Claude Code — hands-on review 2 months (fall 2025), then daily use since returning in May 2026
I had scored Claude Code 7/10 and documented my move to OpenCode. The 2026 tool is not the one I left: skills, hooks, sub-agents, persistent memory, MCP — an ecosystem no hand-assembled toolchain replicates. The Anthropic lock-in hasn't moved: it's a cost I chose to pay, while keeping a tested exit route.
Strengths
- + Model quality on Symfony/TypeScript code: solid contextual reasoning, directly usable refactorings
- + Extensibility became real: custom commands/skills versioned in Git, hooks (pre-commit, gates), specialized sub-agents with routing
- + Multi-surface ecosystem: same engine and same configs (CLAUDE.md, MCP) in the terminal, IDE, desktop, web and mobile
- + Persistent memory across sessions and cron tasks on Anthropic's side — workflows impossible to assemble by hand
- + Zero configuration to start — operational in 5 minutes
- + MCP to connect any data source (open protocol)
Limitations
- − Structural vendor lock-in: a single provider, no routing to other models by task or cost
- − Configs (agents, skills, hooks, memory) are in Anthropic's format — none of it is portable to OpenCode or Cursor
- − An order of magnitude more expensive than budget alternatives for comparable model quality on simple tasks
- − Proprietary: the code isn't inspectable, unlike OpenCode (MIT) or Kilo Code (Apache 2.0)
- − If Anthropic changes its terms — it already happened with the subscription restriction in third-party tools — you take the hit. That's the exact reason for my first departure
July 2026 update: the return
This review said in February: “good entry point, but the lock-in becomes a real constraint” — 7/10, migration to OpenCode documented. The lock-in hasn’t changed. The tool has. I came back in May 2026, on a Max subscription, in daily use — and this page owns the round trip rather than rewriting it. The full reasoning is in the dispatch Anthropic, the Apple of AI.
Test context
First test: fall 2025, 2 months, head-to-head with OpenCode on a Symfony 7 + TypeScript/React project of about 80k lines — followed by a migration to OpenCode, documented in this dispatch. Second round: daily use since May 2026, same project, Max subscription.
What it actually does
Claude Code connects the terminal to the Claude model — that was its description in 2025, and it’s become insufficient. The 2026 tool is extensible and multi-surface: custom commands and skills versioned in Git, hooks that gate commits, specialized sub-agents, persistent memory across sessions, MCP servers shared between the CLI, the IDE and the desktop app, cron tasks running on Anthropic’s side with your machine off.
The model’s strength is what it was: on a Symfony service refactoring with business context, the output is coherent and often directly usable.
What works
The integrated ecosystem. My current setup — specialized agents with routing, morning/evening crons that read GitLab and write to Obsidian, hooks that block a commit without review — holds together because everything shares the same engine and the same configs. I had tried the equivalent as an assembly (OpenCode + scripts + n8n): every piece was maintained separately.
Versioned workflows. What was most missing in 2025 — custom commands shareable through Git — now exists as skills. Team knowledge stops being tribal: it lives in the repo, reviewable like code.
Onboarding simplicity, unchanged: one-command install, productive immediately.
What doesn’t work
The lock-in is structural, and it grows with your investment. Every skill, every hook, every accumulated memory is in Anthropic’s format. This isn’t theoretical: the subscription restriction in third-party tools — a unilateral change of terms — is the exact reason for my first departure. The risk didn’t disappear because I came back.
The cost. A Max subscription costs an order of magnitude more than a budget alternative for comparable model quality on mechanical tasks. If you compare model against model, Max doesn’t justify itself — the ROI is elsewhere.
Transparency. The tool remains proprietary: you don’t know what’s logged or how.
Who is it for?
Claude Code is the right choice if you’re willing to invest in the tooling — the integration is what pays, not the model alone. Using it as a plain chat CLI is the worst of both worlds: the price of the garden without the fruit; in that case OpenCode or Kilo Code with a budget model does the same job. And if multi-provider is a governance requirement where you work, the question is already settled: this isn’t your tool.
Recommended for
- ✓Those willing to invest in agentic tooling (agents, hooks, automations) — the ROI is in the integration, not the model alone
- ✓Teams already committed to the Claude ecosystem that want versioned conventions (CLAUDE.md, shared skills)
- ✓Complex architecture tasks where reasoning depth outweighs cost
- ✓First discovery of coding agents — immediate onboarding
Not recommended for
- ✗Basic chat-CLI usage without leveraging the ecosystem — you pay the price of the garden without picking the fruit
- ✗Workflows where multi-provider is a governance requirement — OpenCode remains the reference
- ✗Tight budgets on simple tasks: a ~$18/month alternative does the same mechanical generation work
Frequently Asked Questions
- Why come back to Claude Code after leaving it?
- Not for the model — budget alternatives score close on coding. For the ecosystem: versioned skills and hooks, sub-agents, persistent memory, the same engine on every surface. I documented both the departure and the return in two dispatches — both stay online; the honest part is the full arc.
- Does Claude Code work with Claude Max?
- Yes, that's the main usage mode today. The Max subscription ($100 or $200/mo depending on tier) covers intensive Claude Code use without token-by-token billing. For sustained daily use it's far more predictable than the direct API.
- Is the Anthropic lock-in still a problem?
- Yes, and it has even grown: the richer your setup gets (agents, skills, hooks, memory), the less portable it is. My answer: keep a tested, working multi-provider exit route, and treat the lock-in as a priced cost, not a taboo.
- Can Claude Code work across multiple files at once?
- Yes — it reads the filesystem, edits multiple files per session, has Git access, and can parallelize through sub-agents. On very large codebases, task decomposition remains the key, whatever the context window.