OpenClaude

You will like this if you learned Claude Code's habits, slash commands, agents, MCP, streaming, and then found yourself paying for a different provider. The daily trait is /provider: a guided setup that saves profiles, so switching from Groq to a local model is a menu rather than an environment-variable ritual. The README admits tool quality depends on the model and small local models struggle with long tool loops.

Where it hurts is trust: a fork of a vendor's code, maintained by a small group, that upstream can reshape at any time. Pick it to run Claude Code's shape on cheap models. Pick OpenCode for the same on cleaner ground.

The risk is the default. A fresh install's startup provider is Gitlawb Opengateway, a gateway run by the project's maintainer that requires an API key from gitlawb.com, and the provider notes add rules about which base URL to pin. A fork whose first-run path routes prompts through its own author's proxy has a conflict of interest written into the config, whatever the intent.

The consequence: run /provider before the first prompt and choose something you already trust. What it does right: --fork-session is documented as conversation branching only, with an explicit note that it creates no filesystem isolation and no worktree, which is the candour other feature lists lack.

The retrieval and verification paths are documented, mostly as caveats. 1. Context: a repository map ranked by PageRank is injected only when a REPO_MAP flag is set, at a default of 2048 tokens. 2. Web search for non-Anthropic models scrapes DuckDuckGo and may be rate-limited or blocked; Firecrawl is optional. 3. A verification agent exists but is gated behind two feature flags, one of them named tengu_hive_evidence. 4. Ollama requests a 32768-token window per call so history is not silently truncated.

The only benchmark with a published methodology is for launcher startup: thirty warm runs, ten cold, median and IQR reported. Rigorous, and about the wrong thing.

The README has a Partners section before it has a Quick Start, and the ten logos are inference gateways and resellers: AI/ML API, Novita, ApiSmart, Concentrate, Atlas Cloud, Xiaomi MiMo and the rest. That is the model: a popular front end that routes tokens to whoever pays for placement. The repository is mirrored on Gitlawb's own hosting, and the disclaimer states that Claude and Claude Code are Anthropic's trademarks, which is a sentence lawyers wrote.

Moat: none; the upstream is someone else's product. Pricing power: none, the value is free. Likely outcome: a rename, or a quiet fade when upstream changes shape. Position: short, and keep a migration path.

The demo is a Claude Code lookalike on a cheap model. For sixty seats the software is free and the meter is whichever provider each engineer picks, which is sixty meters. Provider profiles and credentials are saved in .openclaude-profile.json, and sixty of those files in sixty checkouts is a secrets audit waiting to happen. The headless gRPC server for CI is started from the repo with npm run dev:grpc, which is a developer command, not a deployment.

If we route it through the GitHub Copilot subscription we already pay for, subagents run serially to conserve premium requests, so throughput drops. No SSO, no audit log, support is Discord. Not yet.

Read LICENSE twice: MIT covers the contributors' modifications, and the derived Claude Code portions remain Anthropic's, so what I can fork is the diff, not the whole. Running local is two exports, CLAUDE_CODE_USE_OPENAI=1 and OPENAI_BASE_URL pointing at my Ollama, with OPENCLAUDE_OLLAMA_NUM_CTX if I want a different window. agentModels and agentRouting in ~/.openclaude/settings.json send each built-in agent to a different model, which is the routing I usually have to write myself. MCP, slash commands and skills survived the fork intact.

Grudging respect: the tools are Claude Code's tools, and they are good. The ownership question is the one I cannot settle from the config.