The event connects software agents to physical systems
OpenClaw will bring its maintainers, contributors and hardware partners to Shenzhen for Claw Hack and ClawCon from 4 to 6 December 2026. The first two days are structured as a build event, followed by a conference on open-source agents, open-weight models and physical AI. OpenClaw says core team members will pair-program, review pull requests and troubleshoot with attendees. The event is more than a community gathering because it concentrates several questions that determine whether open agents become dependable tools: integration quality, evaluation, hardware access and the real cost of operation.
Claw Hack has separate software and hardware tracks. Software projects can cover skills, plugins, integrations, local knowledge tools and new interfaces. The hardware track includes devices and sensors from partners such as M5Stack and Seeed Studio, including a robot arm and other physical interfaces. That combination tests an agent outside the forgiving boundary of a chat window. A mistaken answer is inconvenient; an incorrect tool call to a physical system can move equipment, consume energy or damage a workpiece.
The practical value of a build event is rapid feedback between users and maintainers. Setup failures, unclear permissions and brittle plugin assumptions become visible when people try to make a real project under time pressure. Pairing can turn those problems into reproducible issues and reviewed patches. The strongest outcome will not be the most theatrical demo, but fixes that improve installation, diagnostics, recovery and safe defaults for people who were not in the room.
A hackathon can expose real integration friction
Skills and plugins are especially important security boundaries. They extend what an agent knows and what it may do, often by connecting local files, cloud services or command execution. Hackathon teams should document required permissions, external calls and data flows, then publish test cases alongside code. A plugin that works only with a maintainer’s credentials or machine is not a reusable contribution. OpenClaw’s community should reward narrow permissions and understandable failure behaviour as much as feature breadth.
Physical-agent projects need explicit safety design. Developers should define allowed actions, speed or force limits, emergency stops and a simulation path before hardware execution. Sensor readings and model outputs should be treated as uncertain, with deterministic checks around dangerous transitions. Human approval remains appropriate when an action is irreversible or difficult to recover. A successful weekend prototype can demonstrate feasibility, but it should not be confused with a certified control system.
Open models need honest operating economics
ClawCon will include talks and live demonstrations on building and evaluating agents and moving them beyond the screen. Evaluation is a welcome focus because agent demos often hide retries, manual corrections and favourable task selection. Speakers can make the discussion more useful by publishing task definitions, failure rates, hardware configuration and intervention counts. Without that evidence, audiences cannot tell whether a workflow is robust or whether a skilled operator quietly rescued it.
OpenClaw also promises discussion of what it takes and costs to run agents locally and in the cloud. Open weights do not eliminate compute, power, maintenance or support expenses. Local deployment may improve privacy and control but requires suitable hardware and operational skill. Cloud models offer elastic capacity and managed updates while creating recurring charges and external data paths. Honest comparisons should include idle resources, tool calls, retrieval, retries and the human time required to keep a system reliable.
Community work should produce durable artefacts
Shenzhen is a relevant setting because its hardware ecosystem can shorten the path from a software idea to a physical prototype. Partners listed by OpenClaw span model infrastructure, chips, devices, robotics and academic communities. That breadth can encourage interoperable interfaces rather than vertically integrated demonstrations. It can also produce fragmented projects if every team invents a bespoke protocol. Contributors should favour documented standards and small reference implementations that others can reproduce after the event.
Applications close on 25 November, with separate registration for the hackathon and conference. For the wider community, the event’s lasting value will be measured in merged fixes, public documentation, reusable test fixtures and transparent lessons from failed experiments. Maintainers should publish a post-event ledger connecting demonstrations to repositories, issues and release notes, including projects that remain incomplete. Remote contributors would then be able to reproduce work, challenge assumptions and continue development without depending on private conversations from the venue. Hardware projects should include bills of materials and safe simulation instructions wherever licensing permits, alongside exact software versions and supported device configurations. OpenClaw’s open model gives users unusual visibility into how an agent platform evolves. Claw Hack and ClawCon can strengthen that advantage if excitement around robots and models is matched by disciplined permissions, evaluation and maintenance. The best showcase would be an agent ecosystem that becomes easier to inspect and safer to extend after everyone goes home.