we ditched openclaw. claude got the job.
we didn’t switch from openclaw to claude because we got bored.
we switched because the risk got too loud.
openclaw was useful. moved fast. made model access feel simple. but simple stopped meaning safe.
when your agent is running customer work on a real server, “probably fine” isn’t a strategy. it’s how you wake up to broken auth, suspended accounts, and a support thread nobody wants to read.
so we moved 5dive’s preferred agent to claude code. not a vibe shift. an operating call.
the risk story
the big one: subscription oauth got hot.
anthropic’s claude code legal page is blunt now. building a third-party product? use api keys or a supported cloud provider. routing free, pro, or max credentials through someone else’s harness isn’t permitted, and they can enforce that without notice.
that’s a problem because openclaw-style setups often ran on exactly that pattern: take a user’s subscription login, put it behind a third-party wrapper, let the wrapper drive the model.
in april 2026 the grace period ended. venturebeat reported claude pro and max limits would stop covering third-party agent tools like openclaw as of april 4.
translation: if you were running production-ish work that way, you were one enforcement sweep from downtime.
google showed the same shape with antigravity/gemini oauth. different vendor, same lesson. consumer subscription auth isn’t a stable foundation for somebody else’s harness.
that’s not edgy. that’s just bad ops.
wrapper risk isn’t theoretical
the deeper issue wasn’t one policy page. it was the shape of the dependency.
when the harness isn’t the model vendor’s own product, every model update, auth update, safety update, rate-limit change, and billing tweak becomes a coin flip.
maybe the shim still works. maybe it breaks on a friday. maybe the vendor decides the traffic pattern isn’t okay anymore. maybe the user gets blamed for using a tool that looked normal yesterday.
claude code is anthropic’s product. it’s the thing they’re paid to keep working. they can change it, but they can’t deprecate themselves without owning the blast radius.
that’s the difference.
we wanted claude anyway
the risk forced the move. the product made it easy to stay.
claude code isn’t a thin chat wrapper. it has the stuff serious agents need:
- skills for reusable behavior.
- hooks for guardrails, notifications, workflows.
- subagents for splitting work without losing the main thread.
- mcp for wiring real tools instead of pretending copy-paste is automation.
- plugins and connectors for the stuff we’d otherwise have to build ourselves: vercel, linear, github, private mcp servers, our own telegram workflows.
we rewired a lot of 5dive’s ops around that stack in a weekend.
that’s the test. not “does the demo look cool.” it’s “can we put our own business on it without inventing half the platform first.”
claude passed.
long context actually matters here
agent work dies when context collapses mid-incident.
you’re debugging deploy logs, service state, config, repo history, a half-finished patch, three earlier decisions. then the model forgets the important part and starts cosplaying a fresh intern.
claude’s 1m context window changes the feel of long ops sessions. room to keep the incident, the codebase, and the plan in one place for longer.
not forever. not magically. but long enough that real work feels less fragile.
for us that beats a benchmark flex.
billing got cleaner
the old subscription-harness game was always weird. everyone knew it.
paying one company for a consumer plan, handing that login to a different tool, and hoping the usage pattern stays acceptable isn’t a business foundation. it’s a loophole with a countdown.
claude gives us cleaner lanes:
- use claude code where claude code is the product,
- use api keys where api keys are the right contract,
- use cloud or partner routes where teams need that setup,
- stop laundering production usage through consumer auth.
less cute. also how grown-up systems stay online.
what changed in 5dive
claude is now our preferred agent. openclaw isn’t the default we recommend for serious work.
want to experiment? cool. we’re not here to police your playground. but for customer vms, long-running agents, telegram/discord workflows, subagents, skills, hooks, anything you expect to survive a bad week. claude’s the lane.
the upside is simple: less auth drama, less wrapper risk, more to build on, clearer billing, and a model vendor whose own product is in the loop.
we didn’t ditch openclaw because it was uncool. we ditched it because 5dive agents need to keep working after the hype thread ends.
claude’s the boring choice now. and boring is what you want when the agent has the keys.
Sources: Anthropic Claude Code legal and compliance, VentureBeat on the April 4, 2026 OpenClaw change, reported Google/OpenClaw enforcement, Claude plugins, Claude MCP docs, Claude 1M context announcement.