent, and open-claudex-computer-use
• Report issues or request features in the relevant project repository.