openclaude-improved

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

OpenClaude — Improved Version

runs anywhere. uses anything.

An open-source coding agent for the CLI. Cloud APIs, gateways, and local models — same tools, same agents, same workflow.

图片：node 图片：bun 图片：platform 图片：license

Install · Quick start · Providers · Sessions · Config · Docs

---

Install

Built from source. Windows examples below; macOS and Linux are identical minus the shell syntax.

Prerequisites — Node >=22 (enforced by engines.node) and Bun.

node --version
bun --version

No Bun? winget install Oven-sh.Bun or bun.sh.

Build and link

cd openclaude-main
bun install
bun run build
npm install -g .
openclaude

That’s it — openclaude is now on your PATH.

---

Quick start

Run /provider inside OpenClaude for guided setup with saved profiles — this is the recommended path. Credentials land in .openclaude-profile.json.

Prefer env vars? Pick one:

OpenAI

$env:CLAUDE_CODE_USE_OPENAI="1"
$env:OPENAI_API_KEY="sk-..."
$env:OPENAI_MODEL="gpt-4o"
openclaude

Ollama — local, no key

$env:CLAUDE_CODE_USE_OPENAI="1"
$env:OPENAI_BASE_URL="http://localhost:11434/v1"
$env:OPENAI_MODEL="qwen2.5-coder:7b"
openclaude

OpenClaude talks to Ollama’s native chat API and requests a 32768-token window per request, so same-session history isn’t silently trimmed by the OpenAI-compat shim. Override with OPENCLAUDE_OLLAMA_NUM_CTX or OLLAMA_CONTEXT_LENGTH.

GitHub Models

Run /onboard-github inside OpenClaude. Interactive, credentials saved.

macOS / Linux syntax

export CLAUDE_CODE_USE_OPENAI=1
export OPENAI_API_KEY=sk-...
export OPENAI_MODEL=gpt-4o
openclaude

Project .env files are not auto-loaded. Use --provider-env-file .env for provider vars, or export runtime knobs from your shell.

---

Providers

Provider	Setup	Key detail
OpenAI-compatible	/provider · env	Any /v1 server — OpenRouter, DeepSeek, Groq, Mistral, LM Studio
Ollama	/provider · env	Local, no API key
Gemini	/provider · env	API key only
GitHub Models	/onboard-github	Saved credentials
Codex / Codex OAuth	/provider	Browser sign-in or existing Codex CLI auth
Gitlawb Opengateway	default · /provider	Startup default on fresh installs; get a key
Bedrock · Vertex · Foundry	env	Anthropic-family routes

Full provider list (12 more)

Provider	Endpoint / key	Default model
Z.AI GLM Coding Plan	api.z.ai/api/coding/paas/v4 · OPENAI_API_KEY	glm-5.2
AI/ML API	api.aimlapi.com/v1 · AIMLAPI_API_KEY	gpt-4o
Hicap	api-key auth, Responses mode for gpt-*	—
Fireworks AI	FIREWORKS_API_KEY	276 curated models
LongCat	api.longcat.chat/openai/v1 · LONGCAT_API_KEY	LongCat-2.0
ClinePass	api.cline.bot/api/v1 · CLINE_API_KEY	5h / weekly / monthly caps
OpenCode Zen	opencode.ai/zen/v1 · OPENCODE_API_KEY	48 models, PAYG
OpenCode Go	opencode.ai/zen/go/v1 · OPENCODE_API_KEY	13 models, $10/mo
Xiaomi MiMo	mimo.mi.com · MIMO_API_KEY	mimo-v2.5-pro
NEAR AI	cloud-api.near.ai/v1 · NEARAI_API_KEY	Claude / GPT / Gemini + TEE
Cloudflare Workers AI	api.cloudflare.com/.../ai/v1 · CLOUDFLARE_API_TOKEN	—
Atomic Chat	/provider · bun run dev:atomic-chat	Auto-detects loaded models

Gotchas worth knowing

• Anthropic-only features don’t exist on every backend. Tool quality tracks model quality — small local models struggle with long multi-step tool loops.
• Some providers cap output below CLI defaults; OpenClaude adapts where it can.
• Opengateway uses one base URL — switch models with /model, don’t pin the URL to /v1/xiaomi-mimo.
• GLM reasoning: glm-5.2?reasoning=high, ?reasoning=xhigh, or ?thinking=disabled.
• MiMo uses api-key header auth and has no /usage reporting yet.
• GitHub Copilot serializes sub-agents by default to save Premium Requests — see agent routing.

---

What you get

Tools	Bash, read/write/edit, grep, glob, agents, tasks, MCP, slash commands
Streaming	Live tokens and tool progress
Tool loops	Multi-step: model call → execution → follow-up
Vision	URL and base64 images where the provider supports it
Repo map	PageRank-ranked structural map, auto-injected behind the REPO_MAP flag. Inspect with /repomap (docs)
Agent routing	Per-agent provider/model overrides, maxSteps caps, routable built-ins (Explore, Plan, verification) (docs)
Web	WebSearch via DuckDuckGo free by default; drop in FIRECRAWL_API_KEY for JS-rendered pages
gRPC	Headless bidirectional-streaming server for CI and custom UIs (npm run dev:grpc) (docs)
VS Code	Bundled extension: launch integration, Control Center, in-editor chat, Foundry/Azure config

---

Sessions

openclaude --continue                             # most recent, this directory
openclaude --resume 
openclaude --resume  --fork-session   # branch history, new ID

Forking branches conversation history only — no worktree, no filesystem isolation.

Background sessions

openclaude --bg "fix failing tests"
openclaude --bg --name auth-refactor "refactor auth middleware"

openclaude ps
openclaude logs auth-refactor -f
openclaude kill auth-refactor

Plain local child processes — no daemon, no network service. Metadata and logs live in ~/.openclaude/bg-sessions/. Names are reusable once a session is terminal; use the ID to reach older logs sharing a name. attach currently just points you at logs -f.

---

Config

OpenClaude owns ~/.openclaude/ and ~/.openclaude.json. It does not read ~/.claude, project .claude/ directories, or CLAUDE_CONFIG_DIR. Fresh installs start empty and don’t need Claude Code present.

Migrating from a .claude-era setup? Copy only files you wrote — settings, commands, agents, skills, scheduled tasks — into the matching .openclaude path. Don’t blanket-copy, and don’t move credentials; re-run provider setup instead.

OPENCLAUDE_CONFIG_DIR relocates everything.

---

Buddy

/buddy hatches a truecolor pixel-art companion that stands beside your prompt and fires its signature move on every Enter.

/buddy set robinhood    green archer — arrow shot
/buddy set kaio         full-width energy wave
/buddy set strawhat     stretchy snap-back punch
/buddy set merlin       sparkle stream
/buddy set kage         spinning shuriken
/buddy set ember        dragon fire, real heat gradient
/buddy set corsair      cannonball with smoke trail

Respects prefersReducedMotion, degrades to line art on low-color terminals, /buddy mute silences it. Needs ~100 columns for the full sprite.

---

Development

bun run dev        # build and launch from source
bun test           # full suite

Before opening a PR:

bun run build
bun run smoke
bun test path/to/changed.test.ts
bun run test:coverage          # if you touched shared runtime or provider logic

Other commands and layout

bun run test:coverage:ui                        # rebuild HTML report only
bun run test:provider
bun run test:provider-recommendation
bun run doctor:runtime
bun run verify:privacy
bun run security:pr-scan -- --base origin/main

src/                                   core CLI and runtime
scripts/                               build, verify, maintenance
docs/                                  setup and contributor docs
bin/                                   launcher entrypoints
vscode-extension/openclaude-vscode/    VS Code extension

Coverage lands at coverage/lcov.info plus a browsable report at coverage/index.html.

If startup reports ripgrep not found, install ripgrep system-wide and confirm rg --version resolves in the same shell.

---

Docs

Getting started — Non-technical · Windows · macOS / Linux · Android

Going deeper — Advanced setup · Smart auto-routing · Agent routing · Repo map · gRPC server

---

Contributing

Open an issue first for anything large, so scope is settled before code. Bugs and actionable feature work go to Issues; questions and ideas to Discussions. Security reports: SECURITY.md.

---

MIT for contributor modifications; derived Claude Code code remains Anthropic’s — see LICENSE.

An independent community project. Not affiliated with, endorsed by, or sponsored by Anthropic. “Claude” and “Claude Code” are trademarks of Anthropic PBC.