Claude Code vs OpenCode in 2026: A Feature-by-Feature Comparison From the Inside

TL;DR

• Claude Code wins on: bundled test-driven skills (/run, /verify), inter-agent SendMessage primitive, JetBrains + iOS + Web surfaces, mature Windows native install, and a AskUserQuestion MCQ pattern that few other agents ship.
• OpenCode wins on: MIT-licensed open source (181k GitHub stars, direct fork model), 75+ model providers via BYOK, built-in LSP for ~32 languages with no per-language MCP wiring, JS/TS plugin ecosystem with npm access, and exhaustively configurable keybinds (~100 documented actions).
• They tie on: MCP servers, project memory (both auto-load AGENTS.md/CLAUDE.md), subagents with parallel spawn, TUI-first UX, MDM/enterprise-managed config, and slash-command extensibility.
• Neither is a wrong answer. The right pick is a function of your provider stack, your license constraints, and whether you want a batteries-included managed product or a hackable open toolchain.

Framing

By mid-2026, terminal-native AI coding agents have consolidated into two dominant patterns: Claude Code (Anthropic, proprietary CLI + Claude Agent SDK) and OpenCode (Anomaly/sst, MIT-licensed, model-agnostic). Cursor and Windsurf still ship, but they're IDE-first; Aider is losing mindshare; Codex has an OpenAI-specific slot. If you type claude or opencode in your terminal, you're at the top of the market.

This post compares them feature-by-feature with a 60-row parity matrix drawn from official docs (fetched 2026-07-02) plus five hands-on tasks run side-by-side on a corporate Windows 11 workspace with both CLIs installed (claude 2.1.196 and opencode 1.17.12).

No sponsorship, no affiliation. Both tools are legitimate; the goal is to make it easy to pick the right one for a given team.

Full parity matrix

#	Feature	Claude Code (v2.1.x, Anthropic)	OpenCode (v1.17.x, Anomaly/sst)
1	Hook events (count)	30+ including SessionStart, SessionEnd, UserPromptSubmit, UserPromptExpansion, PreToolUse, PostToolUse, PostToolUseFailure, PostToolBatch, PermissionRequest, PermissionDenied, Notification, MessageDisplay, SubagentStart, SubagentStop, TaskCreated, TaskCompleted, Stop, StopFailure, TeammateIdle, InstructionsLoaded, ConfigChange, CwdChanged, FileChanged, WorktreeCreate, WorktreeRemove, PreCompact, PostCompact, Elicitation, ElicitationResult, SessionEnd	25+ including session.created/idle/ended/compacted, message.part.updated, tool.execute.before/after, permission.asked/replied, file.edited, command.executed, lsp.client.diagnostics, server.connected, todo.updated, tui.prompt.append
2	Hook handler types	Shell command, HTTP endpoint, MCP tool, prompt, agent	JavaScript/TypeScript async functions inside a plugin module
3	Hook filtering	matcher + optional if conditional (e.g. Bash(rm *)) avoids process spawn on non-match	JS conditionals inside plugin body
4	Hook config location	settings.json (managed / user / project / local)	JS/TS files in .opencode/plugins/ or ~/.config/opencode/plugins/, or npm packages in opencode.json plugin: [...]
5	Hook decisions	Exit code 2 = block; stdout JSON {permissionDecision, additionalContext}; Stop/SubagentStop can inject context	Plugin function mutates output.args in place
6	MCP servers	Native. .mcp.json (project) + ~/.claude.json (user) + managed policy	Native. opencode.json mcp key
7	MCP transports	stdio, HTTP (streamable-http), SSE	Local (stdio) + Remote (HTTP with headers/OAuth via RFC 7591 DCR); auto-401 handling
8	MCP per-tool permissions	Yes — permissions.allow/deny pattern mcp__server__tool	Yes — per-tool permission block
9	MCP tool search (deferred loading)	Yes — ENABLE_TOOL_SEARCH=auto keeps schemas out of prefix cache	Not documented
10	Skills — format	.claude/skills/<name>/SKILL.md + supporting files; frontmatter controls invocation	Markdown in .opencode/commands/ with YAML frontmatter (description, agent, model); supports $ARGUMENTS, !bash, @file
11	Skills — cross-tool	Agent Skills open standard	Reads ~/.claude/skills/ for CC compatibility
12	Bundled skills	/code-review, /batch, /debug, /loop, /claude-api, /run, /verify, /run-skill-generator	None official
13	Skill auto-invoke	Yes — Claude decides based on description	Explicit invocation only
14	Subagents	Explore, Plan, general-purpose, claude, user-defined	build, plan (primary); general, explore, scout (subagent); user-defined
15	Parallel subagent spawn	Yes — multiple Agent calls in one turn	Yes — general designed for it
16	Background subagents	Yes — run_in_background: true on Agent + Bash	Headless opencode serve + SSE; no per-call flag
17	Subagent isolation	isolation: "worktree" (git worktree) or "remote" (cloud)	Child sessions; no worktree flag
18	Model override per subagent	Yes	Yes
19	Inter-agent messaging	SendMessage — named teammates, structured protocol messages	None equivalent; session fork instead
20	Project memory	CLAUDE.md (auto-load, @import); .claude/rules/*.md (path-scoped)	AGENTS.md (auto-load, walks up tree); .claude/CLAUDE.md fallback
21	Auto memory	Yes — Claude writes learnings automatically	Not documented
22	User-global memory	~/.claude/CLAUDE.md	~/.config/opencode/AGENTS.md
23	Org memory	Managed policy (plist, registry, managed-settings.json)	Managed dirs + macOS MDM .mobileconfig (highest priority)
24	Plugins / marketplace	/plugin install, /plugin marketplace add; Anthropic official marketplace	npm-installable + local; community "ecosystem" catalog
25	Provider support	Anthropic API + AWS Bedrock (BYOC) + Google Vertex AI (BYOC)	75+ providers via AI SDK + Models.dev; OpenCode Zen curated hub; local via lmstudio/, ollama
26	Model selection	/model opus|sonnet|haiku	/models picker; /connect OAuth; per-agent frontmatter; --variant
27	Effort / reasoning levels	/effort low|medium|high|xhigh; alwaysThinkingEnabled	Variants (high, max, low, xhigh, custom)
28	Permissions	permissions.allow/deny with wildcards	Per-category (read, edit, bash, glob, grep, task, skill, lsp, question, webfetch, websearch, external_directory, doom_loop) with allow/ask/deny
29	Permission precedence	Merge across scopes	Last-matching-rule-wins
30	Config format	JSON, no comments	JSONC supported; 8-level precedence including remote (.well-known/opencode)
31	Terminal / TUI	Primary; plan mode, diff viewer, thinking summaries	Primary; Tab-toggle Plan/Build, --mini mode, diff-wrap toggle
32	VS Code extension	Yes — official	Yes — official (sdks/vscode/)
33	JetBrains plugin	Yes (IntelliJ, PyCharm, WebStorm)	Not documented
34	Zed integration	Not documented	Yes (.zed/)
35	Desktop app	Yes — macOS + Windows x64/ARM64; scheduled tasks, side-by-side sessions	Beta — macOS/Windows/Linux (dmg/exe/deb/rpm/AppImage)
36	Web / mobile	claude.ai/code + Claude iOS app + Remote Control	Not documented
37	Headless / server	-p non-interactive, --init-only for CI	opencode serve on port 4096, OpenAPI 3.1, SSE, mDNS
38	Slash commands (built-in)	/model, /effort, /fast, /skill, /clear, /compact, /memory, /help, /config, /plugin, /rewind, /mcp	/init, /undo, /redo, /share, /help, /connect, /models, /compact, /clear
39	Custom slash commands	Yes (as skills)	Yes (markdown with frontmatter)
40	@import in memory files	Yes	Not for imports; @file for TUI fuzzy search
41	Background tasks	run_in_background: true; notification on completion; --continue resumes	Headless serve + SSE; session fork API
42	Cost model	Subscription (Pro/Max) OR API pay-as-you-go OR Bedrock/Vertex BYOC	Free, open source, BYOK on any provider; optional OpenCode Zen aggregator
43	Open-source status	CLI proprietary; Claude Agent SDK open source (Node + Python)	Fully MIT — 181k stars, 22.4k forks, 14,719 commits (fetched 2026-07-02)
44	Interactive prompts	AskUserQuestion MCQ tool (4-option cap, ranked recommendation)	Permission ask prompts only; no multi-choice widget
45	Edit tools	Read, Write, Edit (exact-string), NotebookEdit, Glob, Grep (ripgrep)	read, edit, write, patch, glob, grep; each permission-keyed
46	Diff viewer / undo	/rewind (automatic snapshots); inline in TUI + VS Code	/undo, /redo (git-based); session.diff event; diff-wrap toggle
47	Team features	Managed policy; GitHub Actions, Code Review, GitLab CI/CD integrations	Remote config via .well-known/opencode; managed dirs; MDM
48	Verbosity / transcripts	showThinkingSummaries; /config verbose=true	session_export, session_timeline, session.share, POST /session/:id/fork
49	Concurrency	Desktop side-by-side; agent teams via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1	Multi-session with child navigation; server monitors all sessions
50	Streaming	Token + tool-call interleaving	SSE at /global/event; message.part.updated streams
51	Refactor / LSP	Via MCP (e.g. Serena)	Built-in LSP for ~32 languages; auto-download; off by default
52	Test-driven workflow	/verify + /run + /run-skill-generator bundled (v2.1.145+)	Custom commands with !bash injection; no dedicated primitive
53	File handling	Read, Write, Edit, NotebookEdit, Glob, Grep	read, edit, write, patch, glob, grep; @filename fuzzy search
54	External CLI	Bash / PowerShell tool; wildcard permissions	bash tool; plugin $ = Bun shell API
55	Notifications	Notification hook + claude-notifications-go plugin; iOS push	Hook-based via session.idle (community example: osascript)
56	Session lifecycle	/clear, /compact, --continue, /rewind	session_new/list/fork/rename/delete/interrupt/compact/share + child navigation
57	Language support	All (LSP-agnostic; base model + MCP)	32 first-class LSP; general fallback for rest
58	Keybindings	Standard TUI; not deeply configurable	Fully configurable via ~/.config/opencode/tui.json; ~100 documented actions
59	Windows support	Native installer, WinGet, Chocolatey	WSL recommended; native via Choco/Scoop/npm/Docker
60	GitHub activity	anthropics/claude-code closed source	anomalyco/opencode (was sst/opencode) — 181k stars, 22.4k forks, 14,719 commits, 23 language READMEs

Hands-on: five tasks on the same workspace

Tested 2026-07-02 on a Windows 11 workspace with claude 2.1.196 and opencode 1.17.12 installed side-by-side. The workspace has ~40 MCP servers, a rich AGENTS.md/CLAUDE.md, and hooks for token compression and knowledge lookup.

Task 1 — Read AGENTS.md, summarize the top constraints

• Claude Code: Zero-work — AGENTS.md auto-imports on every session per this workspace's config. Answer available before I finished typing the question.
• OpenCode: cd C:/D/oriz && opencode run --auto "..." cold-started, read the file, returned a correct 3-bullet answer in ~25 seconds on opencode/big-pickle (a free OpenCode Zen model). Surprise win: no per-agent config needed — same AGENTS.md picked up automatically.

Task 2 — Grep for a symbol and rename

Both handle text-level rename identically (Grep + Edit with replace_all, or grep + edit). For symbol-level: Claude Code reaches for the Serena MCP server; OpenCode ships built-in LSP for 32 languages, no MCP needed. OpenCode's LSP is off by default, which its own docs justify as a memory trade-off; enable per-language.

Task 3 — Spawn a subagent to research

Both offer parallel subagents. Claude Code has a named researcher subagent pinned to Haiku for cost, invoked via the Agent tool. OpenCode has general, explore, scout — invoke inside a session via @mention. Friction observed: opencode run --agent explore ... failed with "agent 'explore' is a subagent, not a primary agent" — subagents can only be spawned from inside a primary session, not from the shell wrapper.

Task 4 — Fire a hook on user-prompt-submit

• Claude Code: Add a JSON block to settings.json under hooks.UserPromptSubmit. Filter with matcher + if. Well-worn — this workspace uses it for three separate hooks.
• OpenCode: Write a TypeScript plugin in .opencode/plugins/, export an async function returning a hooks object with the event name (session.updated, message.part.updated, etc.). Higher setup cost (need Bun, plugin types), but access to the full npm ecosystem via bun install.

Task 5 — Use an MCP server tool

• Claude Code: .mcp.json at workspace root; 40+ servers wired for this project. mcp__toolbox__search_toolbox for discovery. ENABLE_TOOL_SEARCH=auto keeps schemas out of the cache prefix — meaningful cache-hit-rate win.
• OpenCode: opencode mcp CLI + mcp block in opencode.json. Supports local (stdio) + remote (HTTP with headers or OAuth via RFC 7591 Dynamic Client Registration); auto-handles 401s on remote. No deferred-loading feature documented.

Recommendation matrix

If you…	Pick
Have a Claude subscription and want a smooth managed experience	Claude Code
Want to route to OpenAI, Google, DeepSeek, or local Ollama	OpenCode
Need JetBrains, Web, or iOS surfaces	Claude Code
Want first-class LSP for many languages with no per-language MCP wiring	OpenCode
Want bundled /verify + /run test-driven loops	Claude Code
Prefer JS/TS plugins with npm ecosystem over JSON hooks	OpenCode
Need MDM/enterprise-managed policies	Either — both support it
Are on Windows without WSL	Claude Code (native installer + WinGet)
Want an MIT-licensed open-source tool you can fork	OpenCode
Want deep parallel-subagent orchestration with named messaging	Claude Code
Want to mix free + paid models per task	OpenCode (Zen + BYOK)

Conclusion

The parity between these two products in mid-2026 is closer than the discourse suggests. Both auto-load AGENTS.md/CLAUDE.md. Both ship native MCP support with per-tool permissions. Both have 25-30+ hook events, parallel subagents, headless server modes, and mature TUIs. The differentiators cluster into three groups:

1. Managed vs open — Claude Code is a proprietary, subscription/API product with polish, marketing, and Anthropic's model in the driver's seat. OpenCode is MIT-licensed, provider-agnostic, and moves in the open. Both are legitimate design choices with different failure modes (SaaS lock-in vs BYOK complexity).

2. Batteries-included vs assembly-required — Claude Code bundles skills (/verify, /run, /rewind), the AskUserQuestion MCQ pattern, and an inter-agent SendMessage primitive. OpenCode ships fewer bundled skills but a more powerful plugin runtime (JS/TS with npm access) and built-in LSP for 32 languages.

3. Surface breadth — Claude Code covers CLI, VS Code, JetBrains, desktop, web, and mobile. OpenCode covers CLI, VS Code, Zed, and desktop-beta, plus a first-class headless server that lets you build your own surfaces.

If you already work inside the Anthropic ecosystem, Claude Code is the path of least resistance. If your team uses multiple providers, wants to self-host, or needs to embed an agent in a custom surface, OpenCode is unusually strong. Most teams will benefit from having both installed — they don't conflict, they read each other's memory files, and the friction of switching is close to zero.

References

1. Claude Code — Hooks reference (fetched 2026-07-02)
2. Claude Code — Overview (fetched 2026-07-02)
3. Claude Code — Skills (fetched 2026-07-02)
4. Claude Code — MCP (fetched 2026-07-02)
5. Claude Code — Subagents (fetched 2026-07-02)
6. Claude Code — Settings (fetched 2026-07-02)
7. Claude Code — Memory (fetched 2026-07-02)
8. Claude Code — Plugins (fetched 2026-07-02)
9. OpenCode — Intro (fetched 2026-07-02)
10. OpenCode — Plugins (fetched 2026-07-02)
11. OpenCode — Agents (fetched 2026-07-02)
12. sst/opencode GitHub repository (fetched 2026-07-02)
13. Firecrawl — Claude Code vs OpenCode (2026-06-01)
14. Thomas Wiegold — I Switched From Claude Code to OpenCode (2026-02-17)
15. Danilchenko — OpenCode vs Claude Code vs Cursor (2026-06-09)
16. Turbodocx — Cursor vs Claude Code vs OpenCode 2026 (2026-05-10)
17. Medium (unicodeveloper) — Claude Code vs Codex vs OpenCode (2026-05-22)