Best AI coding agents for Open Source (2026)
For open source, DeepSeek Harness (dsh) is our pick: The license decides this. Teams that want to read, fork or self-host the agent itself: the license, which model providers it can call, and what a hosted option costs if you would rather not run it yourself. Below is the full ranking and the tradeoffs, or read how we score.
If you sign up through links on this page, vsref may earn a commission; programs exist on both sides of most comparisons, and commissions never change verdicts. How we make money
Reviewed by vsref Editorialfacts verified Oct 3, 2026Methodology →
DeepSeek Harness (dsh)WINNER
What matters for open source
Weight ×5 = decisive, ×1 = relevant| Fact | DeepSeek Harness (dsh) | Freebuff | Gemini CLI | Goose | Grok Build |
|---|---|---|---|---|---|
| License×5 | MITOct 2 | Apache-2.0Oct 2 | Apache-2.0Oct 2 | Apache-2.0Oct 2 | Apache-2.0Oct 2 |
| Supported model providers×4 | n/a | Curated hosted catalogOct 2 | Google only: Gemini Code AssistOct 2 | 45+ providers incl. Anthropic, OpenAI, Gemini +20 moreOct 2 | xAI (Grok models); custom models via base URL + API keyOct 2 |
| Runs local models×3 | n/a | n/a | n/a | ✓ YesOct 2 | ✓ YesOct 2 |
| GitHub stars×2 | 242,700Oct 3 | 13,100Oct 2 | 107,200Oct 2 | ~54,870Oct 2 | 27,200Oct 2 |
| Hosted / paid option×2 | n/a | Codebuff paid subscriptions: $100/mo (1x usage)Oct 2 | Gemini Code Assist Standard $19-22.80/user/moOct 2 | falseOct 2 | n/a |
The ranking, tool by tool
The license decides this. Full DeepSeek Harness (dsh) vs Claude Code verdict →
For teams that want to read or fork the agent, the license matters most. Full Freebuff vs Claude Code verdict →
For teams that want to read, fork or self-host the agent, the license matters most, and Gemini CLI is Apache-2.0 while Google Antigravity does not publish a license. Full Gemini CLI vs Google Antigravity verdict →
For teams that want to read, fork or self-host the agent, the license comes first. Full Goose vs Claude Code verdict →
License decides this, and Grok Build is Apache-2.0 while Claude Code is proprietary, with its npm package pointing to a license in its README. Full Grok Build vs Claude Code verdict →
License decides this one. Full Kimi Code vs Claude Code verdict →
The license decides this. Full Letta Code vs Claude Code verdict →
The license is the deciding factor for teams that want to read or fork the agent. Full Mistral Vibe vs Claude Code verdict →
The license decides this, and only OpenHands has an open one: it is MIT licensed with 89,799 GitHub stars. Full OpenHands vs Devin verdict →
The license is the deciding factor: Pi is MIT licensed, while Claude Code is proprietary, pointing to a license in its README on the npm package. Full Pi vs Claude Code verdict →
The license matters most here. Full Qwen Code vs Claude Code verdict →
The license is what matters most here. Full Tabby vs GitHub Copilot verdict →
The license is what matters most here, and it is decisive. Full Warp vs Claude Code verdict →
The license matters most here, and Zed is licensed GPL-3.0-or-later with Apache-2.0 components, while Cursor publishes no license. Full Zed vs Cursor verdict →
The license matters most here. Full OpenCode vs Claude Code verdict →
Both are MIT licensed, so the license does not separate them. Full OpenCode vs Kilo Code verdict →
For teams that want to read, fork or self-host the agent, the license matters most. Full OpenCode vs Crush verdict →
For teams that want to read or fork the agent, the license matters most. Full OpenAI Codex vs Claude Code verdict →
For teams that want to read or fork the agent, the license is what matters most. Full OpenAI Codex vs Cursor verdict →
The license is what matters most for teams that want to read or fork the agent. Full OpenAI Codex vs GitHub Copilot verdict →
The license decides this: Aider is Apache-2.0 while Claude Code is proprietary, so only Aider can be read, forked and self-hosted as source. Full Aider vs Claude Code verdict →
Both carry the Apache-2.0 license, but for OpenAI Codex it covers the Codex CLI, while Aider's whole tool is free and open source. Full Aider vs OpenAI Codex verdict →
Both are permissively licensed, Aider under Apache-2.0 and gptme under MIT, and both run local models. Full Aider vs gptme verdict →
The license matters most here. Full Cline vs Cursor verdict →
Neither is open source: Claude Code is proprietary, and Command Code's npm package declares itself unlicensed. Full Command Code vs Claude Code verdict →
Neither Devin nor Factory publishes a license, a list of supported model providers, GitHub stars or a separate hosted option, so the main open source signals are missing on both sides. Full Factory vs Devin verdict →
Neither tool is a clear open-source choice. Full Google Antigravity vs Claude Code verdict →
Neither tool publishes a license, a list of supported model providers, GitHub stars or a separate hosted option, so the main open source signals are missing on both sides. Full Google Antigravity vs Cursor verdict →
For teams that want to read, fork or self-host the agent, the license matters most, and neither Cursor nor GitHub Copilot publishes one. Full GitHub Copilot vs Cursor verdict →
Neither tool is an open-source agent you can fork. Full GitHub Copilot vs Claude Code verdict →
The license matters most to teams that want to read or fork the agent, and neither Devin Desktop nor GitHub Copilot publishes one, nor GitHub stars or a supported provider list. Full GitHub Copilot vs Devin Desktop (formerly Windsurf) verdict →
Both licenses let you read, fork and self-host: Cline is Apache-2.0 and Kilo Code is MIT, so the license does not separate them. Full Kilo Code vs Cline verdict →
Spec-driven AI IDE + CLI (ex-Q Developer CLI) + web/cloud agent; AWS successor to Amazon Q Developer. No won verdicts for this use case yet; it ranks on ties and near-misses.
Frontier multi-model agent + cloud dev environments (orbs); also terminal_agent (Amp CLI, connects to VS Code/Cursor/Windsurf/Zed/Neovim). OWNERSHIP: no longer branded as Sourcegraph - operated by "Amp Frontier Corporation" (privacy policy, modified Sep 10, 2026); about page: "We are an independent agent research lab"; CLI npm package moved from @sourcegraph/amp to @ampcode/cli (May 14, 2026 changelog). Free Hobby tier with BYOK/ChatGPT-sub since Sep 13, 2026. No won verdicts for this use case yet; it ranks on ties and near-misses.
Context-engine coding platform for large codebases; now led by Cosmos cloud agents + Auggie CLI (also ide_extension: VS Code, JetBrains, Vim/Neovim chat). Brief seeded it as an IDE extension; the 2026 pricing page leads with "Try Cosmos". Flat team pricing (up to 50 seats, no per-seat charge) with a pooled $ usage balance. No won verdicts for this use case yet; it ranks on ties and near-misses.
"The Sovereign AI Lab": coding agent + own post-trained Lumen models, pitched at regulated/air-gapped orgs and niche languages (COBOL, Fortran, Verilog). Formerly branded Genie (Genie CLI introduced Q2 2025). Secondary segment: terminal_agent (cos CLI). No won verdicts for this use case yet; it ranks on ties and near-misses.
DevSecOps-native agents and flows for GitLab Premium/Ultimate; also IDE extensions (VS Code, JetBrains) and a Duo CLI; self-managed and air-gapped self-hosted-model options. No won verdicts for this use case yet; it ranks on ties and near-misses.
JetBrains IDE agent, now also a terminal agent (Junie CLI) and runs in JetBrains Air; metered in JetBrains AI Credits ($1 each) with BYOK at provider rates, zero markup. No won verdicts for this use case yet; it ranks on ties and near-misses.
Private, governed enterprise coding agent with self-hosted and air-gapped deployment. Now a Tricentis product (acquired 2026-07-30). Secondary segment: terminal_agent (Tabnine Plugin for OpenCode). No won verdicts for this use case yet; it ranks on ties and near-misses.
Low-priced AI IDE (Free/Lite/Pro/Pro+/Ultra); product family now split into TraeCode (IDE) and TraeWork (web/desktop/mobile workspace built on SOLO). Owner per brief: ByteDance - the legal entity could not be read from a primary source (privacy policy/ToS are client-rendered); enterprise page cites Douyin as a customer. Chats may be used for training unless Privacy Mode is on. No won verdicts for this use case yet; it ranks on ties and near-misses.
Plan-Code-Verify multi-model agent; desktop app (primary, most changelog activity), VS Code/JetBrains plugins, Slack/Telegram control; increasingly pitched as an 'AI technical cofounder' with app deployment. No won verdicts for this use case yet; it ranks on ties and near-misses.
Agentic AI IDE (ex-Windsurf) inside the Devin family; shares plans with Devin Cloud and Devin CLI. No won verdicts for this use case yet; it ranks on ties and near-misses.
AI IDE with CLI and cloud agents; ships its own Grok/Composer models. Acquired by SpaceX (completed 2026-08-14, per cursor.com/blog/joining-spacex); also segments terminal_agent, cloud_agent. No won verdicts for this use case yet; it ranks on ties and near-misses.
Autonomous cloud coding agent + terminal agent (Devin CLI); plans shared with Devin Desktop. No won verdicts for this use case yet; it ranks on ties and near-misses.
Terminal-first coding agent included in Claude Pro/Max/Team/Enterprise plans or billed via API; also IDE extension, desktop app and cloud sessions (cloud_agent). No won verdicts for this use case yet; it ranks on ties and near-misses.
Source-available (FSL, converts to MIT after 2 years) BYO-model terminal agent; optional Charm-hosted 'Hyper' inference subscription. No won verdicts for this use case yet; it ranks on ties and near-misses.
MIT local-first terminal agent; also web UI, desktop app, REST server, MCP/ACP; persistent autonomous-agent template. No won verdicts for this use case yet; it ranks on ties and near-misses.