Block, the payments company led by Jack Dorsey, has built a free, open-source AI coding agent called Goose that replicates the core functionality of Anthropic's Claude Code — which costs up to $200 per month — and runs entirely on a developer's local machine with no subscription fees, no usage caps, and no data leaving the user's hardware.

The arrival of Goose as a credible alternative coincides with a wave of developer frustration over Claude Code's pricing structure and newly introduced rate limits. Anthropic, founded by former OpenAI executives and headquartered in San Francisco, rolled out weekly token-based restrictions in late July that Pro subscribers ($20/month) and Max subscribers ($100–$200/month) have widely criticised as opaque and restrictive. Independent analysis suggests Pro users receive roughly 44,000 tokens per session, while the top-tier $200 plan yields approximately 220,000 tokens — limits that intensive users report exhausting within 30 minutes.

For a growing community of developers, the limitations of local models are acceptable trade-offs for something increasingly rare in the AI landscape: a tool that truly belongs to them.

Why Claude Code's Rate Limits Ignited a Backlash

Anthropic frames its new restrictions as targeting fewer than 5 percent of users — specifically those running Claude Code "continuously in the background, 24/7." But the company has not clarified whether that figure applies to 5 percent of Max subscribers or 5 percent of its entire user base, a distinction developers argue is critical. On Reddit and developer forums, reports of hitting daily limits within half an hour of serious work have proliferated, with some users cancelling subscriptions and describing the new caps as "a joke" and "unusable for real work."

The confusion stems from Anthropic's decision to express limits in "hours" of usage rather than tokens. When Claude Code's documentation states that a $200 Max subscriber receives "240 to 480 hours of Sonnet 3.5 per week," that figure is not a measure of clock time — it is a proxy for token consumption that varies depending on codebase size, conversation length, and task complexity. Developers accustomed to predictable pricing have found the ambiguity particularly galling.

How Goose Works — and What Makes It Different

Goose operates as either a command-line tool or a desktop application capable of autonomously performing complex development tasks: building projects from scratch, writing and executing code, debugging failures, orchestrating multi-file workflows, and interacting with external APIs — all without requiring constant human input. It uses "tool calling," the ability for a language model to request and execute specific system actions rather than simply generating descriptive text.

The defining architectural difference is model agnosticism. Goose connects to Anthropic's Claude models via API, to OpenAI's GPT series, to Google's Gemini, or to open-source models running entirely on local hardware through tools like Ollama. That last option is what makes a zero-cost, zero-cloud setup possible. With Ollama installed, a developer can download a model such as Qwen 2.5 or Meta's Llama series and run it directly on their own machine — no internet connection required.

"I use Ollama all the time on planes — it's a lot of fun!" said Parth Sareen, a software engineer who demonstrated the tool in a recent livestream, illustrating how local models remove even connectivity as a constraint.

Goose also integrates with the Model Context Protocol (MCP), an emerging standard for connecting AI agents to external services including databases, search engines, and third-party APIs, extending its capabilities well beyond what the base language model provides. The project has reached 26,100 GitHub stars, 362 contributors, and 102 releases, with version 1.20.1 shipping on January 19, 2026.

Setting Up a Free Local Stack

For developers willing to invest in initial configuration, a fully local, zero-cost Goose setup involves three steps. First, install Ollama from ollama.com and pull a coding-capable model — ollama run qwen2.5 downloads and launches the model in a single command. Second, install Goose via its GitHub releases page or a system package manager; pre-built binaries are available for macOS (Intel and Apple Silicon), Windows, and Linux. Third, configure the connection in Goose Desktop by navigating to Settings → Configure Provider → Ollama and confirming the default API host at http://localhost:11434.

Hardware requirements are the most significant barrier. Block's documentation recommends 32 gigabytes of RAM as a baseline for larger models. Smaller model variants can run on systems with 16 GB, though Apple's entry-level MacBook Air with 8 GB would struggle with most capable coding models. A MacBook Pro with 32 GB handles them comfortably, according to the documentation — a configuration increasingly common among professional developers.

The Real Trade-Offs Between Goose and Claude Code

Goose with a local model is not a drop-in replacement for Claude Code, and the gaps are worth stating plainly. Claude 3.5 Opus remains the strongest available model for software engineering tasks, particularly for understanding complex codebases and producing high-quality code on first attempt. One developer who upgraded to the $200 plan captured the difference: "When I say 'make this look modern,' Opus knows what I mean. Other models give me Bootstrap circa 2015."

Context window size is another meaningful gap. Claude Sonnet 3.5 offers a 1-million-token context window, sufficient to load large codebases without chunking. Most local models default to 4,096 or 8,192 tokens, though many support longer contexts at the cost of additional memory and slower processing. Cloud-based inference is also faster than consumer hardware for iterative workflows.

That said, the open-source model landscape is moving quickly. Moonshot AI's Kimi K2 and z.ai's GLM 4.5 now benchmark near Claude Sonnet 3.5 levels and are freely available, suggesting the quality premium that justifies Claude Code's pricing will face continued pressure.

In a broader competitive context, Cursor charges $20/month for Pro and $200/month for Ultra — mirroring Claude Code's pricing — while tools like Cline and Roo Code offer open-source coding assistance with less agentic depth than Goose. None combine genuine task autonomy, full model agnosticism, local operation, and zero cost in the way Goose does.

What This Means

Developers who prioritise cost, data privacy, and workflow flexibility now have a technically serious, actively maintained alternative to premium AI coding subscriptions — and the rapid improvement of open-source models means the quality gap justifying those subscriptions is narrowing with every release.