OpenClaude
runs anywhere.

One workflow

Keep prompts, files, shell commands, agents, MCP servers, slash commands and streamed responses in one place.

Many backends

Switch between compatible cloud endpoints and local runtimes such as Ollama without adopting a different CLI for each one.

More control

Route different agents to different models, cap tool-use steps, inspect repo structure, and expose the engine over gRPC.

Prototype with Ollama

Try a local coding model against a small repo, keeping the same agent/tool loop while iterating on model size and quality.

Split work by model strength

Use a cheaper or local model for exploration, a stronger model for planning, and a specialised route for frontend or review tasks.

Put an agent behind your UI

Expose the core engine over bidirectional gRPC, then build a custom interface, CI/CD integration or internal tool around the stream.

Inject repo structure

Use the optional repo map to surface structurally important files and definitions when a project is too large to read linearly.

Package skills

Install or author skills as folders containing a SKILL.md, then let the agent invoke reusable procedures by name.

Run background sessions

Kick off a detached prompt, inspect its status and logs, and keep the terminal available for other work.

Model routing

Map an agent type such as Explore, Plan, frontend-dev or verification to a named model route.

"agentRouting": {
  "Explore": "deepseek-v4-flash",
  "Plan": "gpt-4o",
  "frontend-dev": "zai-default"
}

Step limits

A custom agent can set maxSteps to cap tool-use steps. When the limit is reached, the agent is asked for a concise summary of work, findings and remaining tasks.

---
name: bounded-researcher
description: Focused research
maxSteps: 8
---

Structural context, ranked

The repo map enumerates files, extracts symbols and references, builds a weighted graph, then uses PageRank to rank structurally important files before rendering within a token budget.

Reusable procedures, named

A skill is a folder with a SKILL.md. Project and user skills load at start; commands support list, show, install, remove, validate and verify.

Credentials

The routing guide warns that api_key values in settings.json are stored in plaintext. Keep the file private and out of version control.

Model variance

Anthropic-specific features may not exist on other providers, smaller local models can struggle with long tool flows, and output caps vary.

Config boundaries

OpenClaude uses its own ~/.openclaude paths and does not automatically load Claude Code configuration. Migrate intentionally.

OpenAI-compatible API

$ export CLAUDE_CODE_USE_OPENAI=1
$ export OPENAI_API_KEY=sk-your-key
$ export OPENAI_MODEL=gpt-4o
$ openclaude

Ollama

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

Route

Add agentModels and agentRouting to map different agent types to models or endpoints.

Map

Enable REPO_MAP=1 when a repository is large enough that compact structural context improves navigation.

Package

Turn repeated procedures into skills under .openclaude/skills, then validate and verify them before wider use.

Functional checks

• Install path works on a clean machine.
• Provider setup can start a session and stream output.
• File read/write, shell, grep and MCP tools behave as expected.
• Resume, fork and background-session commands match your workflow.
• Repo map and skills load only when intended.

Risk checks

• Credentials are not committed or exposed in logs.
• Agent step limits stop work predictably.
• Tool permissions are visible and reviewable.
• Local versus cloud data paths are understood.
• gRPC is not exposed on all interfaces without authentication.

Primary reading

Licence note

The repository licence says modifications and additions by OpenClaude contributors are offered under the MIT License where legally permissible, while underlying derived code remains subject to Anthropic’s copyright and terms.