Claude Code alternatives for agents with local models

OpenCode suits replacing Claude Code repository work with a model-flexible terminal or IDE agent. The development agent edits files, while a planning agent denies edits by default. Automatic language-server context and parallel sessions support project exploration and changes with local or hosted models.

Cline suits people who would rather work from an editor panel than a terminal, though it also has a CLI and a desktop app. It pairs with models on your machine through Ollama or LM Studio, or with Claude, GPT, Gemini and OpenAI-compatible endpoints.

Plan mode proposes an approach before Act mode edits, and you approve edits and commands one by one or let them run. Diffs and checkpoints let you review and undo changes, and the CLI runs without interaction in scripts and CI.

Qwen Code suits Claude Code users who need agent teams, planning and reusable project skills. Its terminal and editor interfaces support local or hosted models, MCP tools and isolated worktrees. Headless operation and SDKs cover scripted coding tasks.

Codewhale covers Claude Code repository edits with planning, permissions and recovery controls. Review file diffs and command results, resume sessions in the terminal or local browser, and use Fleet to divide work among agents. It supports existing Claude Code plugins with a different model backend.

Codex CLI is another terminal agent for exploring a repository, editing files and running commands. Its review mode checks changes without editing them, and noninteractive tasks fit scripts and CI; local execution does not by itself mean offline inference.

Kilo Code covers repository edits, planning and debugging from an IDE or terminal. It supports local models, model switching during a task and parallel work in isolated worktrees.

OpenHands covers delegated and recurring Claude Code engineering tasks. Agents apply changes, investigate failures, run checks and prepare pull requests, with events or schedules starting work. Run its own agent on your infrastructure; using Claude Code as its backend would retain Claude Code.

Pi Desktop brings coding agents into a desktop project workspace. Review a plan before execution, follow commands and tests, and inspect diffs across persistent sessions with the model provider you choose.

CodePilot provides a desktop workspace for coding with file inspection, Git review and separate agent sessions. Its built-in agent works without a separate coding CLI; choosing its Claude Code or Codex mode requires the corresponding CLI and credentials.

Crush is a terminal coding assistant for people who move between model backends. It discovers models from Ollama, llama.cpp, LM Studio and LiteLLM automatically, connects to Anthropic, OpenAI, Gemini and OpenRouter, and accepts any OpenAI- or Anthropic-compatible API.

You can change models during a session without losing its context and keep separate sessions in one project. Language server integrations add code context, and MCP connects other tools.

Aider fits developers who live in a terminal and keep every project in Git. It maps the codebase so the model has context in larger projects, and it commits each change with a message, so you review diffs and undo edits with the Git tools you already use.

It works with local LLMs or cloud models from Anthropic, DeepSeek, OpenAI and others, runs your linters and tests after a change and fixes what they report. Comments in your editor can serve as requests.

Pi suits adapting Claude Code-style terminal work around your own extensions and project instructions. Switch models during a session, branch earlier conversations and package reusable skills or prompts. Planning and subagents require extensions, and Pi inherits the process access it runs with.

gptme edits project files, executes shell commands and Python, and runs tests while retaining session state. ACP connections bring the same agent into compatible editors, with plugins and Markdown skills for recurring work.

jcode is a terminal coding assistant with a choice of local endpoints or hosted providers. Its memory retrieves details from earlier sessions and checks them for stale or conflicting information, helping carry project decisions into later work.

WhipCode suits coding jobs that need recursive delegation and more context than one conversation can hold. Follow the task tree and agent messages while work continues, then return to the same session through its persistent daemon.

Prime Agent suits coding tasks that continue between terminal sessions. It edits project files, keeps a Python environment alive and coordinates subagents; you can disconnect and later reattach to background work.

Goose suits Claude Code users whose tasks mix coding with research, writing or data work. Its desktop app and CLI use local or hosted providers, while MCP extensions connect additional tools. Using Claude Code through an ACP connection would retain it as the backend.

fx covers terminal coding with a shell-style conversation, visible tool activity and preserved scrollback. Skills, MCP tools and subagents extend the agent, while a chosen model endpoint handles inference.

Fast Agent gives terminal coding sessions access to shell commands and MCP tools, with local llama.cpp or other model providers. Its agent workflows can divide a task between specialists and ask an evaluator for revisions.

Pydantic DeepAgents reads and edits project files and runs shell commands from a terminal assistant. It can fork a task into isolated branches so you can compare approaches and review their diffs before continuing.

Stakpak covers the infrastructure side of Claude Code work: generating deployment configuration, debugging Kubernetes and managing CI. It can run ongoing checks and retains incident context, with action logs and backups for file changes.