Qwen Code: The Open-Source Terminal Agent, Without a Meter

Qwen Code is Alibaba's open-source terminal coding agent: an Apache 2.0 fork of Google's Gemini CLI (27,000+ GitHub stars) optimized for the Qwen model family. It reads and edits your repo, runs shell commands, plans multi-step tasks, and supports skills, subagents, Agent Teams, hooks, and MCP servers - plus extras no other free CLI ships: a daemon mode (qwen serve) over HTTP+SSE, SDKs in TypeScript/Python/Java, IDE plugins for VS Code/Zed/JetBrains, a desktop app, and an Agent Arena mode that races two models on the same task. It installs with npm install -g @qwen-code/qwen-code and works with any OpenAI-compatible endpoint.

The CLI is free forever (Apache 2.0), but the hosted tokens are not. The Qwen OAuth free tier - advertised for a year as 2,000 free requests per day - was cut to 100 requests/day on April 13, 2026 and discontinued on April 15, 2026, per notices the project itself posts on its README and npm page. Many 2026 comparison posts still advertise the dead 1,000-request figure. The corrected reality: the tool is free, the tokens come from the Alibaba Cloud Coding Plan (~\$50/mo), OpenRouter or Fireworks BYOK, or your own local models via Ollama/vLLM - which costs \$0 per token.

Qwen3-Coder posts 69.6% on SWE-bench Verified in independent reads - above the Gemini 2.5 Pro class it forked from. 2026 comparisons credit the Qwen 3.6 Plus generation with about 78.8% on SWE-bench Verified, and Qwen3-Coder-Next (80B total, 3B active params) reaches 70.6% Verified / 44.3% on the harder SWE-bench Pro. Qwen 3.7 Max (May 2026) natively speaks the Anthropic Messages protocol: community testing puts it at 69.7 vs Opus 4.6's 65.4 on Terminal-Bench 2.0 and 60.6 vs 57.3 on SWE-bench Pro - while a workflow costing ~\$300 in Opus tokens runs for roughly \$50 on Qwen pricing. Claude's frontier models still lead on the hardest reasoning; Qwen's edge is price-per-point and open weights.

Yes - it is the strongest genuinely-free setup in the 2026 terminal-agent field. Configure the modelProviders block in ~/.qwen/settings.json to point at an Ollama or vLLM endpoint, and Qwen Code runs against your own machine with no API key and no recurring token bill. The practical engine is Qwen3-Coder-Next: an 80B-parameter MoE with only 3B active per token (hybrid Gated Attention + Gated DeltaNet), trained as an agent on real scaffolds, scoring 70.6% SWE-bench Verified while fitting a single workstation. The trade: quality drops with smaller local models, so complex multi-file refactors may need a second prompt.

Different meters for different jobs. Qwen Code: Apache 2.0 framework, open-weight models, multi-protocol provider switching (OpenAI/Anthropic/Gemini/Qwen plus local), the cheapest token path - but the autonomous edit-test-fix loop generally needs more steering than Claude Code, and hosted inference runs on Alibaba's cloud (self-hosting is the compliance workaround). Claude Code: the deepest ecosystem (hooks, skills, custom subagents, enterprise docs) and the strongest verified benchmark scores, but proprietary and subscription-priced from \$20/mo. Many developers run both - which is exactly the VibeFuse shape: both CLIs as widgets on one free Windows canvas, each billing its own vendor account.

Yes, and Moonshot-style env-var swaps are the community favorite. Qwen 3.7 Max natively speaks the Anthropic Messages protocol at the endpoint level, so Claude Code needs only two environment variables (ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY) plus a model ID to run on Qwen - your existing CLAUDE.md, hooks, MCP servers, and slash commands stay untouched. Independent testing puts the token cost at roughly one-sixth of Opus-class pricing with comparable Terminal-Bench and SWE-bench Pro scores. The reverse also works: Qwen Code can be pointed at Claude models via its multi-protocol config.

Qwen Code is one of the five official agent CLIs VibeFuse ships as widgets - the first free widget-based AI harness for Windows. Download the installer, register a free license key, and run Qwen Code next to Claude Code, Codex CLI, Gemini CLI, and Cursor over your actual repositories on one canvas. There is no harness meter: VibeFuse charges \$0, your vendor subscriptions do the token billing, and your code never uploads by default. Because the canvas is vendor-agnostic, you can route a cost-sensitive bulk task to Qwen Code and a high-stakes review task to Claude Code in the same session - plus an open-source marketplace where creators sell widgets, skills, and styling packs for revenue share.

Directly related: Qwen Code is an open-source fork of Google's Gemini CLI, acknowledged in the Qwen Code README. The lineage now runs opposite to the original's fate - Google discontinued consumer Gemini CLI access on June 18, 2026 (the free tier, then all non-enterprise tiers; enterprise Code Assist licenses keep it), while the fork keeps shipping under Alibaba with near-daily releases and a model-agnostic config layer the original never had. If you liked the Gemini CLI terminal shape, Qwen Code is where that architecture still lives as a consumer tool.