Anthropic's terminal agent — the reference agentic loop for Claude models.
Primary + BYO Subscription Proprietary
Providers Anthropic
Pros
- +Best-in-class agentic coding loop tuned for Claude (subagents, effort levels, 1M context, fallback chains).
- +Deep enterprise model-governance controls and first-party Bedrock / Vertex / Foundry routing.
- +Mature, fast-moving tooling with strong MCP and skill support.
Cons
- –Closed source.
- –Claude-only out of the box; other models need a translating proxy, not just a base-URL swap.
Limitations
- !Pointing at non-Anthropic models requires a gateway that speaks the Anthropic Messages API (LiteLLM, OpenRouter, vLLM) via ANTHROPIC_BASE_URL.
- !Needs a paid Claude subscription or metered API key — no free tier.
Claude Desktop app
Anthropic
Claude Code in a desktop GUI — parallel agentic sessions in a graphical workspace.
Single provider Subscription Proprietary
Providers Anthropic
Pros
- +Runs multiple agentic coding sessions in parallel, each isolated in its own Git worktree until you merge.
- +Graphical workspace with visual diff review, an integrated terminal and editor, and an app-preview browser Claude uses to verify its own changes.
- +Shares the Claude Code engine and config (CLAUDE.md, MCP, hooks, skills); hand a CLI session to the GUI with /desktop.
Cons
- –The Code tab requires a paid Claude plan — no API-key or pay-as-you-go path to the coding GUI.
- –Effectively Anthropic-locked; other providers need the CLI, and only Enterprise deployments can route the desktop app elsewhere.
Limitations
- !macOS and Windows only — no Linux build.
- !Missing CLI capabilities: no headless/scripting (--print, Agent SDK), no inline autocomplete, and dialog commands like /permissions or /agents don't work in the Code tab.
OpenAI's open-source terminal agent, strongest on the GPT-5 Codex family.
Primary + BYO BYO API key Apache-2.0
Providers OpenAI
Pros
- +Genuinely open source (Apache-2.0) and inspectable.
- +Strong local/OSS-model story via the --oss flag (Ollama, LM Studio, MLX).
- +Custom OpenAI-compatible providers configurable in config.toml.
Cons
- –Best experience is tied to OpenAI / ChatGPT auth.
- –Non-OpenAI providers need manual TOML configuration.
Limitations
- !Top GPT-5 Codex models still require OpenAI billing.
- !Subscription auth applies per-window task quotas.
OpenAI's desktop GUI for Codex — parallel agentic threads, local and cloud.
Primary + BYO Subscription Proprietary
Providers OpenAI
Pros
- +Runs multiple Codex threads in parallel, each with a project-scoped terminal and isolated Git worktree.
- +Git workflows built into the UI — diffs, commits and pull-request creation without leaving the app.
- +GUI-only surfaces beyond the CLI: computer use, an in-app preview browser, image generation, voice dictation and a plugin marketplace.
Cons
- –Proprietary and effectively gated behind a paid ChatGPT plan; API-key sign-in leaves some functionality unavailable.
- –Pointing at a non-OpenAI provider means hand-editing config.toml, not a GUI dropdown.
Limitations
- !No native Linux build yet (under development).
- !Computer use is unavailable in the EEA, UK and Switzerland at launch.
Cursor's terminal agent — many frontier models through one account.
Model-agnostic Freemium Proprietary
Providers Anthropic · OpenAI · Google · xAI · Cursor · +more
Pros
- +Frontier models from Anthropic, OpenAI, Google, xAI and Cursor from one interface.
- +Runs in the terminal, GitHub Actions and automation scripts.
- +Shares the Cursor ecosystem (Slack, PR review, rules).
Cons
- –Closed source.
- –Usage-credit billing can be opaque, and there is no bring-your-own-key cost-at-provider path.
Limitations
- !Models are routed and billed through Cursor — a Cursor account is required.
- !No self-hosting.
AI-native editor pairing a VS Code-fork IDE with a separate agent-first Agents Window.
Primary + BYO Freemium Proprietary
Providers Cursor · Anthropic · OpenAI · Google · xAI
Pros
- +Tab predicts multi-line and next-action edits across the file — the signature low-latency editor UX.
- +Built as a VS Code fork, so existing extensions, keybindings and themes carry over.
- +A separate Agents Window (Cursor 3) runs and reviews many parallel agents across local, cloud, SSH and mobile — closer to Codex App or Claude Desktop than to an editor.
- +Composer, its proprietary agentic model, runs natively in the editor at lower cost than routing to frontier APIs.
Cons
- –Fully proprietary — no self-hosting, and core AI features require Cursor's cloud backend.
- –Usage-based on-demand billing on top of the subscription makes heavy-agent costs unpredictable.
Limitations
- !Background/cloud agents run on Anysphere-managed machines, not locally.
- !Indexing and agent features depend on connectivity to Cursor's servers — not a purely local tool.
Windsurf (now Devin Desktop)
Cognition
Agentic VS Code-fork IDE with its own SWE models, rebranded from Windsurf to Devin Desktop.
Primary + BYO Freemium Proprietary
Providers Anthropic · OpenAI · Google · xAI
Pros
- +Built on a VS Code fork, so extensions, keybindings, themes and debugging carry over with no relearning.
- +Ships Cognition's own fixed-cost SWE-1.x models alongside first-class Anthropic, OpenAI and Google frontier models.
- +Free tier includes unlimited Tab/Supercomplete autocomplete plus a light agent quota.
Cons
- –Frontier-model access and higher quotas are gated behind paid tiers; the free tier is limited to weaker models.
- –Pro rose to $20/mo and moved to a quota system, eroding the value edge it once held over Cursor.
Limitations
- !Cognition rebranded Windsurf to Devin Desktop on June 2, 2026 — the standalone Windsurf identity is being folded into the Devin line.
- !The local agent was swapped from Cascade to "Devin Local"; legacy Cascade is retired, so behavior differs from the original product.
Google's open-source CLI agent for Gemini models.
Single provider BYO API key Apache-2.0
Providers Google
Pros
- +Open source (Apache-2.0) with a mature agent loop and strong MCP integration.
- +Large Gemini context window.
- +Historically a very generous free quota for individuals.
Cons
- –Gemini-only natively — no first-party multi-provider support.
- –Heavy product churn around tiers and quotas.
Limitations
- !Provider-locked to Gemini unless you run a third-party router (LiteLLM, gemini-cli-router).
- !Google is withdrawing free individual access from June 18, 2026 — confirm the current tier before relying on it.
Google Antigravity
Google
Agent-first IDE where async agents plan, execute and verify coding tasks.
Primary + BYO Freemium Proprietary
Providers Google · Anthropic · OpenAI
Pros
- +Agent Manager orchestrates multiple async agents in parallel across editor, terminal and an integrated browser.
- +Agents emit verifiable Artifacts — task lists, plans, screenshots, browser recordings — you review like a doc rather than raw logs.
- +Genuine model choice in one IDE: Gemini, Claude Sonnet/Opus and GPT-OSS, with a free tier on macOS, Windows and Linux.
Cons
- –Closed-source VS Code fork — no self-hosting or auditing.
- –Gemini-first: non-Gemini models and heavier tiers burn credits faster, steering you toward Google's own models.
Limitations
- !Quota/credit economics are opaque and tightened sharply after the 2026 credit launch, with the credit-to-token rate undocumented.
- !Still an official public preview, not GA — terms, quotas and pricing remain subject to change.
GitHub Copilot CLI
GitHub
Copilot in the terminal, wired into the GitHub workflow.
Model-agnostic Subscription Proprietary
Providers Anthropic · OpenAI · +more
Pros
- +Tight GitHub integration — delegate to Explore, Plan, Code Review and coding agents.
- +Switch between Claude and GPT models with /model.
- +One subscription covers IDE, CLI and cloud agent.
Cons
- –Requires a paid Copilot subscription — no real free agentic CLI tier.
- –2026 usage-based credit billing adds per-prompt cost on top.
Limitations
- !Limited to GitHub's curated model catalog — no custom provider or base-URL.
- !Models are routed through GitHub, not your own keys.
Frontier multi-model agent with no markup on model tokens.
Model-agnostic Freemium Proprietary
Providers Anthropic · OpenAI · +more
Pros
- +Pay-as-you-go with no markup on model tokens.
- +Team-oriented features — shared threads and reusable workflows.
- +Start agents in the terminal and monitor or steer them from a synced web and mobile interface — agent control from anywhere.
- +Integrates with Sourcegraph code search.
Cons
- –Closed source.
- –CLI-first after the VS Code extension was discontinued in 2026.
Limitations
- !Model set and routing are controlled by Sourcegraph — no custom provider.
- !Self-serve free/Pro tiers were curtailed in favour of enterprise sales — confirm current limits.
Model-agnostic terminal agent with the widest provider catalog.
Model-agnostic BYO API key MIT
Providers Anthropic · OpenAI · Google · DeepSeek · Alibaba Qwen · xAI · +more
Pros
- +Provider-agnostic with 75+ providers via the models.dev registry.
- +Clean modern TUI and a client/server architecture you can drive from an editor.
- +Open source (MIT) with active development.
Cons
- –Younger codebase after the 2025 TypeScript/Bun rewrite.
- –The "opencode" naming split with Charm/Crush causes real confusion.
Limitations
- !The optional OpenCode Zen gateway adds a mild upsell layer.
- !You bring your own model access.
OpenCode Desktop
Anomaly (SST)
Open-source desktop and web GUI for the OpenCode agent, sharing state with the TUI.
Model-agnostic BYO API key MIT
Providers Anthropic · OpenAI · Google · DeepSeek · Alibaba Qwen · xAI · +more
Pros
- +Desktop app and web UI share the same server, sessions and live state as the TUI — switch surfaces mid-session.
- +The web UI launches with a single `opencode web` command, no separate install.
- +Fully open source (MIT) and provider-agnostic, inheriting 75+ models via models.dev with your own keys.
Cons
- –The desktop app is labelled Beta on every platform — less mature than the established TUI.
- –No bundled model access — you must supply your own API keys or a Zen/Copilot login.
Limitations
- !The web UI binds to 127.0.0.1 on a random port; exposing it to a team needs manual host/binding config.
- !On Windows the web UI is recommended from WSL rather than native PowerShell.
Git-native terminal pair programmer with the broadest model reach.
Model-agnostic BYO API key Apache-2.0
Providers Anthropic · OpenAI · Google · DeepSeek · +more
Pros
- +Git-native — auto-commits each edit with a descriptive message.
- +Model-agnostic via LiteLLM (100+ providers); maintains its own respected polyglot benchmark.
- +Editor-agnostic, works over plain SSH or any editor.
Cons
- –Pure terminal — no GUI or IDE panel.
- –Repo-map context tuning can need manual file adds on large repos.
Limitations
- !A pair-programming model rather than a fully autonomous multi-step agent.
- !You bring your own API keys.
Autonomous agent inside VS Code with true bring-your-own-key control.
Model-agnostic BYO API key Apache-2.0
Providers Anthropic · OpenAI · Google · DeepSeek · Alibaba Qwen · xAI · Moonshot AI · +more
Pros
- +Strong autonomous agent loop in the editor with plan/act modes.
- +30+ providers with true BYOK cost control; optional hosted inference.
- +Large install base and active development.
Cons
- –Token-hungry — high API spend on big tasks.
- –Needs careful approval supervision.
Limitations
- !Main experience is bound to the VS Code (Electron) host.
- !No fixed cost without watching token usage.
Open autocomplete, chat and agent across VS Code and JetBrains.
Model-agnostic BYO API key Apache-2.0
Providers Anthropic · OpenAI · Google · DeepSeek · xAI · +more
Pros
- +Autocomplete, chat and agent in one extension across both VS Code and JetBrains.
- +Strong local/offline/self-host story.
- +Reusable config via Continue Hub.
Cons
- –Historically more assistant than aggressive autonomous agent.
- –Configuration can be fiddly.
Limitations
- !Bound to supported IDE hosts.
- !The optional Continue Hub adds a commercial layer.
Extensible on-machine agent with a deep MCP ecosystem.
Model-agnostic BYO API key Apache-2.0
Providers Anthropic · OpenAI · Google · xAI · +more
Pros
- +Executes, edits and tests autonomously on your machine, beyond suggestions.
- +Deep MCP extension ecosystem (70+ extensions).
- +Strong governance — moved to the Linux Foundation in 2025.
Cons
- –Broad general-agent scope can feel heavier than a focused coding CLI.
- –On-machine execution needs sandboxing discipline.
Limitations
- !Capability depends on MCP extension availability.
- !Smaller first-class provider list than OpenCode or aider.
The most polished terminal UX, model-agnostic via the Catwalk registry.
Model-agnostic BYO API key FSL-1.1-MIT
Providers Anthropic · OpenAI · Google · xAI · Z.ai · MiniMax · +more
Pros
- +Best-in-class polished terminal UX (Charm's specialty).
- +Switch models mid-session while preserving context.
- +First-class Z.ai and MiniMax support, plus LSP context and MCP.
Cons
- –Source-available under FSL-1.1-MIT — not OSI open source until it converts to MIT.
- –Younger project born from the contentious opencode split.
Limitations
- !The non-OSI license matters if you require OSI-approved tooling.
- !Go binary — less hackable for JS/TS-centric users.
Gemini-CLI fork tuned for the Qwen3-Coder family.
Primary + BYO BYO API key Apache-2.0
Providers Alibaba Qwen · OpenAI · Anthropic · Google
Pros
- +Tuned specifically for Qwen3-Coder — the best Qwen agentic experience.
- +Inherits Gemini CLI tooling (skills, subagents, Claude-Code-like flow).
- +Generous, cheap Qwen access paths (ModelStudio, ModelScope).
Cons
- –Optimized around Qwen — other models work but aren't the design target.
- –As a downstream fork it trails upstream Gemini CLI changes.
Limitations
- !Qwen-primary by design and intent.
- !Non-Qwen providers go through OpenAI/Anthropic-compatible config.
Rust-fast, DeepSeek-centric agent built around prefix-cache economics.
Primary + BYO BYO API key MIT
Providers DeepSeek · Xiaomi MiMo
Pros
- +Rust-fast and local-first.
- +Strong DeepSeek prefix-caching cost optimization.
- +Approval gates, sub-agents, MCP and side-git rollback.
Cons
- –Centered on DeepSeek — best value there, less so elsewhere.
- –Smaller ecosystem and shorter track record than vendor tools.
Limitations
- !Other providers work but without the DeepSeek caching cost advantage.
- !Name collides with unrelated "Whale" projects — confirm the repo.
Minimal, fully model-agnostic agent toolkit you extend in TypeScript.
Model-agnostic BYO API key Open source
Providers Anthropic · OpenAI · Google · xAI · DeepSeek · +more
Pros
- +Minimal four-tool core (Read/Write/Edit/Bash) that self-extends via TypeScript extensions and skills.
- +Truly provider-agnostic (20+ providers) with subscription-login support.
- +Reusable as a toolkit to build your own agents.
Cons
- –Indie project versus the vendor tools.
- –More build-your-own-harness than turnkey; less mature docs.
Limitations
- !You bring model access — no bundled inference.
- !The minimalist core means more setup for advanced workflows.