I've been using Claude Code daily for months. It's good. But when I sat down to map out what else exists in the terminal coding agent space, I realised I was sitting inside one walled garden and barely knew what the others looked like -- or that some of them had doors connecting to each other in ways I didn't expect.
Here's what turned up.
The Big Five (Plus Some You Haven't Heard Of)
There are now at least ten serious terminal-based coding agents. The big five that most people encounter:
| Tool | Maker | Open Source | Model Lock | Monthly Cost |
|---|---|---|---|---|
| Claude Code | Anthropic | No (source-available) | Anthropic only | $20-200 |
| OpenCode | SST/Anomaly | MIT | Any (75+ providers) | Free + API key |
| Copilot CLI | GitHub/Microsoft | No | Multi (Claude, GPT, Gemini) | $10-39 |
| Codex CLI | OpenAI | Apache 2.0 | OpenAI only | $20-200 (via ChatGPT) |
| Pi | Mario Zechner | Yes | Any (15+ providers) | Free + API key |
And a few significant ones that often get missed:
- Gemini CLI (Google) -- 98k GitHub stars, generous free tier, open source. Probably the biggest omission from most people's mental maps.
- Aider -- 42k stars, the original terminal AI pair programmer. Git-native, multi-model, mature.
- Goose (Block/Square) -- 33k stars, Apache 2.0, MCP-native, now a Linux Foundation project.
- OpenClaw -- 322k stars, the viral personal AI assistant layer. Recently got cut off from Anthropic subscriptions (more on that below).
The landscape is bigger than it looks from inside any one tool.
Claude Code -- the one most readers here will know. Source: claude.com
Claude Code: Subsidised, But Hitting the Walls
The Max plans are heavily subsidised. One analysis estimates API-equivalent usage would cost ~$3,650/month for what you get at $200. Anthropic is likely still profitable on average users, though -- typical daily usage works out to roughly $6 in API-equivalent spend, and actual compute costs are a fraction of retail API prices.
The limits have been the story of early 2026. Anthropic admitted in March that people were "hitting usage limits in Claude Code way faster than expected." Max 5x users ($100/month) reported burning through quotas in an hour. A promotional doubling period ended March 28, peak-hour quotas were reduced, and some users suspected cache bugs were silently inflating costs.
Then on April 4-5, Anthropic cut off third-party tools like OpenClaw from using subscription quotas entirely. Boris Cherny from Anthropic: "our subscriptions weren't built for the usage patterns of these third-party tools." Overflow billing (pay-as-you-go at API rates) now exists as a release valve.
OpenCode: 138k Stars and a Cat-and-Mouse Game
OpenCode is the big open-source competitor. MIT licensed, 138k GitHub stars (nearly double Claude Code's ~71k), supports 75+ LLM providers with a clean client/server architecture.
The interesting story is the Anthropic blocking saga. Since January 2026, Anthropic has progressively shut down third-party tools from using Claude subscription tokens -- server-side checks in early January, client fingerprinting and signature verification by late January, and a formal ToS update in February explicitly prohibiting OAuth token use outside Claude Code.
DHH called it "a paranoid attempt to force devs into Claude Code." OpenCode gained 18,000 GitHub stars in two weeks after the ban.
OpenCode's terminal UI -- note the Claude Opus 4.5 model selector in the bottom bar. Source: github.com/anomalyco/opencode
The Copilot Backdoor
Here's where it gets interesting. GitHub officially partnered with OpenCode on January 16, 2026. Copilot subscribers can authenticate directly via /connect and get access to Claude models -- Opus 4.6, Sonnet 4.6, Haiku 4.5 -- through the Copilot provider.
So the flow is: pay for GitHub Copilot Pro ($10/month) or Pro+ ($39/month), connect it to OpenCode, and use Claude models at a fraction of a direct Anthropic subscription.
The catch? Copilot reports truncated context windows for Claude (144k instead of the real 1M) and only exposes a single thinking level. Whether Anthropic considers this a sanctioned path or a loophole they haven't closed yet is genuinely unclear -- GitHub's integration is official, but there's no explicit Anthropic statement blessing this specific use case.
Copilot CLI: The Dark Horse
This one was a surprise. GitHub Copilot CLI went GA on February 25, 2026, and it's a full agentic coding environment -- not just a shell helper. Plan mode, Autopilot mode, sub-agents, session memory, background cloud delegation.
Copilot CLI -- a full agentic environment, not the shell helper you might expect. Source: github.com/features/copilot/cli
The key differentiator: multi-model by default. Claude, GPT, Gemini -- all through one subscription. The default model is actually Claude Sonnet 4.5. Pricing starts at $10/month for 300 premium requests, with a free tier at 50. If you want multi-model access without managing multiple subscriptions, this might be the best value in the space right now.
Codex CLI: Your ChatGPT Sub Has Free Compute
If you're already paying for ChatGPT Plus, you have coding agent compute sitting idle. Codex CLI is OpenAI's open-source (Apache 2.0) terminal agent, built in Rust, bundled with Plus and Pro subscriptions.
Codex CLI's splash screen -- Rust-based, open source, with a plan-and-execute workflow. Source: github.com/openai/codex
It's locked to OpenAI models, but the unique feature is cloud agent mode -- it can work autonomously in a sandbox, commit code, and open PRs while you do something else. If you're already paying that $20/month, there's compute on the table.
Pi: The Minimalist's Choice
Created by Mario Zechner (of libGDX fame), Pi ships with exactly four tools: read, write, edit, bash. No MCP, no sub-agents, no permission popups, no plan mode. The system prompt is under 1,000 tokens.
The philosophy is radical minimalism and self-extensibility: need the agent to do something new? Ask it to extend itself. Armin Ronacher (Flask creator) calls it "the coding agent I use almost exclusively" and built OpenClaw's agent layer on top of it.
Reviews praise its token efficiency -- roughly 4,500 tokens per task versus ~17,000 for Gemini CLI on equivalent work. Smaller community (~25k stars) but growing fast, partly on the back of YouTube coverage.
Pi's session tree -- four tools, sub-1000-token system prompt, and a URL you won't forget. Source: shittycodingagent.ai
VS Code: The Multi-Agent Switchboard
VS Code isn't just running Copilot anymore. It's become a multi-agent platform where different backends run side by side in the same chat view:
- Copilot's built-in agent -- VS Code's own harness, supports Claude/GPT/Gemini via the model picker
- Copilot CLI sessions -- background agents that persist even when VS Code closes
- Claude third-party agent -- runs the Claude Agent SDK, which is actually Claude Code under the hood
- Codex third-party agent -- similar setup for OpenAI
There's a subtle but important distinction: selecting Claude as the model in Copilot's picker just swaps the LLM behind Copilot's own agent harness. Selecting Claude as a third-party agent spins up Claude Code's actual runtime via the Agent SDK.
And then there's Anthropic's standalone Claude Code extension (9.2M installs) -- the full Claude Code experience in its own sidebar, billed through your Anthropic account. It doesn't share state with any of the above.
The Weird Architecture: SDK Wraps CLI
The most surprising finding was architectural. The Claude Agent SDK -- the library for building custom agents programmatically -- doesn't call the Claude API directly. It bundles the Claude Code CLI binary and spawns it as a subprocess.
The Python SDK README: "The Claude Code CLI is automatically bundled with the package - no separate installation required!" The official docs subtitle: "Build production AI agents with Claude Code as a library."
This is backwards from how you'd normally design it:
Textbook: Reality:
Core Library Claude Code CLI (monolith)
^ ^ ^
| | | subprocess
CLI SDK Agent SDK (wrapper)Claude Code was built product-first, iterates constantly, and wrapping the CLI gave Anthropic an SDK quickly with a single source of truth. The v0.1.x version numbers suggest this probably isn't the final architecture -- the rename from "Claude Code SDK" to "Claude Agent SDK" and breaking changes in v0.1.0 point toward a cleaner separation eventually.
What I Still Don't Know
- Rate limits for Claude via Copilot in OpenCode. Sources confirm limits exist but nobody quantifies them.
- Whether the Copilot-to-OpenCode-to-Claude path violates Anthropic's ToS. GitHub's integration is official. Anthropic partnered with GitHub. But nobody's explicitly said this specific use case is fine.
- Whether the March 2026 cache bugs were real. Users claimed reverse-engineering revealed bugs inflating costs 10-20x. Anthropic described the issues as "intentional capacity management."
- Feature parity between Claude-via-Copilot and standalone Claude Code. The truncated context window and limited thinking modes are confirmed, but there may be other differences.
- How the Agent SDK architecture evolves. Will the CLI-wraps-SDK inversion get unwound, or is this the permanent shape?
Where This Leaves Us: The DOS Era
Every tool in this post is a command-line interface. All of them. The most popular coding agents in the world are terminal programs with text input and text output.
This feels like the DOS era of AI-assisted development. The power is there, but the interaction model hasn't evolved yet. What will be the Windows and macOS of coding agents?
Three candidates are emerging:
Forked IDEs are the current frontrunner. Cursor has 2M+ users and $2B ARR -- by far the largest install base. Windsurf (now owned by Cognition/Devin) has 1M+. These take VS Code and rebuild the experience around AI. But they're essentially DOS with a nicer shell -- the AI is woven into the editing experience rather than offering a fundamentally different interaction model.
VS Code as platform is Microsoft's play. Since January 2026, VS Code has explicitly positioned itself as "the home for multi-agent development" -- Claude, Codex, and Copilot all run side by side in the same chat view. This is the Windows strategy: be the platform other agents plug into rather than fork.
Standalone mission control is the newest and perhaps most interesting pattern. Purpose-built GUIs for managing agents, not editing code. Vibe Kanban (24.5k stars) dispatches work to Claude Code, Codex, and Amp via a kanban board. Claude Code Desktop added "Dispatch" -- assigning tasks from your phone. Sculptor runs parallel agents in Docker sandboxes with visual orchestration.
This last category feels like it's solving the right problem. As agent count per developer goes from one to five or twenty running in parallel, the interaction shifts from "talk to an agent" to "manage a fleet." That needs a genuinely new interface, not a terminal with a GUI skin.
The online debate still skews CLI because the people loudly discussing these tools are comfortable in terminals. But the numbers tell a different story -- Cursor's 2M users dwarf OpenCode's 138k stars. The GUI users are just quieter.
EMAOrchestrator -- one more for the pile.
In the spirit of full disclosure: I'm building one too. EMAOrchestrator (Everybody's Making An Orchestrator) -- because at this point, the surest sign that a tooling category has arrived is that everyone and their dog is building their own version of it. One more for the pile.
Written by Claude as dictated by Harald.
Comments (0)
No comments yet. Be the first!