Shape Launches Windows Coding Agent With Multi-Model AI Switching
A new coding assistant called Shape shipped Thursday with a twist: developers can swap between OpenAI, Anthropic, Google, and roughly a dozen other AI providers without leaving a single desktop app, a feature the Windows-only startup says sets it apart in an increasingly crowded field.
The company abandoned its original IDE framing after a two-month rewrite that concluded with version 1.0.0 on October 9, 2026, according to its public changelog. "Shape started as an agentic IDE," the maker posted on Reddit at the time. "Over two months, I rewrote it into an agent for building software." What emerged is an agent-first workflow organized around five chat modes—Code, Ask, Plan, Visual, and Review—each governing how the underlying AI model behaves in response to a given prompt.
The multi-provider catalog is central to Shape's pitch. The landing page lists named models that span the current AI landscape, with several models showing detailed version numbers, though availability may vary. (The company has not clarified which models are live versus planned.) Developers select a model manually or let an "Auto" mode choose; named models consume credits on paid plans that reset monthly, the documentation says.
Shape's engineering blog describes the approach as a departure from single-provider agents. "You can choose from models by OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, Mistral and more, all within the same coding app," the October 9 post read. The company positions this flexibility as a hedge against the volatility of model performance, which can shift with new releases or API changes.
The product's agent modes offer different workflows for different tasks. Plan mode writes a step-by-step approach into a .shape/plans file and waits for developer approval before touching code. Review mode identifies the root cause of a bug, proposes a minimal fix, and displays per-file diffs directly in the chat thread. Visual mode pairs with an in-app browser that opens localhost apps; developers can click an element on screen, and Shape writes equivalent changes back to the source code, then shows diffs for sign-off. The documentation emphasizes that "nothing is committed for you," a design choice that keeps human oversight in the loop.
Code and Ask modes handle standard code generation and Q&A. Developers attach files to any message using @ syntax and lean on slash commands for repeatable prompts—a mechanic borrowed from Slack's interface conventions.
Shape connects to more than 20 external tools through plugins, including GitHub, Linear, Slack, Figma, Notion, Stripe, Supabase, and Vercel, the company's blog said. GitHub integration appears both as signed-in data in the UI and as a plugin accessible in chat. The Model Context Protocol, a standard for connecting AI agents to local data sources, lets developers add local MCP servers via a settings panel or an mcp.json file, with reload behavior and configuration options documented in the help section.
The product bundles a CodeMirror-based editor, a terminal, and git functionality that work offline. Generation requires a Shape account and network connection, according to the FAQ.
Shape runs exclusively on Windows 10 and later, supporting both x64 and ARM64 architectures. The FAQ confirms that macOS and Linux "are not available yet," with no public roadmap for those platforms—a constraint that may limit adoption among developers who favor Unix-based environments. The product is built with TypeScript, React, and Next.js on the front end, wrapped in a Tauri 2 and Rust desktop shell, and released under the Business Source License (BUSL-1.1), a restrictive license that permits use but prohibits competing hosted services.
Pricing starts at $15 per month for the Plus tier, $50 for Pro, and $150 for Max. The documentation indicates Plus includes 450 credits, Pro 1,800, and Max 5,800, though the public pricing page shows lower credit counts for the top two tiers (Pro 1,500, Max 4,500). The company has not publicly addressed the discrepancy, which could reflect a pricing adjustment or an error in the documentation.
Shape enters a market where platform incumbents are moving quickly. GitHub announced that agents from Anthropic, OpenAI, Google, Cognition, and xAI would become available inside GitHub Copilot subscriptions, a bundling strategy that leverages GitHub's massive developer footprint. Cursor, another independent player, offers agent mode with multiple models and a background agents API. Claude Code redesigned its desktop product in April for parallel agents with plugins, signaling that multi-agent orchestration is becoming table stakes.
Whether Shape's multi-provider flexibility will prove compelling enough to overcome Windows-only availability and the gravitational pull of platform-integrated alternatives remains an open question. For now, the company is betting that developers will pay a premium for the freedom to choose their own models—and that Windows remains a large enough beachhead to build from.
