Open-source alternatives to Devin

The highest-scoring open-source projects in Coding agents and assistants. They are picks, not exact replacements, so read the weaknesses before you switch.

Open-source alternatives to Devin
Rank	Project	Score
1	opencodeTerminal coding agent with build and plan modesCoding agents and assistants, 212.8k stars, MIT	78 out of 100
2	orcaDesktop app that runs CLI coding agents in parallel git worktreesCoding agents and assistants, 90.1k stars, MIT	75 out of 100
3	oh-my-openagentTerminal coding agent that runs parallel sub-agents and keeps git-based memoryCoding agents and assistants, 70k stars, custom license	75 out of 100
#4	gemini-cliTerminal coding agent that runs Gemini models with file, shell and MCP toolsCoding agents and assistants, 107.3k stars, Apache-2.0	73 out of 100
#5	OpenHandsWeb control center for running coding agents and scheduled automationsCoding agents and assistants, 90.7k stars, MIT	73 out of 100
#6	ArchonYAML workflow engine that runs coding agents in isolated worktreesCoding agents and assistants, 23.7k stars, MIT	73 out of 100
#7	Open SWELangChain coding agent that plans, implements and reviews pull requestsCoding agents and assistants, 11k stars, MIT	73 out of 100
#8	clineCoding agent for VS Code, JetBrains, terminal and desktopCoding agents and assistants, 70.2k stars, Apache-2.0	72 out of 100
#9	openinterpreterTerminal coding agent, a Codex fork tuned for low-cost modelsCoding agents and assistants, 68.6k stars, Apache-2.0	72 out of 100
#10	herdrTerminal multiplexer that keeps coding agents running and shows which are blockedCoding agents and assistants, 43.5k stars, Apache-2.0	72 out of 100

Reviews

178

opencode

Terminal coding agent with build and plan modes

212.8k stars, MIT, last commit Oct 2026

Runs an AI coding agent in the terminal with two built-in agents: build (full access) and plan (read-only, asks before running bash), plus a general subagent for multi-step searches. Installs via a curl script, npm, Homebrew, Scoop, Chocolatey, pacman, mise or Nix, and ships a beta desktop app for macOS, Windows and Linux. For developers who want an open, configurable coding agent.

Strengths

MIT license; installable from npm, Homebrew, Scoop, Chocolatey, pacman, mise and Nix
Plan agent denies file edits and asks before bash, for safe codebase exploration
Desktop app (beta) for macOS, Windows and Linux alongside the terminal UI

Weaknesses

README covers install only; providers, config and server mode are in external docs
No Dockerfile or compose file in the repo
Desktop app is still beta

no GPU

Documentation ↗ Website ↗ Repository on GitHub ↗

275

orca

Desktop app that runs CLI coding agents in parallel git worktrees

90.1k stars, MIT, last commit Oct 2026

Orca is a cross-platform desktop app (macOS, Windows, Linux) that runs several CLI coding agents side by side, each in its own isolated git worktree. It bundles terminals with splits, an editor, an embedded Chromium browser, GitHub and Linear task views, and diff annotation. A mobile companion app and an `orca` CLI let you monitor and script agent workflows.

Strengths

Works with any agent that runs in a terminal, including Claude Code, Codex, OpenCode, Pi
Each agent gets its own git worktree; fan one prompt across several and compare
Agents can run on a remote machine over SSH with port forwarding
MIT license; desktop builds for macOS, Windows, Linux plus Homebrew and AUR packages

Weaknesses

No Docker or self-hosted server install; it is a desktop app
Mobile pairing relay lives in a separate pnpm workspace under cloud/
Collects anonymous usage telemetry; opt-out is documented but not the default
Ships daily and the README says its feature list lags behind

no GPU
Needs git, CLI coding agent of your choice

Documentation ↗ Website ↗ Repository on GitHub ↗

375

oh-my-openagent

Terminal coding agent that runs parallel sub-agents and keeps git-based memory

70k stars, custom license, last commit Oct 2026

OmO ships a single `omo` command, a native binary built on senpi, the project's fork of pi. It takes a prompt and plans, runs and checks the work. Adding `ultrawork` or `mass ulw` makes it fan the job out to many agents, each on a model chosen for that step. Memory is stored as markdown files in a git repository, and `/login` supports Claude, ChatGPT, Kimi and GLM subscriptions.

Strengths

Native binary installs via curl script, bun or npm; no Docker needed
Logs in with Claude, ChatGPT, Kimi and GLM subscriptions
Memory is plain markdown in a git repo, so it can be inspected
`omo setup` migrates keys, MCP servers and skills from the OpenCode edition

Weaknesses

License is SUL-1.0, not a standard OSI license; check terms before commercial use
Install is a curl-pipe-to-bash script
Computer use is documented as experimental
Hardware needs and supported model list are not stated in the README

Models: Claude, ChatGPT, Kimi, GLM

Documentation ↗ Website ↗ Repository on GitHub ↗

#473

gemini-cli

Terminal coding agent that runs Gemini models with file, shell and MCP tools

107.3k stars, Apache-2.0, last commit Oct 2026

Gemini CLI is a TypeScript terminal agent, installed from npm, Homebrew, MacPorts or conda, that sends prompts to Gemini models and can edit files, run shell commands, fetch web pages and ground answers with Google Search. It supports MCP servers, GEMINI.md context files, conversation checkpointing, a headless mode with JSON and stream-JSON output, and a GitHub Action for PR review and issue triage.

Strengths

Free tier with Google login: 60 requests/min, 1,000 requests/day
1M token context window with Gemini 3 models
Headless mode emits plain text, JSON or newline-delimited stream-JSON
Auth via Google login, Gemini API key or Vertex AI

Weaknesses

Only talks to Gemini models; no other providers listed in the README
Needs a Google account, API key or Vertex AI credentials; no local models
Preview and nightly channels may contain regressions
Free-tier quotas and terms are governed by Google, not the project

no GPU
Docker
Needs Node.js, Gemini API, Google account login or Vertex AI
Models: Gemini 3, Gemini 2.5 Flash

Documentation ↗ Website ↗ Repository on GitHub ↗

#573

OpenHands

Web control center for running coding agents and scheduled automations

90.7k stars, MIT, last commit Oct 2026

Agent Canvas is a web frontend that starts and manages conversations with coding agents. It runs the OpenHands agent by default and can drive Claude Code, Codex, Gemini CLI, Pi, OpenCode, or any ACP-compatible agent. It connects to one or more Agent Server backends (local, Docker, VM, or OpenHands Cloud) and pairs with an Automation Server for scheduled and webhook-triggered runs.

Strengths

Switches between local, remote and cloud Agent Server backends from one UI
Works with third-party agents through ACP, not only OpenHands
Option to run each conversation in its own Docker container
Automations can run on a schedule or on webhook events

Weaknesses

Marked beta in the README
Non-sandboxed install gives the agent full filesystem access
Needs Node.js 24+ and uv for non-Docker installs
Automation and agent server live in separate repositories

Docker
Needs Node.js 24+, uv, Docker (optional)
Models: Any LLM (via LLM profiles)
port 8000

Documentation ↗ Repository on GitHub ↗

#673

Archon

YAML workflow engine that runs coding agents in isolated worktrees

23.7k stars, MIT, last commit Oct 2026

Defines development processes (plan, implement, validate, review, PR) as YAML workflows and runs them through Claude Code, Codex or Pi, each run in its own git worktree. Deterministic nodes mix with AI nodes and human approval gates; runs start from the CLI, a web console, Slack, Telegram, Discord or GitHub webhooks, with state in SQLite or PostgreSQL. For teams standardizing how agents ship code.

Strengths

Every run isolated in a git worktree; parallel fixes without conflicts
Bundled sdlc pack: ship, triage, investigate, plan, deliver, review, validate, upkeep
Adapters for web, CLI, Slack, Telegram, Discord and GitHub webhooks
Telemetry documented field by field; DO_NOT_TRACK=1 or CI=true disables it

Weaknesses

Requires Claude Code (or Codex, Pi) installed separately; binaries need CLAUDE_BIN_PATH
Anonymous telemetry is on by default
x64 quick-install binaries require AVX2; older CPUs must build from source
Workflows from v0.11.1 and earlier no longer ship and must be copied manually

no GPU
Docker + Compose
Needs Bun, Claude Code (or Codex or Pi), GitHub CLI, SQLite or PostgreSQL
Models: Claude Code, Codex, Pi

Documentation ↗ Repository on GitHub ↗

#773

Open SWE

LangChain coding agent that plans, implements and reviews pull requests

11k stars, MIT, last commit Oct 2026

LangGraph-based agent that investigates a repository, implements changes in a per-thread Linux sandbox, validates them and opens a pull request, then reviews PRs and watches CI with /baby-sit. Work starts from a dashboard, GitHub issues or PR comments, Slack or Linear. Deploys into your infrastructure with a backend, dashboard, GitHub and Slack apps, for teams building an internal coding-agent service.

Strengths

Covers build, review, investigate and operate flows, with subagents for parallel work
Durable execution and thread state via LangGraph; sandboxes persist per thread
Configurable models, reasoning effort, skills, MCP integrations and sandbox providers
Push approvals gate detected git pushes; PR chat excludes mutation tools

Weaknesses

Production standalone Agent Server deployments require a license key
Maintainers are not accepting issues or contributions; breaking changes expected
LangSmith is the default sandbox and tracing provider; alternatives need configuration
CLI mode executes commands locally as you, with no sandbox isolation

no GPU
Docker + Compose
Needs LangSmith (default sandbox and tracing), GitHub App, Slack app (optional), model provider credentials
Models: configurable LLM providers

Repository on GitHub ↗

#872

cline

Coding agent for VS Code, JetBrains, terminal and desktop

70.2k stars, Apache-2.0, last commit Oct 2026

Cline is a coding agent that reads a project, edits files across it, and runs shell commands while watching the output. It ships as a VS Code extension, JetBrains plugin, CLI with headless mode, macOS/Windows desktop app, and a Node.js SDK. Edits and commands need approval by default, with Plan and Act modes, checkpoints, and optional auto-approve.

Strengths

Same agent engine across VS Code, JetBrains, CLI, desktop app and SDK
Works with Anthropic, OpenAI, Gemini, Bedrock, Vertex, OpenRouter, Ollama and LM Studio
Headless CLI accepts piped input and emits JSON for CI/CD scripts
Supports MCP servers, SDK plugins, cron-scheduled agents and multi-agent teams

Weaknesses

JetBrains plugin source is not open-sourced
VS Code extension code is still migrating to the new layout
Hosted model providers need your own API keys; local models via Ollama or LM Studio
Diff review and checkpoints are described only for VS Code and JetBrains

Needs Node.js, VS Code or JetBrains IDE (for extensions), LLM provider API key or local model server
Models: Anthropic, OpenAI, Google Gemini, OpenRouter, Vercel AI Gateway

Documentation ↗ Website ↗ Repository on GitHub ↗

#972

openinterpreter

Terminal coding agent, a Codex fork tuned for low-cost models

68.6k stars, Apache-2.0, last commit Oct 2026

Open Interpreter is a Rust fork of OpenAI's Codex that runs as a terminal coding agent. It can emulate other agent harnesses (claude-code, kimi-code, qwen-code, swe-agent and others) via /harness, so cheaper models are driven the way their providers recommend. It also runs as an ACP agent for editors and accepts the Codex exec protocol.

Strengths

Switchable harness emulation: native, claude-code, kimi-code, qwen-code, swe-agent, minimal and more
Runs under native sandboxing on macOS, Linux and Windows
Works as an ACP agent via `interpreter acp`; Codex SDK users can swap the binary
Reads shared AGENTS.md and .agents/skills; supports MCP, hooks and permissions

Weaknesses

Rewrite of the original Python project, which is now only a community fork
Built-in computer use relies on external tools: agent-browser and trycua
README gives no RAM, GPU or supported-model minimums
Install is a curl or PowerShell pipe-to-shell script; no Docker image

Needs agent-browser, trycua
Models: OpenAI-compatible Chat Completions providers, Kimi K3, DeepSeek, Z.AI GLM

Documentation ↗ Website ↗ Repository on GitHub ↗

#1072

herdr

Terminal multiplexer that keeps coding agents running and shows which are blocked

43.5k stars, Apache-2.0, last commit Oct 2026

herdr is a Rust terminal multiplexer for running several coding agents such as Claude Code, Codex, Cursor and opencode. A background server keeps panes alive when the client detaches or SSH drops, and each pane is marked working, blocked or idle. Agents can also drive it through a CLI and socket API, and saved SSH machines appear in the same window.

Strengths

Panes keep running in a background server after detach or SSH disconnect
Per-pane working, blocked or idle status across local and saved SSH machines
Agents can spawn panes and prompt each other via CLI and socket API
Single Rust binary; installs via curl script, Homebrew, mise or PowerShell

Weaknesses

After a server or machine restart, processes are lost; only layout and supported agent sessions resume
Resume works only for supported agents; the README does not list which
Windows support is described as beta in the docs link
Install is a curl-pipe-to-shell script by default

no GPU

Documentation ↗ Website ↗ Repository on GitHub ↗