OpenCode vs Cursor vs Claude

Aspect	OpenCode	Claude Code	Cursor
Category	agent runtime	coding CLI assistant	AI IDE
Open source	yes	no	no
Model support	many	Anthropic only	many
Primary interface	CLI / TUI	CLI	GUI editor
Agent autonomy	high	medium	low
Multi-agent orchestration	native	limited	minimal
LSP integration	yes	no	yes
IDE features	none	none	full
Tool execution	built-in	built-in	limited

Agents vs Skills vs Subagents

Capability	OpenCode	Claude Code	Cursor
persistent agents	yes	limited	no
multi-agent workflows	yes	partial	no
tool orchestration	strong	moderate	limited
autonomous loops	yes	partial	no
IDE integration	none	none	full

Terminology is inconsistent across these tools.

OpenCode agents are markdown files. You drop a .md file in .opencode/agent/, give it a frontmatter description, and the body becomes the agent's system prompt. You can optionally set a model in the frontmatter to pin the agent to a specific model. Permissions work through a per-agent allowlist/denylist of tools. Agents don't pick their tools, they inherit all tools and you restrict from there. The runtime exposes these as named agent types that can be spawned via the Task tool.

Claude Code skills are also markdown files, but they inject context into the current session rather than spawning a new one. Claude Code does have a Task tool for spawning subagents. It works similarly to OpenCode's, but you don't define custom subagent types the same way. You get whatever named agent types the system provides.

Cursor subagents don't really exist as a user-facing concept. Cursor has Agent mode where it takes more autonomous steps, and Background Agents for longer-running tasks. Neither lets you define a named agent type with its own model or permission set. You get one agent, driven by whatever model you picked, working in your IDE.

The practical difference: in OpenCode you can define a reviewer agent pinned to one model and a coding agent on another, then have a primary agent spawn both in parallel via the Task tool and coordinate the output. In Claude Code you can spawn subagents but you can't define custom named agent types with their own models. In Cursor you press a button and hope for the best.

About LSP Support

OpenCode runs Language Server Protocol clients in the background. The agent gets a tool that can jump to definitions, find references, inspect types, and walk call hierarchies. Same operations you'd do in an IDE, but available to the agent.

Claude Code doesn't have this. It navigates code with grep, glob, and file reads. That works surprisingly well for most tasks, but in a large typed codebase where save appears in 40 files, LSP gets you to the right one faster.

Cursor has LSP by default since it's a full IDE. OpenCode is the only CLI tool that offers it.

For blog writing or small projects, this doesn't matter. It shows up in large TypeScript or Python codebases where the type hierarchy and call graph are complex.

OpenCode as a Coding Assistant

The simplest way to use OpenCode is as a drop-in replacement for Claude Code. Install it, point it at a provider, and start prompting.

Install

# pick one
curl -fsSL https://opencode.ai/install | bash
npm i -g opencode-ai@latest
brew install anomalyco/tap/opencode

Connect a provider

Run opencode in your project directory. Then run /connect in the TUI to add an API key. It walks you through picking a provider and stores the key in ~/.local/share/opencode/auth.json.

OpenCode supports 75+ providers through the AI SDK: Anthropic, OpenAI, Google, Bedrock, Groq, local models via Ollama, and others. They also offer "Zen", their own curated model list if you just want something that works without picking a provider.

Set a model

In opencode.json at your project root (or ~/.config/opencode/opencode.json globally):

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-sonnet-4-6"
}

Here are the main model options and what they cost through their respective APIs. Prices are per million tokens, as of March 2026.

Provider	Model	Input $/MTok	Output $/MTok	Notes
Zen	Big Pickle	free	free	`/connect` → OpenCode Zen
Zen	GPT 5 Nano	free	free	`/connect` → OpenCode Zen
Anthropic	Claude Opus 4.6	$5	$25	console.anthropic.com
Anthropic	Claude Sonnet 4.6	$3	$15	console.anthropic.com
Anthropic	Claude Haiku 4.5	$1	$5	console.anthropic.com
OpenAI	GPT-5.4	$2.50	$15	Best all-rounder, includes coding
OpenAI	GPT-5.3 Codex	$1.75	$14	Coding-specialized
Google	Gemini 3.1 Pro	$2	$12	Flagship, replaces 3 Pro
Google	Gemini 3 Flash	$0.50	$3	Fast and cheap
Google	Gemini 2.5 Flash	free	free	Free tier at lower rate limit
xAI	Grok 4	$3	$15	2M context window
xAI	Grok 4.1 Fast	$0.20	$0.50	Budget, very fast
Groq	Llama 4 Maverick	$0.20	$0.60	Fast inference platform
Groq	Llama 3.3 70B	$0.59	$0.79	Solid open-source option

Big Pickle is OpenCode's "stealth model". They won't say what it actually is. The community thinks it's GLM-4.6 under a different name. It launched as a free weekend stress test and stuck around. 200K context window, free through Zen, and it's the default model you get when you /connect without paying anything.

Note: Your $20/mo Claude Pro subscription does NOT give you API access. OpenCode has an unofficial "Claude Pro/Max" auth option in /connect that piggybacks on your subscription, but Anthropic says it's unsupported. For reliable use, get a separate API key or use OpenCode Zen.

Google's free tier is worth knowing about. Gemini 2.5 Flash is free at the lower rate limit, which is fine for trying things out.

What about local models? If you're on a 16GB laptop (M2 Mac, etc), the free cloud tiers beat anything you can run locally. You're limited to ~8B parameter models on 16GB RAM, maybe a quantized 14B if you squeeze. A local Llama 8B or Qwen3 8B via Ollama will be noticeably worse at coding tasks than Big Pickle or Gemini Flash, and slower too. Run local if you need privacy or offline access, not for quality.

Init and go

cd /path/to/project
opencode

Run /init to generate an AGENTS.md file that maps out your project structure. This gives the agent context about the codebase, similar to a CLAUDE.md file. Commit it to the repo.

From there, it's the same workflow you'd use in Claude Code. Ask questions, request changes, reference files with @. Hit Tab to switch between build mode (full tool access, makes changes) and plan mode (read-only, just analyzes). /undo and /redo work as expected.

The TUI is where OpenCode diverges from Claude Code. It's a terminal app with a real UI, not a pure CLI. Built by neovim users and the terminal.shop team, so the keybindings and layout feel like a terminal-native tool rather than a chat window.

OpenCode Multi-Agent Architecture

Note: I'm entirely new to this and going on early research and testing, not some deep pool of experience. This is just, like, my opinion, man.

The general idea: a planning agent decomposes jobs, dishes them out to named agents, and each does their thing. A validator or reviewer takes a look and either declares victory or loops back to try again.

graph TD
    coordinator --> decompose[task decomposition]
    decompose --> research[research agent]
    decompose --> coding[coding agent]
    decompose --> tests[tests agent]
    research --> reviewer
    coding --> reviewer
    tests --> reviewer
    reviewer --> finalize

Shared State Between Agents

Each agent runs in its own session with its own context window. Sessions are persisted in a local SQLite database. Messages, tool call results, todos, and diffs are all stored there. That's the actual shared substrate, not some in-memory object store.

What agents share in practice is the filesystem. When a coding agent writes a file and a test agent runs tests against it, that works because they're both operating in the same working directory. The spawning agent passes task descriptions and file paths in the prompt; it doesn't share memory references.

Subagent sessions have a parent_id linking them to the session that spawned them. The Task tool supports a task_id parameter so a spawning agent can resume the same subagent session later rather than starting fresh each time.

Context doesn't flow automatically between agents. If a research agent finds something important, the spawning agent has to explicitly include it in the next task prompt. A coordinator that passes rich context works well. One that just says "go code the thing" leaves subagents flying partially blind.

Note: I've seen claims that OpenCode has "Shared Memory Objects", "Message Queues", and a "Graph Execution Engine". I checked the source and none of those exist. Don't believe everything ChatGPT tells you about a codebase.

Ruler Support for OpenCode

If you use Ruler to manage rules across multiple coding agents, it does support OpenCode. Here's what it handles and what it doesn't.

What ruler does:

Instructions, writes your .ruler/ markdown rules to AGENTS.md, which OpenCode reads for project context.
MCP servers, merges MCP server definitions into opencode.json alongside your existing config.
Skills, copies .ruler/skills/ to .opencode/skills/. This is experimental in ruler, but it works. Skills are SKILL.md files that inject context into the agent session.

What ruler does not do:

Agent definitions. The .opencode/agents/*.md files that define named agents with their own model, mode, temperature, and tool permissions are OpenCode-specific. Ruler's interface has no concept of these because no other tool has an equivalent. You manage agent definitions directly.

In practice this means ruler handles the shared stuff (rules, MCP, skills) and you maintain your agent configs by hand in .opencode/agents/.

MVP

Custom agents are markdown files. The filename is the agent name. Drop them in .opencode/agents/ in your project root (or ~/.config/opencode/agents/ for global agents). Subdirectories work too. OpenCode globs agents/**/*.md, so in a monorepo you can scope agents by project: .opencode/agents/blog/, .opencode/agents/api/, etc.

The agents below live in .opencode/agents/blog/ in the monorepo.

Why not just use one agent?

For most tasks, a single Claude or Opus session is faster and simpler. The swarm adds complexity. Here's when it earns that complexity back.

Independent verification. A single agent that researches, writes, then fact-checks its own work has confirmation bias. It wrote the claim, so it finds evidence supporting it. A separate fact-checker starts with a blank context and verifies independently. That's genuinely better.

Context isolation. A single agent doing all four jobs accumulates a huge context full of research notes, draft revisions, and tool call noise. By step 4 the early research is getting pushed out or diluted. Each subagent gets a focused context with just its task.

Model matching. Web search tasks on free Gemini, prose on Big Pickle. A single agent is stuck on one model for everything. Mixing models saves money and plays to each model's strengths.

Tool restriction. The researcher can't write files. The reviewer can't edit the draft. A single agent has all tools and you're trusting the prompt to prevent misuse.

When it's not worth it: speed (it's slower sequentially), simplicity (more moving parts), or straightforward tasks where you already know the facts. For a post about something you just did, a single session is faster and probably just as good.

Blog Agent Swarm Example Design

The goal: write a blog post that's factually correct without manual fact-checking. Four agents, each with a specific job.

graph TD
    build[build agent - coordinator] -->|1. gather facts| researcher
    build -->|2. draft from brief| writer
    build -->|3. verify claims| fact-checker
    build -->|4. check style| reviewer
    fact-checker -->|issues found| build
    reviewer -->|issues found| build
    build -->|fix and re-check| writer

How they invoke each other: they don't, directly. The built-in build agent acts as coordinator. It uses the Task tool to spawn each subagent sequentially, passing the output of one as context to the next. The subagents never talk to each other. They return results to build, which decides what to do next.

The flow:

You tell build what post to write
build spawns researcher with the topic. Researcher gathers facts from docs, repos, and the web. Returns a structured brief with sources and confidence levels.
build passes the brief to writer. Writer produces a markdown draft. Every claim must trace back to the brief. Gaps are marked TODO: need source for X.
build passes the draft to fact-checker. Fact-checker independently verifies every claim against primary sources. Returns a report: VERIFIED, INCORRECT, OUTDATED, or UNVERIFIABLE per claim.
build passes the draft + fact-check report to reviewer. Reviewer checks style, structure, and readability only.
If fact-checker or reviewer found issues, build passes the feedback back to writer for a revision, then re-runs fact-checker. Loop until clean.

Why this split matters: the researcher and fact-checker both search for sources, but at different stages. The researcher works from a blank slate. The fact-checker works from a finished draft and catches things the researcher missed or the writer added. Two independent passes catch more errors than one.

First-Draft Agent Specs

.opencode/agents/blog/researcher.md, read-only. Pinned to Gemini 2.5 Flash (free tier) because it has built-in web search (grounding), useful for research tasks:

---
description: Gathers facts from docs, repos, and the web. Read-only.
mode: subagent
model: google/gemini-2.5-flash
tools:
  write: false
  edit: false
  bash: false
---
You are a research agent for blog post writing. Your job is to
gather accurate, sourced information on a given topic.

For each claim or fact you return:
- State the fact clearly
- Cite the source (URL, repo path, or doc reference)
- Note your confidence level (verified, likely, uncertain)

Use web search and file reading tools to find primary sources.
Prefer official docs and source code over blog posts and forums.
Do not speculate. If you can't verify something, say so.

Return a structured research brief the calling agent can hand
to a writer. Group findings by subtopic.

.opencode/agents/blog/writer.md, can write files. Using Big Pickle here since it's free and handles prose fine:

---
description: Drafts blog posts from research briefs. Can write files.
mode: subagent
model: opencode/big-pickle
tools:
  bash: false
---
You are a blog post writer. You receive a research brief and
produce a markdown draft.

Style rules:
- Casual, first-person, working engineer voice
- No em-dashes. Use commas or periods instead.
- Sentences: 8-15 words typical. Fragments fine for emphasis.
- Paragraphs: 1-5 sentences. No walls of text.
- No AI writing tells: no filler affirmations, no hedging
  qualifiers, no summarizing paragraphs, no conclusion sections
- Code blocks should be copy-paste ready

Every factual claim in the draft must come from the research
brief. Do not add facts the researcher did not provide. If the
brief has gaps, note them with "TODO: need source for X" inline.

Write the draft to the file path specified in your task prompt.

.opencode/agents/blog/fact-checker.md, read-only. Also on Gemini 2.5 Flash so it can independently verify claims via web search:

---
description: Verifies claims in a draft against primary sources.
mode: subagent
model: google/gemini-2.5-flash
tools:
  write: false
  edit: false
  bash: false
---
You are a fact-checking agent. You receive a blog post draft and
verify every factual claim in it.

For each claim:
1. Identify the specific assertion
2. Search for the primary source (official docs, source code, API)
3. Mark it as: VERIFIED, INCORRECT, OUTDATED, or UNVERIFIABLE
4. If incorrect or outdated, provide the correct information with
   a source link

Pay special attention to:
- Version numbers and release dates
- API pricing (changes frequently)
- CLI commands and flags
- Feature comparisons between tools
- Config file formats and field names

Return a structured report. List verified claims briefly. Expand
on anything incorrect, outdated, or unverifiable.

.opencode/agents/blog/reviewer.md, read-only, low temperature. Also Big Pickle, style review doesn't need web search:

---
description: Reviews draft for style, structure, and readability.
mode: subagent
model: opencode/big-pickle
temperature: 0.1
tools:
  write: false
  edit: false
  bash: false
---
You are a blog post reviewer. You check drafts for style and
structure, not factual accuracy (that's the fact-checker's job).

Check for:
- AI writing tells: filler affirmations, hollow transitions,
  motivational sign-offs, restating conclusions
- Em-dashes (should be commas or periods instead)
- Walls of text (paragraphs over 5 sentences)
- Generic section headers
- Sentences that are too long or too uniform in length
- Missing code blocks where a command should be copy-pasteable

Return specific, actionable feedback. Quote the problematic text
and suggest a fix. Don't rewrite the whole post.

Locking down the swarm

The build agent can spawn any agent by default. Restrict it to only these four in opencode.json:

{
  "agent": {
    "build": {
      "permission": {
        "task": {
          "*": "deny",
          "researcher": "allow",
          "writer": "allow",
          "fact-checker": "allow",
          "reviewer": "allow"
        }
      }
    }
  }
}

Kicking it off

From your project root:

opencode

Run /connect to set up your providers. You need Zen (free, gives you Big Pickle for writer/reviewer) and Google (free tier Gemini 2.5 Flash for researcher/fact-checker). Make sure you're in build mode, not plan. Hit Tab to toggle.

As a first test, have it write a post about Linear's MCP server. Add the Linear MCP connector to opencode.json:

{
  "mcp": {
    "linear": {
      "type": "remote",
      "url": "https://mcp.linear.app/mcp"
    }
  }
}

Then paste a prompt like:

Write a blog post about using Linear's MCP server for
AI-assisted project management. Follow this pipeline:
1. Spawn researcher to look up Linear's MCP docs and
   changelog for current capabilities. Wait for the brief.
2. Spawn writer with the brief. Write the draft to
   apps/blog/blog/markdown/posts/linear-mcp.md
3. Spawn fact-checker with the draft path.
4. Spawn reviewer with the draft path.
5. If either found issues, spawn writer again with the
   feedback. Re-run fact-checker until clean.

The build agent handles orchestration from there. The researcher gets web search via Gemini and pulls the Linear MCP tool list, changelog, and setup docs. The fact-checker independently verifies whatever the writer produces. You get a post you didn't write, about a tool the agents researched themselves. The result of this test is Linear MCP: From Tab-Switching to Flow.

Kyle Pericak

"It works in my environment"

Categories

Tags