Key takeaways
- The only published study of agent tool picks covers Claude Code alone. It found a few near-monopoly defaults (GitHub Actions, Stripe, shadcn/ui, Vercel) and a surprising 12% share of "build it yourself" answers.
- I found no equivalent study for Codex or OpenCode. Anything said about their default picks is inference from how they work, and I will mark it as such.
- OpenCode has no model of its own. Its recommendations are whatever the underlying model would say, so the harness matters less there than the model you plug in.
- Instruction files (CLAUDE.md, AGENTS.md) are the strongest lever you have, though practitioners report they are sometimes ignored.
- If you sell developer tools, the practical question is whether models and agents can find and trust your docs, GitHub presence, and review profiles.
Why this question matters now
A few years ago a developer picked a library by reading comparison posts, asking colleagues, or copying whatever the last project used. Now a growing share of that decision happens inside a terminal, when an agent writes npm install on your behalf. Nobody sees a ranked list. The agent just picks.
That is a big deal for two groups. Developers should know what defaults they are inheriting, because an agent that quietly adds a paid feature-flag service or a GitHub Actions workflow can cost real money. And tool vendors should care because the agent is now a distribution channel, one with its own biases.
The frustrating part is how little hard evidence exists. Plenty of people have compared these agents on speed, price, and UX. Almost nobody has measured what they recommend. So this guide separates what is measured, what is reported anecdotally, and what is reasonable inference.
The three agents in one paragraph each
Claude Code is Anthropic's agentic coder. It runs in the terminal, in VS Code and JetBrains, as a desktop app, on the web, and in Slack. It reads your files, runs shell commands, and executes tests. It keeps project memory in CLAUDE.md files and has the deepest extensibility stack of the three: hooks, skills, plugins, subagents, and Agent Teams.
Codex is OpenAI's multi-surface agent: an open-source CLI written largely in Rust, a cloud and web app, IDE extensions, and mobile. It reads AGENTS.md, a community convention that Cursor, Windsurf, and OpenCode also respect. Its standout feature is OS-level sandboxing, which the Tensorlake benchmark author called the strongest of the group.
OpenCode is the MIT-licensed, terminal-first agent maintained by Anomaly (formerly the SST team). It is model-agnostic, with 75+ providers including local models through Ollama. It runs as a client and server, so sessions survive a dropped SSH connection. It also reads .claude/skills, so a Claude Code setup mostly carries over.
If you want the broader feature and pricing comparison, the Medium and HackerNoon comparisons are decent starting points. Here is the first one:

Those articles mostly answer "which agent is better." This guide asks a narrower question: which stack does each one steer you toward?
What the only real study found: Claude Code's defaults
Amplifying ran 2,430 open-ended prompts with no tool names in them, across 20 tool categories, 4 repos, and 3 models (Sonnet 4.5, Opus 4.5, Opus 4.6), three runs each. After discarding unparseable runs, 2,073 primary picks remained. Model agreement was around 90%, meaning 18 of 20 categories landed in the same ecosystem regardless of model. You can read the full write-up at the Amplifying study.
Near-monopolies
A handful of tools were picked almost every time they were relevant:
- GitHub Actions: 93.8% of CI picks (152 of 162)
- Stripe: 91.4% of payments picks (64 of 70)
- shadcn/ui: 90.1% of UI component picks (64 of 71)
- Vercel: 100% of JavaScript deployment picks (86 of 86)
Strong defaults
Tailwind (68.4%), Zustand (64.8%), Sentry (63.1%), Resend (62.7%), Vitest (59.1%), and PostgreSQL (58.4%) all won clear majorities in their categories.
The part that surprised me: build it yourself
The single most common "pick" was custom or DIY code. It accounted for 252 of 2,073 primary picks, 12%, and showed up in 12 of 20 categories. Feature flags went custom 69% of the time, against 21% for LaunchDarkly. Python auth was 100% custom (JWT plus bcrypt). Overall auth was 48% custom.
If you sell a feature-flag or auth product, that stings. Your competitor is not another vendor. It is a hundred lines the agent writes on the spot.
Who gets ignored
Some very popular tools barely register. Redux got 0 of 88 primary picks, though it was mentioned 23 times. Express was absent. Jest got 7 of 171. yarn got 1 of 135. The big clouds (AWS, GCP, Azure) received zero primary deployment picks; Netlify and Cloudflare Pages showed up only as alternatives.
Deployment was stack-determined: Vercel for JavaScript, Railway for Python and FastAPI (82%).
Models differ, sometimes sharply
| Model | Behavior in the study |
|---|---|
| Sonnet 4.5 | Conventional: Prisma for 79% of JS ORM picks, Redis 93% for Python caching, Celery 100% for Python jobs |
| Opus 4.5 | Most likely to name a specific tool (86.7%) |
| Opus 4.6 | Picks Drizzle 100% of the time as a JS ORM, never Prisma; builds custom most often (11.4%) |
So "Claude Code recommends X" really means "this Claude model, in this harness, recommends X." A model upgrade can flip a category overnight.
The study's own caveats
The authors are upfront about limits, and you should be too if you quote these numbers. The picks reflect training data, RLHF, and the system prompt, not developer consensus. Quality cannot be separated from how often a tool appeared in training data. And each prompt ran against an existing codebase, so some picks are tautological: you get FastAPI because the repo already is FastAPI.
What we do not know about Codex and OpenCode
I looked for a comparable study on Codex and OpenCode and did not find one. That is a real gap, not a footnote.
What I can offer is reasoning from architecture, labeled as inference:
- Codex runs OpenAI models (GPT-5.5 is the recommended model in the sources I read). Its picks should reflect OpenAI's training data and post-training, which may overlap heavily with Claude's on mainstream choices (Postgres, Stripe, Tailwind) but could diverge on newer or niche tools. Nobody has measured this.
- OpenCode is aggressively model-agnostic. Point it at Claude and you will probably get Claude-like picks; point it at a DeepSeek or Kimi model or a local Ollama model and you get that model's habits. The harness is thin: roughly 6,900 tokens of prompt and tool schema before your first message, versus a community-measured 33,000 for Claude Code. A thinner system prompt means less harness-level steering.
The Tensorlake author put it neatly: pick the harness first, pick the model second. For tool recommendations I would flip that. The model sets the taste, and the harness decides how much of your own context overrides it.
Where the harness does matter: context and instruction files
What an agent recommends depends heavily on what it reads before it answers. Three things shape that.
Instruction files
Claude Code reads CLAUDE.md, with hierarchical resolution (most specific file wins), @path imports, and an auto-memory system. Codex reads AGENTS.md, which is a shared convention that OpenCode, Cursor, and Windsurf also respect. Builder.io's author gave Codex the win on this point because AGENTS.md is portable. Firecrawl's comparison argues CLAUDE.md is more powerful inside Claude Code.
OpenCode reading both AGENTS.md and .claude/skills is a quiet advantage if you switch tools often.
Existing codebase
The Amplifying study ran inside real repos, and the stack you already have is the strongest signal. An agent in a Next.js repo reaches for Vercel and Tailwind. In a FastAPI repo it reaches for Railway. If you want a different answer, change the context, not the prompt.
Project memory and compaction
Each agent handles long sessions differently. Claude Code auto-compacts. Codex compacts server-side and has a manual /compact. OpenCode never deletes old messages; it marks them as compacted so the model stops seeing them. An early instruction like "use bun" can quietly fall out of context in a long session, which brings us to the failure stories.
What practitioners report on Hacker News
This is anecdotal, so treat it that way. In the discussion of the Amplifying study:
- One user had Claude Code propose NeonDB and Fly.io even though CLAUDE.md specified AWS EC2. Another saw it suggest Neon while they already used PlanetScale.
- Others say explicit preferences work reliably: "always use bun for web projects, uv for python," or specifying Hono with no Tailwind.
- One commenter said they have fought with Claude, Codex, and OpenCode over AGENTS.md or CLAUDE.md rules not being followed.
- One user reported Claude adding GitHub Actions unasked, which cost money. A diff view that hides dotfolders can mask this.
- Some commenters speculate vendors will eventually pay to be recommended. No evidence was offered, and I would not repeat it as fact.
My read: instruction files steer the average case well and fail at the edges. Never treat them as enforcement.
A practical comparison for steering recommendations
| Claude Code | Codex | OpenCode | |
|---|---|---|---|
| Instruction file | CLAUDE.md (hierarchical, imports) | AGENTS.md | AGENTS.md, plus reads .claude/skills |
| Model behind picks | Anthropic models only | OpenAI models | Any of 75+ providers, including local |
| Published pick study | Yes (Amplifying) | None found | None found |
| Hard guardrails | Hooks (26 to 29 lifecycle events depending on source), deny-first permissions | OS-level sandbox | Config via opencode.json |
| System prompt overhead | ~33,000 tokens (community test) | Not measured in my sources | ~6,900 tokens (community test) |
| Best lever for tool choice | CLAUDE.md plus hooks to block unwanted installs | AGENTS.md plus sandbox limits | Choose the model, then AGENTS.md |
The hook count differs between sources (one says 29, another 26), and the default model name conflicts too (Opus 4.7 in one, 4.8 in another). Check Anthropic's docs before quoting either. I also did not verify current pricing for any of the three, so confirm that on each vendor's site.
How to steer what your agent recommends
- Write the stack down. Name your database, host, auth approach, test runner, and package manager in CLAUDE.md or AGENTS.md. Be explicit: "Use Postgres on our existing RDS instance. Do not add new hosted services without asking."
- Ask for a plan first. Commenters recommend telling the agent to ask clarifying questions and propose a plan before touching files. It surfaces a surprise vendor before it is installed.
- Pick a side on build versus buy. Tell it to prefer off-the-shelf libraries, or to write things custom. Without that, you get the 12% DIY wildcard.
- Use hard controls for the costly stuff. In Claude Code, hooks can block certain installs or file writes. In Codex, the sandbox limits what a bad choice can do. Budget caps on any cloud account the agent can reach are worth setting.
- Review dotfiles. Make sure your diff view shows
.github/and similar folders so a surprise workflow cannot hide. - Retest after model upgrades. Opus 4.5 and 4.6 disagreed on JS ORMs. An upgrade can change your defaults without warning.
If you build developer tools: how to get recommended
Here the picture connects to AI search visibility. Coding agents are one more AI surface that recommends tools, alongside ChatGPT, Gemini, and Perplexity. They will not all behave the same way, and I cannot tell you from data how Codex or OpenCode rank vendors. But the inputs that shape any model's opinion are fairly consistent: documentation, GitHub presence, and third-party review profiles.
Promptwatch Data has no report on which libraries coding agents recommend, so I will not pretend otherwise. The closest proxy is its ChatGPT software citations report for Sep 28 to Oct 4, 2026. In it, GitHub is the most cited software domain at 1.2% of citations, followed by Microsoft Learn at 0.98% and G2 at 0.73%. By source type, docs and help centers earn the largest share at 2.64%. That is ChatGPT web citations, not agent behavior, so read it as a rough signal for what a model treats as authoritative on software questions.
The practical takeaways for a vendor are modest and sensible: keep crawlable docs, keep a public GitHub presence, and keep your G2 and Capterra profiles current. It also helps to check whether AI crawlers can actually read your docs. Promptwatch is an AI visibility platform that tracks how ChatGPT, Gemini, Claude, Perplexity, and Google AI Overviews mention and cite a brand, and it also tracks how Codex, Claude Code, and OpenCode recommend tools, SDKs, and docs, which matters if you sell to developers.

Its AI crawler logs show whether bots can read your docs at all, and its free AI visibility report is a no-login way to see how a few major models describe you today. For a wider look at the category, the GEO software directory at bestgeosoftware.com lists alternatives.
Which agent should you use if tool choice matters to you?
- Choose Claude Code if you want the most control over behavior. Hooks and CLAUDE.md let you shape and block choices, and it is the only agent with measured pick data you can plan around.
- Choose Codex if sandboxing and cross-surface continuity matter more. Its OS-level isolation limits the damage from a bad install, and AGENTS.md is portable.
- Choose OpenCode if you want to control the model. Different models have different tastes, and OpenCode lets you test that directly, even running fully local for NDA-bound work.
Many heavy users route work by task: Claude Code for fast interactive iteration, Codex Cloud for unattended runs, OpenCode for cheap or local-model work. If you do that, put the same stack rules in AGENTS.md and let the other files import or mirror it, so every agent hears the same instructions.
The bottom line
Agents recommend tools the way a confident junior engineer does: from what they have seen most, shaped by the repo in front of them. The evidence for Claude Code shows strong defaults, a 12% habit of rolling its own, and real variation between model versions. For Codex and OpenCode, the honest answer is that nobody has published the numbers yet. Write your stack into the instruction file, ask for a plan, and put guardrails on anything that costs money.