Best Open Source Claude Code Alternatives

Claude Code's public repository license says all rights reserved. That is a documented constraint with practical consequences. The published repository does not provide an open-source, forkable, self-buildable implementation. When terms change or the service has an outage, there is no fallback you can compile yourself.

Model inference routes through Anthropic's service or documented Claude gateways, not local models and not arbitrary third-party providers. GPT-4o, Gemini, and local Ollama runs are not supported paths. The official documentation covers Claude's own deployment surfaces, including Bedrock and Vertex routes, but those are still Claude models delivered differently. For work that depends on a specific non-Claude model, Claude Code is not the tool. OpenCode covers this directly: 75+ providers plus local models under an MIT license.

The subscription structure can make the economics awkward. Pro runs at $20 a month and Max starts at $100, with metered API usage as a separate route. For consistent high-volume work that may be reasonable. For occasional terminal sessions, you may be paying for capacity you do not use.

Configuration is portable in the sense that CLAUDE.md, .mcp.json, skills, hooks, and settings are plain files you can copy. What is not documented is a single-command neutral export that packages your full configuration plus recent sessions for handoff to another tool. Claude Code stores local session transcripts and documents /export; no single command bundles all configuration and sessions for migration. For anyone with a substantial Claude configuration on disk, oh-my-pi is the most direct answer: it auto-discovers those files and loads them on first run.

If any of these constraints are the reason you're looking, the nine MIT- or Apache-2.0-licensed tools in this guide address different combinations of them. The deciding question is whether you want provider freedom, configuration portability, or both.

The deciding question is what you most need to preserve from your current setup: your Claude subscription, your configuration files, or your session history.

If you have an active Claude Pro or Max subscription and significant configuration on disk, I'd choose oh-my-pi. It auto-discovers CLAUDE.md, MCP servers, skills, hooks, and rules on first run. You re-approve what already exists rather than converting it. If you also want Goose as an option, it documents ACP reuse of an existing Claude Code subscription, but ACP sessions do not support fork/resume, so check that limitation before committing to it for long-running work.

If you have no existing Claude config and want maximum provider freedom, OpenCode is the common-case pick. It covers 75+ providers plus local models, treats CLAUDE.md as a fallback when AGENTS.md is absent, and has a GitHub PR workflow built in. Most people starting from zero should start here.

If you're on a ChatGPT plan and want to bring recent Claude Code session history along, Codex has /import for that. It covers the last 50 chats from 30 days, maps skills, hooks, and MCP configuration, and runs on the same ChatGPT subscription.

If you need the Claude transcript import path and also want to move between IDE and terminal, Kilo Code covers both. The /resume-claude command imports Claude transcripts in the CLI, and the VS Code, JetBrains, and CLI clients read the same shared configuration.

If you want one agent across the editor and terminal that reuses your Claude subscription and carries your skills, Cline fits. It signs in with a Claude Max or Pro plan through the Claude Code provider, auto-discovers .claude/skills, and imports your Claude Code sessions via its SDK. Its one gap is CLAUDE.md, which it does not load automatically, so if that file matters most, oh-my-pi or Kilo Code read it directly instead.

For a deliberately minimal Git workflow where each change commits under Aider's default auto-commit setting and /undo reverts the last Aider commit, I'd pick Aider. Check its maintenance dates before building a team standard around it.

For server-grade pipelines, SDK-driven subagent delegation, or GitHub PR workflows, OpenHands is the strongest fit. The headless always-approve mode and the enterprise boundary from the MIT core are the two things to know going in.

If you want a very small agent to script in a pipeline or embed in another program, fx is a 6.17 MiB native binary with fx ask --json for structured output and an ACP and WebAssembly embedding surface. Its own README labels it experimental, and the model and authentication pages do not explain the local routes its docs index advertises, so I'd keep it for scripting and embedding rather than daily terminal work.

I'd group the migration effort into three tiers. Direct/automatic: oh-my-pi discovers CLAUDE.md, .mcp.json, rules, skills, commands, hooks, and tool settings on first run. Partial/configured with documented paths: OpenCode, Kilo Code, Goose, Codex, Cline, and fx each document compatibility for some files but need manual work for others. Manual rebuild: for Aider and OpenHands no automatic importer is documented; I'd supply instruction files explicitly and recreate MCP, hooks, and skills by hand.

Before you start: keep a clean copy of the repository config. Remove API keys, machine-specific paths, and tokens before copying anything. Inspect every instruction file, MCP definition, skill, and hook as executable configuration before enabling it.

What moves by target:

oh-my-pi discovers CLAUDE.md, AGENTS.md, project .mcp.json, rules, skills, commands, hooks, and tool settings on first run. Revalidate provider authentication and approval modes.

OpenCode treats CLAUDE.md as a fallback when AGENTS.md is absent and recognizes .claude/skills. Translate MCP servers into OpenCode's config format and re-approve one at a time; rebuild Claude hooks as OpenCode plugins.

Kilo Code recognizes CLAUDE.md and AGENTS.md and loads .claude/skills with compatibility enabled. Use /resume-claude for session transcripts; it skips and reports unsupported outputs. Translate MCP servers into Kilo's format and approve each one.

Goose loads AGENTS.md by default. Add CLAUDE.md via CONTEXT_FILE_NAMES. Goose discovers .claude/skills and .claude/agents. No native import of MCP servers or Claude hooks; recreate them as Goose extensions and Open Plugins.

Codex /import covers supported Claude Code setup/project material, skills, hooks, MCP config, subagents, and up to 50 chats from 30 days. The import is unavailable during active tasks, remote sessions, or daemon attachment; review tool restrictions and hook semantics afterward.

Cline auto-discovers .claude/skills, reuses a Claude subscription through its Claude Code provider, and imports Claude Code sessions from ~/.claude/projects via its SDK. CLAUDE.md is not auto-loaded, so move it into .clinerules by hand and recreate MCP servers and hooks under .cline/.

No automatic importer is documented for Aider. Supply a reviewed instruction or conventions file with --read. --restore-chat-history restores Aider's own previous sessions; Claude session history is not documented as importable. No migration path for MCP, hooks, skills, or subagents is documented.

OpenHands supports CLAUDE.md as optional permanent context and recommends AGENTS.md. Hook format is Claude-compatible, but use .openhands/hooks.json and review path/event differences. No direct .mcp.json importer; recreate or bundle in a plugin.

fx reads AGENTS.md natively; move CLAUDE.md content there. It accepts a Claude-compatible .mcp.json and discovers .claude/skills; hooks and Claude session import are not documented.

Staged move: I'd start in read-only or ask mode, add one MCP server at a time and re-approve it, then run a disposable repository task before enabling write and shell access. Keep your Claude Code setup available until the replacement is reliable.