Pick a terminal coding agent in 2026 and you are not really picking a model — the frontier models have largely converged, so the harness wrapped around them decides the daily experience. Four have earned a serious look in the last eighteen months, and they split along one line: Claude Code is the proprietary platform product; the other three are open source but pull in different directions — goose is the foundation-governed generalist, OpenCode is the full-featured contender with IDE-grade code intelligence, and Pi is the minimal core you extend yourself.

This is a documentation-grounded comparison, not a scored bake-off. Each tool is measured against its primary sources — official docs, repos, and release notes — on the axes that actually decide a daily driver: model lock-in, cost and auth, safety defaults, and how much setup you inherit versus build.

Scope note: Figures below are verified against primary sources as of June 25, 2026. Star counts move daily; pricing, permission modes, and subscription policies change between releases.

TL;DR comparison

Axes are ordered by how often they break a pick. Multi-item cells lead with the shared capabilities (e.g. MCP · subagents first) so you can scan straight down a column and see which harness adds or misses a given feature. Jump to goose, Claude Code, OpenCode, or Pi for receipts.

	Model flexibility	Cost model	Claude subscription auth	License / governance	Default safety	Extensibility	Code intelligence	Interface	Maturity	Signature move
goose	15+ providers, BYO key (providers docs)	API metered; no subscription bundle	No — API key only (providers docs)	Apache 2.0; AAIF / Linux Foundation (Apr 2026)	Sandbox mode, prompt-injection detection	MCP · subagents · YAML recipes	Text + tool output; no built-in LSP	CLI · desktop · API	~50K stars; setup rougher than commercial tools	Portable recipes for repeatable workflows
Claude Code	Anthropic-first; any Anthropic-API-compatible endpoint via ANTHROPIC_BASE_URL (Bedrock, Vertex, self-hosted gateways)	Subscription (interactive) + separate API credit pool (headless) (billing pause)	Yes — native OAuth (approved client) (auth docs)	Proprietary; Anthropic product	default → acceptEdits → plan → auto → dontAsk → bypassPermissions (permission modes)	MCP · subagents · skills · hooks · workflows (preview)	Strong via tools; no first-party LSP feed	CLI · IDE · desktop · web · PR bots	Most polished first-party UX	Batteries-included Anthropic stack
OpenCode	75+ models via Models.dev, BYO key	API metered; optional Go / Zen tiers	No — subscription auth removed Mar 2026 (legal)	MIT; SST / Anomaly team	Build vs Plan agents; per-tool permission matrix (agents)	MCP · subagents · custom agents (JSON/md)	LSP diagnostics fed to agent when enabled (LSP docs)	TUI · IDE extensions · desktop	~178K stars; fast-moving surface	Open-source Claude Code ergonomics + LSP
Pi	20+ providers, BYO key (npm)	API metered or subscription login	Conditional — /login works, under Anthropic’s paused metering policy	MIT core; Mario Zechner → Earendil Works (Apr 2026)	Four tools by default; add guardrails via extensions	MCP (via extension) · no subagents · TS extensions + packages	Runs tsc/lint via bash; LSP only as extension	TUI/CLI · RPC · SDK	~65K stars; power-user niche	Minimal harness you extend, not adopt

goose

goose started inside Block’s open-source program as “codename goose” in early 2025. In April 2026 Block donated it to the Agentic AI Foundation under the Linux Foundation — same umbrella as Anthropic’s MCP spec and OpenAI’s AGENTS.md. The repo lives at github.com/aaif-goose/goose (~50K stars as of June 25, 2026).

What stands out: goose is not positioning as “the open Claude Code.” It is a general-purpose, local-first agent that happens to be good at coding. Recipes — portable YAML workflow templates you can share and automate — are the feature people cite when they switch from a subscription agent. Pair a strong API model with BYO keys and community threads often report per-task cost an order of magnitude below Claude Max tiers, though that depends on model choice and loop length.

Auth: goose is API-key-only for Anthropic. There is no Claude Pro/Max subscription login — a long-standing feature request for Anthropic OAuth was closed without shipping. You bring an ANTHROPIC_API_KEY (or any of 15+ other providers) and pay per token.

Extensibility: MCP is the integration surface. Block’s docs list 70+ first-party extension patterns; the community MCP catalog is much larger. Subagents and YAML recipes round out the surface. The desktop app lowers the terminal barrier for teammates who will not live in a TUI.

Rough edges: Hacker News threads praise recipes and cost flexibility but flag orchestration that lags the underlying model — the harness feels less refined than Claude Code or OpenCode even when you point the same frontier model at both. Setup (profiles, providers, extensions) takes more fiddling than signing into a proprietary CLI.

Verdict: Best when you want vendor-neutral governance, any-model flexibility, and workflow packaging beyond one-off chat sessions. Worst when you want the smoothest out-of-box coding UX.

Claude Code

Claude Code is the polished commercial benchmark. Anthropic ships terminal, VS Code/JetBrains, desktop, web, and GitHub/GitLab PR workflows from one product surface. If you are already paying for Claude and want the least friction, this is the default incumbents measure against.

Permission modes matter more than feature checklists. The docs list six, verified against the current permission-modes page: default (reads without asking; prompts on edits and shell), acceptEdits, plan (read-only, propose don’t edit), auto (a classifier reviews each action — see our auto-review vs YOLO post for the Cursor parallel, not a duplicate deep-dive), dontAsk (only pre-approved tools, for locked-down CI), and bypassPermissions (everything; containers/VMs only). One nuance the marketing skips: only default → acceptEdits → plan sit in the Shift+Tab cycle. auto appears once your account and version qualify (research preview, v2.1.83+), dontAsk is set via --permission-mode, and bypassPermissions unlocks only behind a launch flag.

Extensibility: MCP, skills, hooks, subagents, and dynamic workflows (research preview) are first-party. The ecosystem is Anthropic-shaped — powerful, but you are inside one lab’s opinion of what an agent should do.

Billing and subscription (June 2026): Interactive Claude Code sessions draw from your subscription pool, and Claude Code is one of only two Anthropic-approved subscription clients (the other is claude.ai), so its OAuth /login is stable in a way third-party harnesses’ is not. Programmatic usage — Agent SDK, claude -p, GitHub Actions — was scheduled to move to a separate monthly API credit on June 15. Anthropic paused that change on day one; treat the split as directionally true while the timeline stays soft. If your workflow is mostly interactive terminal sessions, this matters less than if you run headless agents in CI.

Lock-in is softer than it looks. Claude Code defaults to Anthropic models, but it is not hard-wired to them. Point ANTHROPIC_BASE_URL at any gateway that speaks the Anthropic Messages API and set ANTHROPIC_AUTH_TOKEN, and you can front Bedrock, Vertex, a LiteLLM or vLLM proxy, or a self-hosted model (LLM gateway docs). That routes traffic — and billing — to your endpoint instead of your subscription, so the “Anthropic-only” ceiling is a default, not a wall. The real coupling is to Anthropic’s product opinions (modes, skills, hooks), not strictly its weights.

Verdict: Best polish and tightest integration with Claude’s model improvements. Worst out-of-box model flexibility — escapable through a gateway, but only by giving up the subscription economics that make it attractive in the first place.

OpenCode

OpenCode from the SST/Anomaly team is the open-source answer to “why can’t I have Claude Code ergonomics on whatever model I want?” The repo anomalyco/opencode had ~178K GitHub stars on June 25, 2026 — star velocity matters less than whether the TUI and permission model match how you work.

LSP is the differentiator. OpenCode can start language servers for dozens of file types and feed diagnostics back into the agent loop (LSP docs). LSP is disabled by default; OpenCode’s own docs warn that language servers can desync, eat memory, and slow sessions. It pays off most on typed codebases like TypeScript, where compiler feedback beats raw file diffs; on small scripts the overhead may not be worth it — let the agent run pnpm check via bash instead.

Agents: Tab switches between Build (full tools) and Plan (edits and bash default to ask). Subagents — General, Explore, Scout — handle parallel research without polluting the main session. Permissions are granular per tool (edit, bash, lsp, MCP wildcards).

Models, auth, and privacy: OpenCode routes through Models.dev with 75+ providers (BYO key). One sharp edge for Claude users: subscription auth is gone. In March 2026, after Anthropic’s legal push against third-party subscription use, OpenCode stripped its Anthropic OAuth plugin from the official build — you authenticate Claude with an API key, not a Pro/Max login. Code residency depends on the path you run: BYO-key and self-hosted paths keep traffic on your own infra and provider; hosted tiers (OpenCode Go at $5/mo, Zen) route through their servers — check the mode you actually use.

Tradeoffs: Community reports suggest iteration can feel slower than Claude Code on the same model — more harness overhead, more configuration surface. The contributor count is enormous; breaking changes land frequently.

Verdict: Open-source Claude Code ergonomics with real code intelligence when LSP is worth the cost. The standout choice when you want MIT license, any model, and IDE-grade agent modes without Anthropic lock-in — provided you are fine paying Claude per-token rather than on a subscription.

Pi

Pi is Mario Zechner’s deliberately tiny harness — the same engineering mind behind libGDX, and the core loop inside OpenClaw. In April 2026 Zechner announced Earendil Works; the MIT core stays MIT. The repo earendil-works/pi sat at ~65K stars on June 25, 2026.

Install note: Pi ships as an npm package, but the scope moved with the company. The current package is @earendil-works/pi-coding-agent; the old @mariozechner/pi-coding-agent scope is deprecated and frozen at 0.73.1, with 0.74.0 the first release under the new home (migration post). Old pins still resolve, but new installs should use the Earendil scope.

Minimal by design: Pi ships four tools by default — read, write, edit, bash (author post). No built-in subagents, plan mode, or MCP. The system prompt is intentionally short; Zechner’s argument is that frontier models are already RL-trained for coding — extra instructions mostly burn context. Optional read-only tools (grep, find, ls) exist if you want to restrict write access further.

Extensibility is the product: TypeScript extensions hot-reload. Skills use progressive disclosure so prompt cache stays warm. Pi packages distribute extensions via npm or git. Fifty-plus examples ship in-repo. Want MCP? Write an extension or point Pi at the docs and let it scaffold one — there is no vendor marketplace gate.

Auth — read this before you rely on it: Pi keeps a /login flow for Claude Pro/Max and ChatGPT subscriptions alongside BYO API keys. The subscription path is real but fragile. Anthropic restricted third-party harnesses from drawing on Claude subscriptions in early 2026, enforced it on April 4 (third-party calls were rejected with “draws from your extra usage, not your plan limits”), then paused the metering change on June 16. So subscription login works today, under a policy Anthropic has already changed three times this year. Budget for an API key as the stable fallback.

Modes: Interactive TUI, print/JSON for scripts, RPC for embedding, SDK for building your own agent. Armin Ronacher’s write-up on Pi inside OpenClaw captures the power-user appeal: you read the extension code, you own the behavior.

Tradeoffs: Not mass-market polish. No background bash — Pi expects tmux if you want long-running processes with full observability. The subscription /login paths reduce API friction but add auth surface — and policy risk — you must accept.

Verdict: Reshape the harness, do not adopt a product. Best when you treat the agent loop as infrastructure you version-control.

Decision guide

No single winner — four different bets on how much product you want wrapped around the model.

• Pick Claude Code if you already live in Anthropic’s subscription, want the least setup friction, and value first-party polish, hooks, and PR integrations — and note its subscription auth is the only one here Anthropic guarantees.
• Pick OpenCode if you want open source (MIT), any model, Build/Plan agent ergonomics, and LSP-fed diagnostics on typed codebases — and you are fine paying Claude with an API key, not a subscription.
• Pick goose if foundation governance matters to your org, you need recipes for repeatable cross-team workflows, or you want a generalist agent beyond pure coding with BYO keys across 15+ providers.
• Pick Pi if you are building agent infrastructure — minimal core, TS extensions you own, token-efficient prompts — and you will invest setup time once to save lock-in forever.
• Pick none of them as your only tool if you still do most editing in an IDE. These are terminal defaults, not replacements for Cursor tab-complete or a language server’s inline fixes.

One axis to pressure-test before migrating a team default: cost and auth predictability. Subscription agents feel cheap until you run autonomous loops; BYO API feels expensive until you model per-task spend. And the auth ground keeps moving — only Claude Code has guaranteed subscription access; goose and OpenCode are API-key-only for Claude, and Pi’s subscription path rides a policy Anthropic keeps revising. Log usage on one representative task in each harness, and price the API-key fallback, before you commit a team.

Scope and limits

Figures here come from primary sources, not a controlled benchmark:

• No Terminal Bench or SWE-bench scores — none are invented.
• Vendor adoption metrics (Block’s ~60% internal goose use, OpenCode’s “6.5M monthly developers”) are vendor claims, not independent measurements; GitHub stars stand in for size.
• Subscription-auth status reflects Anthropic’s June 16, 2026 pause and can shift again — re-check before you bet a workflow on it.
• Desktop UX, Windows/WSL edge cases, and enterprise SSO get only shallow coverage.

For permission-tier nuance on Claude Code and Cursor, see auto-review vs YOLO and safe mode. For harness orchestration tradeoffs on Claude Code specifically, see when to let Claude write the harness.

---

Which axis mattered most in your last agent pick — cost, lock-in, LSP feedback, subscription auth, or how much harness you wanted to own?