16 tools that write software with you — or for you. Grouped by how they fit into your workflow. Model quality matters, but so does the harness around it.
Agents that live in your shell — they read the repo, run commands, edit files and iterate until the task is done.
Anthropic · Claude models
The agentic coding tool that defined the CLI-agent category. Runs in the terminal, IDE extensions, desktop and web; supports sub-agents, hooks, MCP servers and skills. Its rise through 2025–26 made terminal-native agents the default way many engineers ship code.
Usage-based (API) or Claude Pro/Max from $20/mo
OpenAI · GPT-5.x / codex variants
OpenAI's coding agent, spanning an open-source CLI, IDE extension and a cloud sandbox for delegated tasks. Tightly coupled to GPT-5.x codex models; strong at parallel cloud task execution reviewed via PRs.
Included with ChatGPT Plus/Pro; API usage-based
Google · Gemini models
Google's open-source terminal agent with a famously generous free tier. Brings Gemini's 1M+ context windows to repo-scale work, with built-in Google Search grounding and MCP support.
Generous free tier; paid via Gemini API
Open source · Any model (LiteLLM-compatible)
The original open-source terminal pair programmer. Git-native (every change is a commit), efficient with tokens via repo-map context, and famously model-agnostic — its polyglot leaderboard is itself a respected coding benchmark.
Free (bring your own API key)
Sourcegraph · Multi-model (frontier)
Sourcegraph's opinionated agentic coding tool, built by the team behind Cody. Known for an uncompromising 'let the agent drive' philosophy, multi-thread workflows, and drawing on Sourcegraph's code-intelligence heritage.
Usage-based; free tier via ads
Block · Any model
Block's open-source local agent framework, extensible well beyond coding via MCP extensions. Runs entirely on your machine and automates anything from engineering tasks to data workflows; a reference example of MCP-first design.
Free (bring your own API key)
SST · Any model
A fully open-source terminal coding agent with a polished TUI, popular as the model-agnostic alternative to Claude Code. Pairs naturally with budget open-weights models, and its client/server design enables remote driving.
Free (bring your own key); Zen plans
Full development environments built around agents, with inline edits, background tasks and multi-file awareness.
Anysphere · Multi-model (frontier + in-house Composer)
The AI-native IDE that became one of the fastest-growing dev tools ever. A VS Code fork with deeply integrated agent mode, background agents, and its own speed-optimized Composer models alongside frontier model choices.
Free tier; Pro from $20/mo
Cognition · Multi-model + in-house SWE models
An agent-first IDE built around Cascade, its autonomous flow system. Now owned by Cognition (makers of Devin), pairing IDE ergonomics with deep autonomous-agent DNA and in-house SWE models.
Free tier; Pro from $15/mo
Amazon (AWS) · Claude models (primary)
AWS's spec-driven agentic IDE. Kiro pushes a structured workflow — requirements, design docs and task lists before code — appealing to teams that want agent output anchored to engineering process rather than vibes.
Free tier; Pro from $19/mo
Agentic power inside the editor you already use — VS Code and JetBrains extensions with plan/act workflows.
GitHub / Microsoft · Multi-model (OpenAI, Anthropic, Google)
The largest-scale AI coding assistant, evolved from autocomplete into a full platform: agent mode, a cloud coding agent that takes GitHub issues to PRs, and code review. The default choice inside enterprises standardized on GitHub.
Free tier; Pro $10/mo; Business $19/user/mo
Cline Bot Inc. · Any model
The leading open-source agentic extension for VS Code, with plan/act modes, browser use and MCP support. A favorite pairing with cheap open-weights models like GLM and DeepSeek for budget agentic coding.
Free (bring your own API key)
Roo Code Inc. · Any model
A Cline fork that grew into its own platform, known for configurable modes (architect, code, debug) and orchestration of multi-step tasks. Popular with power users who want fine-grained control over agent behavior.
Free (bring your own API key); Cloud plans
Fire-and-forget agents that take a ticket, work in a cloud sandbox and come back with a pull request.
All Hands AI · Any model
The leading open-source autonomous software engineering platform (formerly OpenDevin). Agents get a full sandboxed environment — shell, browser, editor — and resolve issues end-to-end. A reference platform for SWE-bench-style evaluation.
Free self-hosted; cloud from $20 credits
Cognition · Proprietary stack
The product that coined 'AI software engineer'. Devin works fully autonomously in a cloud workspace — planning, coding, testing, opening PRs — and is now positioned for parallel fleets of agents handling backlog tickets.
From $20/mo (Core); Team $500/mo
Google · Gemini models
Google's asynchronous coding agent: assign it GitHub issues and it works in a cloud VM, returning PRs with audio changelogs. The 'fire and forget' counterpart to the interactive Gemini CLI.
Free tier; paid plans via Google AI