Codex, Claude Code, OpenCode, and Cursor all take a task, read your repository, edit files, run commands, and keep going until the job is done. A year ago they looked very different. In October 2026 they share much of the same feature list: subagents, hooks, skills, MCP, and parallel runs in Git worktrees. The differences that remain are about where you work, which models you can use, how the subscription subsidizes your tokens, and how much of the stack you own.
So which AI coding agent should you use? This guide compares all four on the things that change your day: price, models, interfaces, safety controls, extensibility, and the kind of work each one handles best. We checked every price, plan, model name, and feature below against official documentation on October 18, 2026. Pricing in this category changes often, and Cursor changed its entire plan lineup six days before we published, so check the linked pages before you buy.
TL;DR
- Claude Code is the pick for long, autonomous, multi-file work. It runs Anthropic's Opus 5.5 by default with a 1M-token context window, and it has the deepest agent tooling: agent view,
/goal, dynamic workflows, hooks, skills, and a plugin marketplace. It starts at $20 a month with Claude Pro. - OpenAI Codex is the pick for safe defaults and value. The open-source Rust CLI runs inside an operating-system sandbox with network access off by default, and Codex comes with every ChatGPT plan from Plus ($20) up, with limited access on Free and Go.
- OpenCode is the pick if you want open source and model freedom. It is MIT-licensed, works with 75+ model providers and local models, and costs nothing beyond the model you bring. OpenCode Go adds curated open models for $10 a month.
- Cursor is the pick if you want to stay in an editor. You get Tab completion, visual diffs, the Agents Window, and cloud agents. New customers now buy it through The X Subscription from $8 a month, and Claude, GPT, and Gemini models are billed at API list price on Premium ($30) and above.
- Mixing tools is normal. All four speak MCP, and Codex and Claude Code both install as extensions inside Cursor, so many teams run a terminal agent and an editor side by side.
Codex vs Claude Code vs OpenCode vs Cursor at a glance
| OpenAI Codex | Claude Code | OpenCode | Cursor | |
|---|---|---|---|---|
| Made by | OpenAI | Anthropic | Anomaly (open source) | Anysphere |
| Main surfaces | CLI, ChatGPT desktop app, IDE extension, Codex Cloud | CLI, desktop app, VS Code and JetBrains, web, mobile | Terminal UI, desktop app, web UI | Desktop editor, Agents Window, CLI, cloud agents |
| Default model | GPT-6.1 Sol | Claude Opus 5.5 | Whatever you configure | Composer 2.5 and Grok 4.7 in the plan pool |
| Model choice | OpenAI models only | Anthropic models only | 75+ providers plus local models | Own models, plus Claude, GPT, and Gemini at API price |
| Open source | CLI only (Apache-2.0, Rust) | No | Yes (MIT, TypeScript) | No |
| Entry price | $20 (Plus); limited on Free and $8 Go | $20 (Pro) | Free with your own model; Go $10 | Free (limited); Plus $8 |
| Heavy-use tiers | Pro $100, $200, $500 | Max $100, $200 | Go Plus $40, or API pay as you go | Super $100, Ultra $200 |
| Project instructions | AGENTS.md | CLAUDE.md | AGENTS.md | .cursor/rules and AGENTS.md |
| MCP support | Yes | Yes | Yes | Yes |
| Best for | Sandboxed, parallel, cloud-delegated work | Deep autonomous refactors and long tasks | Model flexibility and full control | Interactive editing with a visual IDE |
How these four tools actually differ
Feature lists won't separate these tools anymore. This year Codex added plugins and a goal mode, Claude Code added a multi-session dashboard, Cursor started reading Claude Code's skills and hooks directly, and OpenCode added worktree tools. Three questions separate them now.
Where do you want to work? Claude Code, Codex, and OpenCode are terminal-first agents with desktop and IDE add-ons. Cursor is an editor first, with an agent-first window and a CLI added on top. If you review changes by reading diffs in an editor, Cursor's interface does the most work for you. If you hand off a task, walk away, and review a pull request later, a terminal or cloud agent fits better.
Who chooses the model? Claude Code runs only Anthropic models, and Codex runs only OpenAI models. That is a real constraint, because the model drives much of the quality difference. Cursor includes its own models (Composer and Grok) in the subscription and charges API list price for third-party models. OpenCode has no model of its own: you connect a provider or a local server and it runs whatever you choose.
How much do you get per dollar? Anthropic and OpenAI price their consumer plans well below the API cost of heavy usage, which is why the $100 and $200 tiers exist. Cursor subsidizes its own models inside a weekly pool but passes third-party model costs straight through. OpenCode has no markup on API models, but you pay normal API rates unless you buy Go or sign in with a subscription it supports.
Claude Code
Claude Code is Anthropic's agentic coding tool. It started as a terminal CLI and now runs in the terminal on macOS, Linux, WSL, and Windows, in VS Code (including inside Cursor) and JetBrains, in a desktop app for macOS and Windows, on the web at claude.ai/code, and in the Claude mobile apps. The default model on paid plans is Claude Opus 5.5, and every current Claude model has a 1M-token context window.
Claude Code leads on long autonomous work. Agent view shows every background session on one screen, each in its own Git worktree. /goal keeps Claude working until a separate model confirms that your stated condition is met. Dynamic workflows let Claude write a script that coordinates dozens or hundreds of subagents for migrations and large refactors.
Where it wins
- Reasoning on hard, multi-file changes. Opus 5.5 scored 66.4% on Terminal-Bench 4.0 at xhigh effort, according to Anthropic's launch post, and it tops Cursor's own CursorBench 4.0 leaderboard at 57.8%.
- Extensibility. Subagents, hooks, skills, slash commands, plugins with marketplaces, and MCP are all mature. The ecosystem is large enough that Cursor now loads Claude Code hooks and
.claude/skillsfolders. - Permission modes. You can choose Manual, accept-edits, plan, auto (a classifier reviews each action), and more, plus a built-in Bash sandbox on macOS, Linux, and WSL2. Checkpoints and
/rewindundo file edits the agent made. - Automation surfaces. GitHub Actions, GitLab CI/CD, automated code review, scheduled Routines, and the Claude Agent SDK let you run the same agent outside your laptop.
Where it falls short
- One model vendor. You can switch between Opus, Sonnet, Haiku, and Fable, but you cannot point Claude Code at GPT or Gemini.
- Closed source. The claude-code repository holds the changelog, plugins, and examples, but the CLI is proprietary under Anthropic's commercial terms.
- Usage limits. Every plan has a rolling five-hour window plus weekly limits. Heavy daily use can run into them on Pro, which is what the Max tiers are for.
- Subscription lock-in. Anthropic's terms do not allow third-party tools to route requests through Free, Pro, or Max credentials, so your Claude subscription only works in Anthropic's own apps.
Pricing. Claude Code is included in Pro at $20 a month ($17 a month billed annually), Max 5x at $100, and Max 20x at $200. Team Standard seats cost $25 a month ($20 annually) and Team Premium seats $125 ($100 annually). Enterprise is $20 per seat billed annually plus usage at API rates. With an API key you pay per token: Opus 5.5 is $4 per million input tokens and $20 per million output tokens.
Choose Claude Code if you hand off big, ambiguous tasks and want the agent to plan, execute, test, and fix its own mistakes with minimal check-ins.
OpenAI Codex
Codex is OpenAI's coding agent, and it now spans several products. The Codex CLI is open source under Apache-2.0 and written mostly in Rust. The IDE extension runs in VS Code, Cursor, and Windsurf. The former standalone Codex app merged into the ChatGPT desktop app in July 2026, and a Linux preview followed in August. Codex Cloud runs tasks in OpenAI's environments, so you can start work from the web, from your phone, from Slack, or from a GitHub comment.
The default model is GPT-6.1 Sol, which OpenAI says matches GPT-6 Astra on DeepSWE at about a fifth of the cost. GPT-6 Astra is the most capable option, and GPT-6 Luna is the fast, cheap tier.
Where it wins
- Sandboxing by default. Codex runs commands inside an operating-system sandbox: Seatbelt on macOS, bubblewrap on Linux and WSL2, and a native sandbox on Windows. Network access is off unless you allow it. An Auto-review preset lets a reviewer agent approve escalations, so you get fewer prompts without giving up control.
- Parallel and delegated work. Subagents are on by default, the "Ultra" effort level fans a task out across subagents, and managed Git worktrees keep parallel runs isolated. Cloud tasks keep running after you close your laptop.
- Open-source CLI. You can read and audit the CLI, though the repository does not accept external pull requests.
- Code review. Codex reviews pull requests on GitHub (generally available) and GitLab (beta), and DevDay 2026 added cross-project and automatic reviews.
- Easy switching.
/importbrings settings over from Claude Code and Cursor.
Where it falls short
- One model vendor. Like Claude Code, Codex runs only OpenAI models.
- Fast-changing product surface. The app merged into ChatGPT, the docs moved to learn.chatgpt.com, and
codex mcp-serverwas removed in September. Expect to relearn parts of the setup. - Usage estimates, not guarantees. On Plus, OpenAI estimates 15 to 160 local messages per five hours with GPT-6.1 Sol, depending on task size, and weekly limits can apply.
Pricing. Codex is included in every ChatGPT plan. Free and Go ($8 a month) get GPT-6 Luna in the desktop app and no Codex Cloud. Plus at $20 a month unlocks the CLI, IDE extension, cloud tasks, code review, and the full model lineup. Pro comes in $100, $200, and $500 tiers with no five-hour limit, and only the $500 tier includes Ultrafast mode. Business costs $20 per user per month billed annually or $25 monthly. You can also run the CLI with an API key: GPT-6.1 Sol is $2 per million input tokens and $10 per million output tokens.
Choose Codex if you want strong defaults around safety, you like delegating several tasks in parallel, or you already pay for ChatGPT.
OpenCode
OpenCode is an open-source coding agent maintained by Anomaly (the repository moved there from sst/opencode). It is MIT-licensed, written in TypeScript, and had over 216,000 GitHub stars on the day we checked. It runs as a terminal UI, a desktop app, and a web UI, all connected to one background server per user.
OpenCode has no model of its own. You connect any of 75+ providers, run local models through Ollama, LM Studio, or vLLM, or sign in with a GitHub Copilot, ChatGPT Plus or Pro, GitLab Duo, or SuperGrok subscription.
Two versions are current. OpenCode 2 shipped on September 11, 2026 as @opencode/cli, and OpenCode 1 still receives updates as opencode-ai. Version 2 rebuilt the server, plugin API, and configuration, and some well-known V1 features did not carry over. V2 does not run language servers, ignores CLAUDE.md in favor of AGENTS.md, and removed session sharing in v2.0.30. Most comparison posts you will find still describe V1.
Where it wins
- Model freedom. You can use a frontier model for planning, a cheap open model for routine edits, and a local model for private code, all in the same tool.
- Open source and inspectable. You can audit, fork, and self-host the whole agent, which matters in regulated environments.
- Cost control. The tool is free, the Console's pay-as-you-go models have no markup, and OpenCode Go bundles open models such as GLM, Kimi, Qwen, DeepSeek, and MiniMax for $10 a month (Go Plus is $40).
- A familiar agent model. Build and Plan primary agents, General and Explore subagents, custom agents in Markdown, skills, plugins with hooks, custom commands, ordered allow, ask, and deny permission rules, and undo snapshots.
- Editor and CI reach.
opencode acpconnects to Agent Client Protocol editors such as Zed, and V1 documents a GitHub integration that responds to/opencodein issues and pull requests.
Where it falls short
- No Claude subscription. OpenCode removed its Claude Pro and Max login in v1.3.0 because Anthropic prohibits it. You can still use Claude through an API key, but at API prices.
- Quality depends on your choices. OpenCode is only as good as the model you connect and the settings you pick. Teams that want a tuned, opinionated default may prefer Claude Code or Codex.
- Version churn. The V1 to V2 migration broke the plugin and server APIs, and a few documented features differ between versions.
- Looser default in V2. V2 allows access to directories outside the project by default, while V1 asked first. Review the permission rules before you point it at a sensitive machine.
Pricing. The software is free. You pay your model provider, a Console balance (pay-as-you-go top-ups start at $20), or Go at $10 or $40 a month.
Choose OpenCode if you want to switch models freely, run local or open models, keep everything open source, or reuse a Copilot or ChatGPT subscription in a terminal agent.
Cursor
Cursor is Anysphere's AI code editor. It keeps a classic editor view that supports VS Code extensions and adds the Agents Window from Cursor 3 (April 2026) for running several agents in parallel, each in its own worktree. It also has a CLI (the command is agent), cloud agents you can launch from the web, iOS, Slack, GitHub, or Linear, and JetBrains support through ACP.
Cursor's business changed this year. SpaceX agreed to buy Anysphere, Cursor's maker, in June 2026. Since October 12, new customers buy Cursor through The X Subscription, which bundles Cursor, Grok, and X in one plan. Cursor's included models are now Composer 2.5 and Grok 4.5 through 4.7. Claude, GPT, and Gemini models remain available but sit outside the included pool.
Where it wins
- Interactive editing. Tab completion, inline edits, and visual diffs make Cursor the fastest of the four for work where you guide each change and want to watch every edit.
- Parallel agents with a UI. The Agents Window shows local and cloud agents side by side, with worktrees, Plan Mode, a browser tool, and Design Mode.
- Compatibility. Cursor reads
.cursor/rules, nested AGENTS.md files, the Agent Skills standard,.claude/skills,.codex/skills, and Claude Code hooks, so a team can share one configuration across tools. - Code review. Bugbot reviews pull requests on GitHub, GitLab, Azure DevOps, Bitbucket, and Cursor's own code host, Origin, with learned rules and Autofix.
- Model breadth. It is the only tool here that offers its own models, Claude, GPT, and Gemini in one interface.
Where it falls short
- Third-party models cost extra. On the new plans, Claude, GPT, and Gemini are billed at API list price from prepaid credit, and only Premium ($30) and above can use them. If Opus 5.5 is your daily driver, a Claude Max plan is likely to give you more Opus usage per dollar than paying API prices through Cursor.
- Plan gating. The $8 Plus plan has no Auto mode, cloud agents, automations, or Bugbot, and runs models at fixed settings.
- Uncertainty. A new owner and a new plan structure in the same quarter means pricing and the model lineup may keep moving. Existing customers keep their legacy Pro ($20), Pro+ ($60), and Ultra ($200) plans for now.
Pricing. On The X Subscription: Free, Plus at $8, Premium at $30 (4x the Plus usage), Super at $100 (15x), and Ultra at $200 (50x). Each paid plan includes a weekly usage pool for Cursor's own models, shared with the Grok apps. Teams seats cost $30, $100, or $200 per seat, with a $0.25 per million token Cursor Token Rate on third-party models. Enterprise pricing is custom.
Choose Cursor if you want an editor-first workflow with the best interactive UX, or a team needs one tool that covers completion, agents, and code review.
Pricing compared (October 2026)
| Plan level | OpenAI Codex | Claude Code | OpenCode | Cursor |
|---|---|---|---|---|
| Free | ChatGPT Free (Luna, desktop app only) | Not included | Free tool, free promo models, local models | Free (limited Agent and Chat) |
| Entry | Go $8 (limited); Plus $20 | Pro $20 | Go $10 | Plus $8 (own models only) |
| Mid | Pro $100 | Max 5x $100 | Go Plus $40 | Premium $30; Super $100 |
| Heavy | Pro $200 or $500 | Max 20x $200 | API pay as you go | Ultra $200 |
| Teams | Business $20 to $25 per user | Team $25 or $125 per seat | Console workspaces | $30, $100, or $200 per seat |
| API option | GPT-6.1 Sol $2 in / $10 out per M tokens | Opus 5.5 $4 in / $20 out per M tokens | Any provider at its rates | Third-party models at API list price |
Three pricing patterns matter more than the headline numbers.
$20 is the real starting point for serious use. Codex Plus and Claude Pro both cost $20 and include a capable frontier model. Cursor Plus at $8 is cheaper, but it only covers Cursor's own models. OpenCode Go at $10 is the cheapest way to get substantial agent usage, provided you are happy with open models.
Heavy users pay $100 to $200. If you run an agent for most of the workday, the five-hour and weekly limits on $20 plans will slow you down. Claude Max and ChatGPT Pro remove most of that friction, and Pro's $100 tier has no five-hour limit at all. Cursor's Super and Ultra tiers scale the pool for Cursor models but not for third-party models, which stay at API prices.
Subscriptions beat the API for daily use. API pricing makes sense for CI jobs, automation, and spiky usage. For an engineer using an agent all day, a consumer plan usually works out cheaper than paying per token.
Head-to-head matchups
Codex vs Claude Code
This is the closest matchup, and it is also the most searched one. Both are terminal-first agents with CLI, desktop, IDE, web, and mobile surfaces, both cost $20 to start, and both offer subagents, hooks, skills, plugins, MCP, worktrees, and goal modes.
The difference is in emphasis. Claude Code puts more weight on reasoning depth and supervised autonomy. Opus 5.5 posts higher vendor-reported Terminal-Bench 4.0 scores than GPT-6 Astra and leads CursorBench, and features like dynamic workflows and agent view suit long, coordinated tasks. Codex puts more weight on safety and parallelism. Its OS-level sandbox with network off by default is the strictest default here, cloud tasks run without your laptop, and the CLI is open source.
Pick Claude Code for hard refactors, debugging sessions where reasoning quality decides the outcome, and large migrations. Pick Codex if you run many small tasks in parallel, want sandboxing without configuration, or already pay for ChatGPT. Plenty of developers keep both, using one to write and the other to review.
Claude Code vs Cursor
Claude Code is an agent you delegate to, and Cursor is an editor you work in. If you spend most of your time typing code and want AI help at every keystroke, Cursor's Tab completion and inline edits make it the better daily driver. If you mostly describe tasks and review the result, Claude Code's autonomy and tooling go further.
Cost is the other factor. Running Opus 5.5 inside Cursor now means paying API list price on Premium or above, while Claude Pro and Max include Opus usage at a much lower effective rate. You can also install the Claude Code extension inside Cursor and get Cursor's editor with Claude Code's agent and your Claude subscription.
OpenCode vs Claude Code
OpenCode feels a lot like Claude Code in the terminal: a TUI, plan and build agents, subagents, skills, hooks, and MCP. The difference is control. OpenCode is open source and works with any model, while Claude Code is closed and works only with Claude.
There is one practical catch. You cannot sign in to OpenCode with a Claude Pro or Max subscription, so using Claude models in OpenCode means paying API rates. If Claude is your main model, Claude Code is the cheaper way to use it. If you want to mix models, use open or local models, or avoid depending on one vendor, OpenCode is the better choice.
Codex vs Cursor
Codex is built for delegation: background cloud tasks, sandboxed local runs, and pull request reviews. Cursor is built for interaction: you watch the agent work in an editor and steer it as it goes. Both handle parallel agents, code review (Codex review vs Bugbot), and mobile handoff. Since the Codex IDE extension runs inside Cursor, you can use both in one window. Choose by where you want the work to happen: in the cloud or the terminal (Codex), or in the editor (Cursor).
OpenCode vs Cursor and Codex vs OpenCode
OpenCode is the open alternative to both. Compared to Cursor, it gives up Tab completion and a polished editor in exchange for open source, model freedom, and lower costs. Compared to Codex, it gives up OS-level sandboxing defaults and a tuned model in exchange for provider choice. Notably, OpenCode can sign in with a ChatGPT Plus or Pro subscription, so you can use OpenAI models in an open-source harness.
What the benchmarks say, and what they don't
Every vendor publishes benchmark numbers, and almost none of them are directly comparable. Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 in its launch post and 89.9% on SWE-bench Pro in its system card. OpenAI reports GPT-6 Astra at 57.9% on Terminal-Bench 4.0 and 74.1% on DeepSWE v1.1. Cursor publishes the CursorBench 4.0 leaderboard, where Opus 5.5 at max effort leads at 57.8%, ahead of Fable 5.1 (51.8%) and Cursor's own Grok 4.7 (46.3%).
Use these numbers with three caveats:
- They measure models, not tools. The harness matters. The same model can score differently in Claude Code, Cursor, and OpenCode because the prompts, tools, and context handling differ.
- Effort settings vary. Vendors often report their best effort level. In daily use you may run lower settings to stay under your limits.
- Benchmarks lag the products. SWE-bench Verified, the standard benchmark a year ago, no longer appears in Anthropic's or OpenAI's current launch materials.
The most useful benchmark is your own. Take three real tickets from your backlog, such as a bug fix, a feature, and a refactor, and run them through two tools. Compare the diffs, the number of times you had to step in, and what it cost. An afternoon of testing will tell you more than any leaderboard.
Which AI coding agent should you choose?
| If you... | Choose |
|---|---|
| Hand off large refactors, migrations, or long autonomous tasks | Claude Code |
| Want sandboxed execution and parallel cloud tasks by default | Codex |
| Already pay for ChatGPT Plus or Pro | Codex (or OpenCode with your ChatGPT login) |
| Want open source, local models, or any provider | OpenCode |
| Have the tightest budget | OpenCode Go at $10, or Cursor Plus at $8 for Cursor's own models |
| Prefer to code in an editor with AI at every keystroke | Cursor |
| Lead a team that needs one tool for completion, agents, and PR review | Cursor or Codex |
| Want the best model inside the best editor | Cursor plus the Claude Code or Codex extension |
For most professional developers in October 2026, the strongest setup is a terminal agent for delegated work (Claude Code or Codex), with an editor for the work you want to drive yourself. OpenCode is the right answer when owning the stack and choosing the model matter more than having the most polished default.
FAQs
Is Codex better than Claude Code?
Neither wins outright. Claude Code with Opus 5.5 posts higher vendor-reported Terminal-Bench 4.0 scores, tops Cursor's CursorBench leaderboard, and is stronger on long, complex tasks. Codex has stricter sandboxing by default, an open-source CLI, and stronger cloud delegation. Both start at $20 a month, so the best way to decide is to test both on your own codebase.
Can I use my Claude Pro or Max subscription in OpenCode or Cursor?
No. Anthropic's terms do not allow third-party tools to route requests through Free, Pro, or Max plan credentials, and OpenCode removed that login in March 2026. You can use Claude models in OpenCode through an API key, and in Cursor at API list price on Premium and above. You can also run the Claude Code extension inside Cursor with your subscription.
Is OpenCode really free?
The software is free and MIT-licensed. You still pay for the model unless you run one locally or use one of the free promotional models. OpenCode Go costs $10 a month for curated open models, and OpenCode itself supports signing in with GitHub Copilot, ChatGPT Plus or Pro, GitLab Duo, or SuperGrok.
Does Cursor still include Claude and GPT models?
Cursor still offers Claude, GPT, and Gemini models, but on the plans launched in October 2026 they are not part of the included usage. Premium, Super, and Ultra subscribers pay API list price for them from prepaid credit, and the Free and Plus plans cannot use them. Existing customers on legacy Pro, Pro+, or Ultra plans keep their previous terms.
Which AI coding agent is cheapest for heavy daily use?
For frontier models, Claude Max and ChatGPT Pro at $100 to $200 a month give the most usage per dollar for full-time use. For open models, OpenCode Go at $10 or Go Plus at $40 is far cheaper. Paying API rates is usually the most expensive way to run an agent all day.
Can I use more than one of these tools at once?
Yes, and many teams do. Codex and Claude Code both install as extensions in Cursor, all four read some form of project instruction file, and Cursor can load Claude Code skills and hooks. A common split is Cursor for interactive work, a terminal agent for delegated tasks, and one agent reviewing the other's pull requests.
Do all four support MCP?
Yes. Codex, Claude Code, OpenCode, and Cursor all connect to remote MCP servers with OAuth, so one MCP server can give every agent on your team the same tools.
Give your coding agent live web access with Context.dev
All four tools share one weakness: they are only as current as the context you give them. Agents read docs, changelogs, API references, and issue threads constantly, and the built-in web tools often return either too little or far too much.
We measured this when we launched Context Highlights. On 250 real URL-and-question pairs, Claude Code's built-in WebFetch kept enough evidence to fully answer the question 50.4% of the time. A single native page open in Codex reached 57.6% but returned an average of 6,889 tokens per page. Context Highlights reached 76.0% groundedness with an average of 229 tokens. That means more accurate answers and far less context spent on navigation menus and cookie banners.
Context.dev gives any of these agents web search, JavaScript-rendered scraping to clean Markdown, site maps, crawls, and schema-shaped JSON extraction through one hosted MCP server. Setup takes one command or config entry per tool:
# Claude Code
claude mcp add --transport http --scope user context https://mcp.context.dev/mcp
# Codex
codex mcp add context --url https://mcp.context.dev/mcpFor Cursor, add the server to ~/.cursor/mcp.json or install the Context.dev Cursor plugin:
{
"mcpServers": {
"context": { "url": "https://mcp.context.dev/mcp" }
}
}For OpenCode 2, add it under mcp.servers in opencode.jsonc, then run opencode mcp auth context to sign in:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"servers": {
"context": { "type": "remote", "url": "https://mcp.context.dev/mcp" }
}
}
}Each client handles sign-in through OAuth, so you never paste an API key into a config file. To make Context.dev the default for every search and page read, follow the default search and scrape guide, which shows how to turn off the built-in web tools in Claude Code and Codex and add routing instructions to CLAUDE.md or AGENTS.md.
Teams building their own coding agents use the same API. Aldena gave every one of its AI coding agents unified web search and fetch in 30 minutes, and freecode cut modeled agent costs by 8.9% after switching to Highlights. Whichever agent you choose, create a free Context.dev account to get 1,000 free credits a month (no credit card required), and give it the same reliable view of the web.