OpenClaude CLI: Multi-Provider Coding Agent Setup

The problem

You may want a coding agent without locking the model provider to one cloud harness. Demand is strong for keeping a Claude Code-style terminal UX—tools, agents, MCP, slash commands—while switching across OpenAI-compatible APIs, Gemini, Ollama, and similar backends.

📑Table of Contents

1. What OpenClaude Is (a multi-provider terminal harness)
2. Install path (npm → /provider → smoke test)
3. Providers, features, and practical levers
4. Comparison axes vs Claude Code / OpenCode / Gemini CLI / Codex-style tools
5. License, security, and pre-adoption checklist
6. FAQ
7. Summary and next actions

---

What OpenClaude claims to be

OpenClaude (Gitlawb/openclaude) is an open-source coding-agent CLI for cloud and local model backends. The official README describes one terminal-first workflow for prompts, tools, agents, MCP, slash commands, and streaming.

---

Research-time scale

At research time, the install and scale picture is:

• Distribution: npm package @gitlawb/openclaude
• Runtime: Node.js ≥22
• Stable line: v0.26.0 (GitHub Release published 2026-07-27)
• GitHub scale: roughly 30.4k stars / 8.9k forks

---

Boundaries

Boundaries matter.

• The project is a substantially modified fork derived from Claude Code and is not affiliated with Anthropic
• The README Disclaimer/License state that OpenClaude modifications are MIT while derived Claude Code remains Anthropic’s
• It is not the same product as OpenCode (different repository and distribution)

---

What this article covers

This article covers (1) official positioning and install, (2) provider setup and operational tips, (3) comparison axes versus Claude Code / OpenCode / Gemini CLI / Codex-style tools, and (4) license, secrets, and a pre-adoption checklist so you can trial or defer with a clear next step.

---

What OpenClaude Is (a multi-provider terminal harness)

OpenClaude is not a thin chatbot wrapper or an IDE-only plugin. It is a terminal agent loop that reviews code changes while routing work across multiple model providers.

Official definition and supported providers

The official README lists the following under Supported Providers:

• OpenAI-compatible APIs
• Gemini
• GitHub Models
• Codex OAuth / Codex
• Ollama
• Atomic Chat
• Bedrock / Vertex / Foundry, and more

The project site (openclaude.gitlawb.com) emphasizes a reviewable terminal agent loop and multi-model positioning.

README “Why OpenClaude” highlights include:

• one CLI UX across providers
• guided /provider profiles
• bash/file/grep/glob/agents/tasks/MCP/web tools
• a repo-bundled VS Code extension path

---

Config cutover (~/.openclaude)

Config cutover is operationally important.

• Current README defaults: ~/.openclaude and ~/.openclaude.json
• By default it does not read ~/.claude, project .claude/, or CLAUDE_CONFIG_DIR
• Migrate only intentional config
• Do not bulk-copy Claude Code auth

Some third-party docs still mention Claude Code path reuse. Treat current official cutover as source of truth, and use third parties as ecosystem adoption evidence.

---

Independent positioning

Independent sources align on positioning.

• Actionable Notes: lists OpenClaude among post-leak runnable forks
• Sébastien Dubois and catalog-style writeups such as OpenApps: multi-provider fork that keeps Claude Code-adjacent UX
• ByteRover CLI tools docs: OpenClaude as a connector target—evidence beyond GitHub/npm alone

Source

---

One-line summary table

---

Product and distribution

Item	Detail	Source
Product	Multi-provider coding-agent CLI	GitHub README
Distribution	npm @gitlawb/openclaude / AUR openclaude	npm / AUR / README
Runtime	Node.js ≥22 (source builds may use Bun separately)	README / npm engines

---

Scale, status, and license

Item	Detail	Source
Latest stable (research time)	v0.26.0 (2026-07-27)	GitHub Releases
Scale (research time)	~30.4k★ / 8.9k forks	GitHub API 2026-07-28
Official status	Unofficial independent community project	README Disclaimer
License	MIT on modifications; derived Claude Code remains Anthropic’s	README License

Sources: Gitlawb/openclaude, npm @gitlawb/openclaude, Release v0.26.0 (as of July 2026).

---

Install path (npm → /provider → smoke test)

The safest first success path is: version-checked npm install → trial keys only via /provider → small task in a disposable repo → review with /diff.

Prerequisites

• Node.js ≥22
• rg (ripgrep) on PATH. If missing after install, install system rg and restart the shell.

---

Official Quick Start style steps

1. npm install -g @gitlawb/openclaude@latest
2. Confirm 0.26.x with openclaude --version and npm view @gitlawb/openclaude dist-tags
3. Start with openclaude
4. Run /provider in-session (profiles land in .openclaude-profile.json-style files). GitHub Models uses /onboard-github
5. Optional: /init in a project → small task → /diff (same arc as the TechLatest masterclass)

---

Env shortcuts (official)

• OpenAI-style: CLAUDE_CODE_USE_OPENAI=1 + OPENAI_API_KEY + OPENAI_MODEL
• Ollama: add OPENAI_BASE_URL=http://localhost:11434/v1 and a model name (notes mention native chat API and 32768 context expectations)
• No automatic .env load. Prefer /provider, or openclaude --provider-env-file .env

On Arch Linux, the AUR package openclaude (0.26.0-1 at research time, updated 2026-07-27) is community packaging evidence beyond npm.

• Example: paru -S openclaude
• Meaning: a non-npm distribution trail that tracks the same line

---

Long-running jobs and VS Code

Long jobs use:

• openclaude --bg "...", plus ps / logs / kill
• local child processes, not a network daemon
• metadata usually under ~/.openclaude/bg-sessions/
• VS Code integration as vscode-extension/openclaude-vscode in the repo

---

Common pitfalls

Common pitfalls:

• Node below 22
• missing rg
• committing plaintext API keys in profile/config files
• expecting Anthropic-only features on every provider

Source

---

Providers, features, and practical levers

How to read the provider table

Read the provider table as setup paths plus notes, not a promise of feature parity.

• OpenAI-compatible endpoints cover many /v1 hosts (OpenRouter, DeepSeek, Groq, Mistral, LM Studio, and similar)
• Capabilities are asymmetric across providers

---

Cost, web search, and headless

For cost control, route subagents with these keys in ~/.openclaude.json:

• agentModels
• agentRouting

Other practical levers:

• GitHub Copilot-related notes: serializing subagents to limit Premium Requests
• Non-Anthropic web search: may fall back to DuckDuckGo with rate-limit/ToS caveats
• Optional Firecrawl for stability
• Headless integration: gRPC server path (npm run dev:grpc) for other apps/CI

---

Independent hands-on evidence

Independent hands-on evidence:

• TechLatest: install → /provider → many slash commands → permissions → /init → /diff /rewind /cost → --bg
• Zenn (mskbhd): LSP diagnostics / /diagnostics on v0.23-class builds via OpenRouter (watch interactive vs -p differences)
• Practitioner notes: per-agent assignment in ~/.openclaude.json and plaintext key storage risks

---

Do-not-over-expect frame

Official provider notes are the right “do not over-expect” frame:

• smaller models struggle with multi-step tools
• prefer strong tool/function-calling models

---

Comparison axes vs Claude Code / OpenCode / Gemini CLI / Codex-style tools

The goal is role split—official harness versus multi-provider fork—not a scoreboard. Do not conflate OpenClaude with OpenCode.

Comparison table

---

Provider, UX, and config

Axis	OpenClaude	Claude Code (official family)	OpenCode / other harnesses	Decision question
Providers	Many backends in one CLI	Primarily Anthropic official path	Product-specific model policy	Do you need to unlock providers?
UX ancestry	Heavily modified Claude Code lineage	Official product	Often different lineage	Keep Claude Code-like terminal habits?
Config dirs	Default ~/.openclaude after cutover	~/.claude family	Product-specific	Can you avoid config bleed?

---

Extensibility, legal, and local LLMs

Axis	OpenClaude	Claude Code (official family)	OpenCode / other harnesses	Decision question
Extensibility	MCP / skills / agents / slash	skills / MCP / hooks, etc.	Product-specific	Reuse existing MCP assets?
Legal/trademark	Unofficial + dual-layer license	Official terms	Each OSS/commercial license	Will legal review pass?
Local LLMs	Ollama called out in official quickstarts	Product-policy dependent	Sometimes stronger	Is offline/on-prem required?

Sources: OpenClaude README Disclaimer/Providers plus independent notes (dsebastien, OpenApps, note.com-style writeups), July 2026 research window. OpenCode is a separate product.

---

Selection heuristics

Heuristics:

• Official support and in-policy Anthropic operation first → Claude Code official
• Keep Claude Code-adjacent terminal UX while testing other providers → trial OpenClaude in isolation
• Want a different architecture/UX → compare OpenCode or vendor CLIs separately (this article centers OpenClaude)

---

Related reading and discovery signals

For background on harness design and safer official CLI operation, see related pieces such as the Claude Code harness layering guide, Claude Code vs Codex CLI comparison, and Claude Code security setup on this site.

Social chatter about post-leak forks is a discovery signal only—not factual evidence by popularity.

---

License, security, and pre-adoption checklist

Disclaimer and dual-layer license

Start with the README Disclaimer and License.

• The project is unaffiliated with Anthropic; Claude/Claude Code are Anthropic trademarks
• Modifications are MIT; derived Claude Code remains Anthropic’s—a dual layer
• Enterprise adoption needs primary LICENSE review and legal judgment

---

Secrets

Secrets policy to lock early:

• never commit /provider profiles or ~/.openclaude.json
• assume plaintext keys may exist in config; control directory permissions and backups
• never put production keys or customer data in the first trial

---

Permissions and web search

Permission-wise, agents can run bash and edit files.

• Habituate permission gates and /diff review
• Understand DuckDuckGo fallback risks for web search
• Consider Firecrawl if needed

---

Reader checklist (next actions)

• ✓ Decide whether the goal is unlock providers or keep official support first
• ✓ Prepare Node.js ≥22 and ripgrep
• ✓ After npm i -g @gitlawb/openclaude@latest, confirm 0.26.x
• ✓ Configure trial keys only via /provider and keep them out of git
• ✓ Use a disposable repo or worktree for the first task
• ✓ Review with /diff before any commit
• ✓ Read LICENSE/Disclaimer against company policy
• ✓ Align the team that Anthropic-only features may be missing on other providers
• ✓ Keep production data and credentials off the trial path

If unsure, run one harmless task on Ollama or a cheap OpenAI-compatible endpoint, then add MCP/subagents/background only after that succeeds.

---

FAQ

Q1. Is OpenClaude an official Anthropic product?

No. The README Disclaimer states it is an independent community project and unaffiliated.

---

Q2. Is it the same as OpenCode?

No. Gitlawb/openclaude is a different repository and distribution (@gitlawb/openclaude on npm). Similar names cause confusion.

---

Q3. What is the minimal install path?

With Node.js ≥22: npm install -g @gitlawb/openclaude@latest → openclaude → /provider. Confirm the 0.26.0 line via Release/npm.

---

Q4. Can I reuse ~/.claude as-is?

Under current cutover defaults, no automatic read of ~/.claude. Move only intentional custom config into the OpenClaude side; avoid bulk-copying auth.

---

Q5. What matters most for company use?

Dual-layer licensing, trademarks, where secrets live, and agent file/shell permissions. Do not ship to production without legal review.

---

Q6. Can I run local LLMs only?

Yes—README includes an Ollama quickstart. Official provider notes still warn that smaller models may struggle with multi-step tools.

---

Summary and next actions

Takeaway

OpenClaude is an OSS CLI for using a Claude Code-adjacent terminal agent UX across multiple providers. The v0.26.0 npm line and large community scale make it easy to try—but because it is unofficial, legal review, secrets hygiene, and permission review travel together.

---

Pick one next action

Pick one next action:

1. Trial: disposable repo → npm install → /provider → one small refactor → /diff
2. Compare: use the table axes to choose official Claude Code continuity vs OpenClaude trial
3. Defer: if official support/terms are mandatory, stay on Claude Code and only track the fork

---

Primary sources

Primary sources: GitHub README, Release v0.26.0, npm, project site.

Related articles:

• 11300
• 13580
• 13623