Best Open Source Claude Code Alternatives

Key Features

• 75+ providers plus local models; ChatGPT Plus/Pro subscription via browser sign-in
• AGENTS.md for project instructions; rename CLAUDE.md to AGENTS.md when migrating; .claude/skills as a documented fallback path
• Session resume, fork, export/import JSON, and share URL
• TUI and desktop beta; IDE/ACP server mode
• Subagents, MCP, plugins/hooks with file-edited, tool-execution, and session-lifecycle events; skills; custom commands via configuration

Pros

• MIT repository with many install routes: curl, npm-family packages, Homebrew, Scoop, and Nix
• AGENTS.md is the instruction file; rename CLAUDE.md to AGENTS.md to migrate project instructions without rebuilding the content
• Any-provider freedom lets you swap Claude for a local model or a ChatGPT plan without changing your workflow
• Per-tool allow/ask/deny permissions let you gate individual tools without blocking everything else

Cons

• Claude Pro/Max subscription reuse is explicitly not supported; the first-party provider page says Anthropic prohibits the route; oh-my-pi or Goose cover that path instead
• MCP servers from .mcp.json need manual translation into OpenCode's config format and re-approval; no direct importer is documented
• Hosted provider services are separate from the core MIT repository

License & Hosting

• License: MIT, fully open; modify and self-host freely
• Where it runs: local-first; no server required from anyone else; available as a TUI, desktop beta, and IDE/ACP server
• Setup: install via curl, npm-family packages, Homebrew, Scoop, or Nix; source: github.com/anomalyco/opencode
• Costs: provider API key required for inference; Claude Pro/Max subscription sign-in is explicitly not supported

Best For

I'd start here for most people who want provider freedom in a terminal coding tool. Rename CLAUDE.md to AGENTS.md to bring project instructions across; .claude/skills compatibility means your skills move without rewriting. Skip OpenCode if you need Claude Pro/Max subscription inference to keep working; that route is explicitly not supported, so oh-my-pi or Goose are the right places to look instead.

Key Features

• Auto-discovers CLAUDE.md, AGENTS.md, project .mcp.json, rules, skills, commands, hooks, and tools/settings on first run
• Anthropic OAuth and subscription login; OpenAI Codex/ChatGPT subscription login
• 60+ providers including Ollama; subscription or API key
• omp commit Git command; subagents; RPC mode; no-session for headless and automation
• ACP IDE integration

Pros

• MIT repository; the most direct Claude configuration compatibility of any tool on this shortlist
• CLAUDE.md, MCP servers, skills, and hooks load without manual conversion; you re-approve rather than rebuild
• Active Claude subscription keeps working without a separate API key

Cons

• Own-session resume only; Claude conversation history import is not documented
• No managed hosted tier is documented; local client with BYOK or subscription auth

License & Hosting

• License: MIT, client code fully open; compute is not included
• Where it runs: local-first; TUI, RPC, and no-session headless modes run on your machine; no documented managed hosting
• Setup: install via curl, Homebrew, Bun, Nix, or Windows package routes; source: github.com/can1357/oh-my-pi
• Costs: provider API key, local endpoint, or eligible Claude/ChatGPT subscription required for inference

Best For

I'd start here for anyone who has an active Claude Pro or Max subscription and significant Claude configuration already on disk. The rules, skills, MCP servers, and hooks move without manual conversion; you're re-approving what already exists, not rebuilding it. The deciding detail is whether those configuration files are the asset you want to preserve. RPC mode and no-session headless paths are available if you need to drive it from automation without an interactive session. Skip oh-my-pi if you have no existing Claude config and want the widest model agnosticism; OpenCode is the better starting point in that case.

Key Features

• /resume-claude imports Claude Code transcripts; skips and counts unsupported outputs
• AGENTS.md auto-loads; add CLAUDE.md to instructions or copy/rename it to AGENTS.md
• .claude/skills loads with compatibility enabled
• ChatGPT Plus/Pro subscription documented; multi-provider and local models including Ollama and LM Studio
• VS Code, JetBrains, and CLI clients in the MIT core; Slack via the hosted Cloud Agent; subagents/custom modes; MCP; plugins/hooks
• Session resume, export, and fork; headless CLI automation

Pros

• /resume-claude is one of three documented paths for moving Claude session transcripts (alongside Codex's /import and Cline's SDK importer)
• Multi-surface: IDE and CLI share the same configuration, so switching contexts during a task does not mean switching tools

Cons

• Claude consumer subscription sign-in is not available; the documentation requires an Anthropic API key.
• Git commit-message generation is documented for the VS Code extension only, not the CLI or TUI
• MIT label covers the repository code; hosted/team/cloud services have separate terms not included in that license

License & Hosting

• License: MIT (core repository); hosted, team, and cloud services have separate terms not covered by the repo license
• Where it runs: CLI and IDE extensions are local-first; Slack and cloud surfaces run on the hosted Cloud Agent
• Setup: global npm install for CLI; VS Code and JetBrains IDE extensions from their respective marketplaces; source: github.com/Kilo-Org/kilocode
• Costs: Anthropic API key required; consumer Claude subscriptions are not documented as supported; hosted and cloud tiers carry separate costs or terms

Best For

I'd choose Kilo if you want the /resume-claude import path and plan to move between IDE and CLI. The combination of transcript import and multi-surface availability makes it the clearest fit for mixed IDE/terminal workflows. Session resume, export, and fork are also available within Kilo's own sessions, so you do not lose continuity after the initial transcript import. Skip it if you need Claude subscription sign-in (not documented), or if you expect the MIT label to extend to the hosted services.

Key Features

• ACP reuse of an existing Claude Code subscription; ChatGPT Plus/Pro via ACP, with no new API account required for those paths
• AGENTS.md loads by default; CLAUDE.md configurable via CONTEXT_FILE_NAMES
• .claude/skills and .claude/agents discovered automatically
• Subagents (sequential and parallel); MCP; lifecycle hooks under .agents/plugins; recipes; headless JSON modes
• Apache-2.0, maintained under AAIF

Pros

• If you have a Claude Code or ChatGPT Plus/Pro subscription, Goose can reuse it without separately metered provider API credentials for that path
• .claude/skills and .claude/agents discovery means those artifacts travel without manual relocation
• Recipes provide reusable multi-step workflows without a custom plugin, which can simplify routine automation

Cons

• ACP sessions do not support fork or resume; native Goose sessions do resume, so the limitation only applies to ACP-authenticated work
• CLAUDE.md is not the default instruction file; it requires adding it explicitly via CONTEXT_FILE_NAMES
• Direct project .mcp.json import not documented; recreate MCP configuration as Goose extensions
• Built-in local Git commit/PR workflow is not documented

License & Hosting

• License: Apache-2.0, maintained under AAIF
• Where it runs: local-first; no server required for the core installation; local desktop, CLI, and API surfaces; no native IDE client or managed hosted tier documented
• Setup: shell installer with platform-specific paths; source: github.com/aaif-goose/goose
• Costs: provider API key or Claude/ChatGPT Plus/Pro subscription required; ACP subscription reuse is documented

Best For

I'd start with Goose specifically when the requirement is reusing an existing Claude Code or ChatGPT Plus/Pro subscription without opening a new API account. Keep the ACP session-resume limitation in mind before relying on it for long-running work; if session continuation matters, native Goose sessions are the right path, not ACP.

Key Features

• Apache-2.0 (CLI, SDK, App Server); IDE extension and Codex cloud are not open source
• ChatGPT sign-in or plan access; local Ollama and LM Studio via --oss
• /import from Claude Code: supported setup/project material, skills, hooks, MCP, subagents, and up to 50 chats from last 30 days
• AGENTS.md native; configurable fallback filenames can include CLAUDE.md
• Subagents, MCP, hooks, skills; codex exec headless; JSON output modes

Pros

• First-party /import is one of three structured documented paths for Claude Code handoff, alongside Kilo Code's /resume-claude and Cline's SDK importer
• ChatGPT plan reuse means no separate API account for existing subscribers
• --oss brings Ollama and LM Studio into the same terminal tool without switching clients

Cons

• IDE extension and Codex cloud are explicitly not Apache-2.0; open-source scope is CLI, SDK, and App Server only
• Claude subscription auth not documented; the documented subscription path is ChatGPT
• /import has restrictions: unavailable during active tasks, remote sessions, or daemon attachment; capped at 50 chats from the last 30 days
• Desktop app licensing was not independently characterized beyond the official component table

License & Hosting

• License: Apache-2.0 (CLI, SDK, and App Server); the IDE extension and Codex cloud are explicitly not open source
• Where it runs: local-first; no server required for the open-source components; CLI on macOS, Linux, and Windows; Codex cloud is a separate hosted surface
• Setup: npm or standard install; source: github.com/openai/codex
• Costs: OpenAI API key or ChatGPT plan required; Codex cloud carries separate pricing

Best For

I'd start with Codex if you're already on a ChatGPT plan and want the most structured import of your recent Claude Code sessions. The 50-chat, 30-day cap means recent context moves but older history stays behind. Plan for the authentication review and tool permission checks that follow any import before enabling shell access.

Sessions are branchable trees: resume, branch, compact, or export to HTML or a gist. The print/JSON and RPC modes suit scripted pipelines and embedding. Agent Skills and extension events and commands let you add behaviour without forking. Context loads from AGENTS.override, AGENTS, and CLAUDE filenames in that priority order; CLAUDE.md project instructions are recognized. MCP configuration does not carry: Pi has no built-in MCP by design.

Claude Pro or Max subscription works through Anthropic OAuth. ChatGPT Plus/Pro subscription is also documented. A local llama.cpp route handles offline or cost-sensitive runs. Pi runs with the invoking user's filesystem, process, network, and credential authority; use a Docker container when a hard isolation boundary matters.

Key Features

• Four default tools: read, write, edit, bash (eight built-ins available across supported platforms: adds grep, find, ls, and powershell on Windows)
• Branchable tree sessions: resume, branch, compact, HTML or gist export
• Interactive TUI, print/JSON, and RPC modes; Agent Skills and extension events
• AGENTS.override/AGENTS/CLAUDE context filenames; CLAUDE.md recognized
• Claude Pro/Max (Anthropic OAuth/subscription login), ChatGPT Plus/Pro, and local llama.cpp
• MIT; npm, pnpm, Bun, official installer, or release binaries

Pros

• MIT repository; minimal footprint with no framework or permission layer to configure
• Branchable sessions suit both long-running interactive work and scripted one-shot runs
• RPC and SDK surfaces make Pi a composable core for embedding in a larger program

Cons

• No built-in MCP, subagents, permission popups, plan mode, built-in to-dos, or background bash; all are deliberate core omissions
• Runs with invoking-user authority by design; use a Docker container for a hard permission boundary
• Claude session history and MCP configuration do not carry across automatically

License & Hosting

• License: MIT (repository and coding-agent package)
• Where it runs: local-first; runs on your machine with invoking-user permissions; Docker container setup documented for isolation
• Setup: install via npm, pnpm, Bun, the official installer, or release binaries; source: github.com/earendil-works/pi
• Costs: provider API key required; no hosted tier documented

Best For

I'd start with Pi when you want the smallest possible open-source core to extend yourself. The RPC and SDK surfaces suit embedding and scripted pipelines. Skip Pi if you need built-in MCP, subagents, or per-action permission prompts: those are deliberate omissions. Use oh-my-pi if you want those features. Run Pi in a Docker container before granting write and shell access to real work.

Key Features

• VS Code, JetBrains, CLI, desktop app, SDK, and Kanban; npm install -g cline or a runtime-free packaged binary
• Claude Code provider authenticates with Claude Max or Pro; Anthropic, OpenAI, Google, Ollama, LM Studio, and compatible local endpoints
• Auto-discovers .claude/skills; MCP servers and hooks under .cline/hooks; experimental read-only subagents (cannot edit files, browse, or use MCP); SDK plugins
• SDK session importer reads Claude Code history from ~/.claude/projects; Kanban companion documents Commit and Open PR

Pros

• It reuses the Claude subscription you already pay for through the Claude Code provider; a Claude Pro or Max plan drives it without a separate API key.
• Your .claude/skills come across automatically and your Claude Code sessions import through the SDK, so two of the pieces that usually need a manual rebuild move for you.
• One tool spans the editor and the terminal, so a team split between VS Code and the shell shares one set of rules, servers, and skills instead of two.

Cons

• Automatic CLAUDE.md loading is not documented; you point Cline at .clinerules or .cline/rules, or move that file's content in by hand.
• The CLI defaults to --auto-approve true; pass --auto-approve false if you want approval prompts before each tool call. Know which mode you are in before granting shell access.
• Plugins are documented for the CLI, VS Code, JetBrains, Kanban, and SDK/custom hosts. The JetBrains plugin is not open source.

License & Hosting

• License: Apache-2.0
• Where it runs: local-first; installs and runs on your machine via global npm package or packaged binary; no separate runtime required
• Setup: npm install globally or download a packaged binary
• Costs: model inference billed by your provider or plan; vendor-hosted services are separate from the open-source client

Best For

I'd choose Cline when moving from an IDE-heavy Claude Code workflow and you want to keep a Claude Max/Pro subscription working. The session import and .claude/skills discovery cut the rebuild cost: skills come across automatically and Cline can discover and import selected Claude Code sessions from ~/.claude/projects. Skip it if automatic CLAUDE.md loading is a hard requirement; that gap is documented and remains. For that case, oh-my-pi reads CLAUDE.md directly; Kilo Code requires adding it to instructions or renaming it to AGENTS.md.

Key Features

• Automatic commits, /undo, and /diff after every change
• Multi-file edits; /run and /test for inline shell and test commands
• --message and --yes for scripted one-shot tasks
• Anthropic, OpenAI, Gemini, Ollama, and other cloud/local providers
• --restore-chat-history for own session; /save reconstructs file context
• Apache-2.0; terminal-first with an optional browser GUI flag

Pros

• The commit-and-undo loop is the simplest documented Git workflow on this shortlist; each accepted change lands in history immediately
• Broad API-key cloud provider support plus local providers, with no subscription dependency
• Deliberate narrowness means less to configure and fewer surprises in a team setup

Cons

• The Aider docs do not describe MCP client support, subagents, or lifecycle hooks
• ChatGPT subscription explicitly does not include API access; the Aider docs say so
• Claude subscription login not documented; API key is the documented path
• Default-branch commit dated 2026-05-22 and latest release v0.86.0 from 2025-08-09; materially older maintenance dates than the rest of this shortlist

License & Hosting

• License: Apache-2.0, client code fully open; model compute is not included
• Where it runs: local-first; terminal client and optional browser GUI both install and run locally
• Setup: Python 3.8–3.13 required; install via pip, uv, pipx, or script; source: github.com/Aider-AI/aider
• Costs: hosted model API key billed separately by your provider; local models run on your own hardware; neither is included in the client license

Best For

I'd choose Aider when every change needs to be in Git and the commit-and-undo loop is the central workflow. It is deliberately simpler than the other tools here, and that simplicity is the point. Check the maintenance dates before building a team setup around it; the most recent activity is older than the rest of this shortlist, and that is worth knowing before you commit to it as a team standard.

Key Features

• MIT core; CLI, headless, web/server, SDK, cloud, and ACP IDE surfaces (Zed, JetBrains, community VS Code extension)
• ChatGPT Plus/Pro subscription; provider-agnostic SDK with many LLMs and local model routes
• Subagents via SDK: TaskToolSet for sequential blocking delegation; parallel delegation also available when Agent tool_concurrency_limit is greater than 1
• Claude-compatible hooks at .openhands/hooks.json; plugins can bundle hooks, MCP, agents, skills, and commands
• AGENTS.md recommended; CLAUDE.md supported as optional permanent context
• Cloud/GitHub PR integration; headless mode with JSONL output

Pros

• The SDK documents sequential delegation via TaskToolSet and parallel delegation by configuring Agent tool_concurrency_limit greater than 1, so multi-step work can be split without a custom orchestrator
• Hook format is Claude-compatible, so existing hook logic can often move to .openhands/hooks.json with minimal rewriting
• GitHub Cloud integration handles PR workflows, code review, and issue resolution without a separate automation script

Cons

• Claude subscription login not documented for the standard LLM path; the subscription guide says OpenAI is the current vendor
• Headless mode always approves actions; know this boundary before putting it in an automated pipeline
• Direct project .mcp.json import not documented; recreate MCP configuration or package it inside a reviewed plugin
• The OpenHands/enterprise repository is PolyForm source-available, not MIT; keep it separate from core feature claims

License & Hosting

• License: MIT (core repository); OpenHands/enterprise is a separate PolyForm source-available repository
• Where it runs: local, self-hosted, or OpenHands Cloud; use the main repository; the deprecated standalone OpenHands-CLI is no longer actively maintained
• Setup: install via uv on Python 3.12+, a binary installer, or Docker; source: github.com/OpenHands/OpenHands
• Costs: provider API key or Claude subscription required; OpenHands Cloud carries separate pricing

Best For

I'd choose OpenHands when the goal is a server-grade, SDK-driven pipeline or a cloud PR workflow rather than an interactive terminal session. The headless always-approve behavior is the main constraint to understand before scripting it. Use the current main repository, not the deprecated standalone OpenHands-CLI; and attribute MIT to the core, not to any enterprise features.

Key Features

• Single 6.17 MiB native binary; install with one curl command, no runtime, no container
• AGENTS.md project instructions, global and per-directory, with the narrowest scope winning
• Skills, MCP servers, and subagents; sessions with save, resume, recover, and compact
• fx ask and fx ask --json for noninteractive and scripted use
• fx acp serves the Agent Client Protocol for editors; the libfx npm package embeds the agent core in Node or WebAssembly hosts
• Auto and ask permission modes; file changes, commands, and external paths go through the permission policy

Pros

• A vendor-reported 6.17 MiB native binary with an explicit libfx embedding surface for Node and WebAssembly hosts
• Shell-shaped output that preserves scroll history composes naturally in pipelines and CI jobs
• Despite the small binary, fx supports skills, MCP servers, and subagents

Cons

• Experimental by the vendor's own label; the README says use at your own risk
• Vercel AI Gateway is the default documented provider, with ChatGPT and Grok subscription sign-in; compatible loopback endpoints and fully hermetic local inference are also documented
• Claude subscription login is not documented, and neither is CLAUDE.md; instructions live in AGENTS.md
• No documented hooks or Claude session import; migration means copying instructions into AGENTS.md and reviewing the existing Claude-compatible project .mcp.json

License & Hosting

• License: Apache-2.0
• Where it runs: local-first; single native binary for macOS and Linux (x86_64/arm64); ACP server and libfx embedding modes run locally; hermetic local inference also documented
• Setup: install by script or build from source with Zig; source: github.com/vercel-labs/fx
• Costs: Vercel AI Gateway (default), ChatGPT or Grok subscription sign-in, or compatible loopback endpoint; each carries its own ongoing cost or local resource requirement

Best For

I'd pick fx to script an agent inside a pipeline, a CI sandbox, or another program, where a 6 MiB binary with a JSON output mode is exactly the right shape and none of the bigger tools fit. I would not choose fx as a daily driver replacement for Claude Code: it is experimental by its own README, and Claude subscription login is not documented.