Key takeaways
- Claude Code is the only one of the three with a published study of its tool picks. In a Feb 2026 Amplifying study of 2,430 open-ended prompts, "build it yourself" was the single most common answer (12%).
- Where Claude Code does pick a third-party tool, it converges fast: GitHub Actions took 94% of CI/CD, Stripe 91% of payments, shadcn/ui 90% of UI components.
- Codex and OpenCode have no equivalent study. OpenCode is model-agnostic, so its picks probably track whichever model you plug in. That part is inference, not data.
- Harness differences (instruction files, system prompt size, web search use) shape recommendations as much as the model does.
- If you make a developer tool, public docs, a clear GitHub presence and sensible defaults matter more than any single marketing page.
Why this question matters
A year ago a developer picked a database by reading blog posts, asking colleagues, or copying whatever the last project used. Now a growing share of new projects start with a prompt like "add user authentication" or "how do I deploy this?" and the agent just chooses. If it doesn't pick your product, you never enter the conversation.
The Amplifying researchers call the agent a new gatekeeper. I think that's fair, with one caveat: it's a gatekeeper that is mostly invisible to the vendors it gates. Nobody sends you a notification when an agent skips your SDK.
So how do the three big terminal agents actually decide? The honest answer is that we know a fair amount about one of them and very little about the other two. Let me lay out what is sourced and what is guesswork.
The three agents in short
Before getting to recommendations, a quick orientation. These tools are converging, but they differ in who controls the model and how much scaffolding sits around it.
| Claude Code | Codex | OpenCode | |
|---|---|---|---|
| Maker | Anthropic | OpenAI | Anomaly (open source, MIT) |
| Models | Claude only | OpenAI models (GPT-5.5 recommended per one review) | 75+ providers, including local via Ollama |
| Repo instructions | CLAUDE.md | AGENTS.md | AGENTS.md and project rules; can read .claude/skills |
| Extensibility | Hooks, Skills, plugins, subagents, MCP | MCP, AGENTS.md, kernel-level sandbox | Single opencode.json, subagents, MCP |
| Rough system prompt overhead | ~33,000 tokens (community test) | Not measured in our sources | ~6,900 tokens (community test) |
| Tool recommendation data | Yes (Amplifying study) | None found | None found |
One wrinkle: the comparison articles disagree on Claude Code's default model. A May 2026 Medium piece says Sonnet 4.6 on Pro and Opus 4.7 on Max, while a July 2026 HackerNoon piece says Opus 4.8. Check Anthropic's current docs before you quote a model name.

What Claude Code actually recommends
The best evidence here is the Amplifying study (v1.0, February 2026). The design is worth understanding because it limits what you can conclude.
- 2,430 open-ended prompts with no tool names in them, such as "i need a database, what should i use".
- 20 tool categories across 4 greenfield repos: a Next.js 14 SaaS, a FastAPI API, a Vite/React SPA and a Node CLI.
- Three models (Sonnet 4.5, Opus 4.5, Opus 4.6), three runs each, on Claude Code CLI v2.1.39.
- LLM subagents extracted the picks. 2,073 responses had an identifiable primary tool, an 85.3% extraction rate.
- The repo was reset between prompts so one answer couldn't contaminate the next.
Agents build, not buy
The headline finding surprised me. The most common primary pick wasn't Vercel or Stripe. It was custom/DIY, at 12% (252 of 2,073 responses), and it was the leading outcome in 12 of 20 categories. Feature flags are a good example: LaunchDarkly got 21% against 69% for rolling your own.
Response time correlated with this. The average was 88 seconds, ranging from 32 seconds for deployment to 245 seconds for auth, and the slower categories had higher build-it-yourself rates. Auth is the category where the agent will happily write you a login system from scratch.
The default stack
When it does pick a third-party tool, a pattern shows up: Vercel, PostgreSQL, Stripe, Tailwind CSS, shadcn/ui, pnpm, GitHub Actions, Sentry, Resend and Zustand. Stack-specific picks follow the ecosystem: Drizzle for JavaScript ORMs, SQLModel for Python, NextAuth.js for Next.js auth, and Vitest or pytest for testing.
Some categories are close to locked up:
| Category | Top pick | Share |
|---|---|---|
| CI/CD | GitHub Actions | 94% |
| Payments | Stripe | 91% |
| UI components | shadcn/ui | 90% |
| Package manager | npm | 23% primary |
| Feature flags | LaunchDarkly | 21% (vs 69% custom) |
And some well-known names barely register. Redux had zero primary picks (2 alternates, 23 mentions). Express was absent from both primary and alternate recommendations. Prisma got zero picks from Opus 4.6.
Context beats phrasing
The study found that what you're building matters more than how you ask. Vercel was the pick for Next.js, Railway for Python. Across five phrasings of the same question, picks were 76% stable. The three models also agreed on the top tool in 18 of 20 categories within an ecosystem, roughly 90% consensus. Caching and real-time were the real disagreements.
Practical read: rewording your prompt rarely changes the answer, but changing the repo does.
What the authors caution about
They are upfront about limits. It's a revealed-preference study of Claude Code only. It can't separate quality signals from training frequency. Picks reflect training data, RLHF and the system prompt, not tool quality. So "Claude Code picks Stripe" doesn't mean Stripe is the best payment tool. It means Stripe is what the model has seen most and been rewarded for suggesting.
What we don't know about Codex
I looked for a Codex equivalent and didn't find one. No one has published a prompt-by-prompt study of what Codex picks for a database or an auth library. Anyone telling you otherwise is guessing.
What we can say comes from the harness, and it's softer evidence:
- Codex reads AGENTS.md, a community-defined convention. If your repo says "use Drizzle and Vitest", it will follow that.
- It has a kernel-level sandbox and runs across several surfaces (CLI, cloud, IDE, mobile), so a lot of its work is unattended. Tool choices made in a cloud task that runs for hours get less human review.
- Reviewers describe it as reasoning longer before editing, while Claude Code commits to an action quickly and iterates. Longer reasoning might mean more considered picks. Nobody has measured it.
That last point is speculation on my part. Treat it as a hypothesis to test, not a finding.
What we don't know about OpenCode
OpenCode is the most interesting case because the harness barely decides anything on its own. It supports 75+ providers and runs Claude, GPT, Gemini, DeepSeek, Kimi or local models through Ollama. Its system prompt overhead is small, about 6,900 tokens before your first message versus roughly 33,000 for Claude Code in one community test.
My inference, not a sourced finding: with a thin harness, OpenCode's recommendations mostly reflect the model underneath. Point it at Claude and you'd probably see Claude-like picks. Point it at a local model and you get whatever that model learned. Two OpenCode users can get different default stacks from the same prompt.
This also means the "agents build, not buy" result may not carry over cleanly. A small local model might rarely build auth from scratch and just name-drop popular libraries.
Why harness and instruction files matter
Three harness features shape tool choice more than most people realize.
Instruction files
CLAUDE.md is hierarchical, supports @path imports and an auto-memory that writes back to project files. AGENTS.md is the open convention that Codex, OpenCode, Cursor and Windsurf respect. I didn't verify how either file steers tool choice in testing, but the mechanism is obvious: a single line like "we use Clerk for auth, never roll your own" overrides the model's default. If you maintain a starter template, put your stack decisions there.
# AGENTS.md
- Database: PostgreSQL via Drizzle
- Auth: use the existing auth module, do not add a new library
- Tests: Vitest
- Deploy: Railway
System prompt and scaffolding
Anthropic trains the model and harness together, and a community write-up notes that more than 80% of Claude Code's system prompt was cut for the Claude 5 generation as models improved. Less scaffolding means more of the decision comes from the model's own priors, which is the thing the Amplifying study measured.
Web search and docs
When an agent can browse, it also draws on public sources. Promptwatch Data tracks what ChatGPT's web search cites for software prompts, which is a different thing from what a coding agent picks, but it's a useful side signal. In Promptwatch's October 2026 software citation report (Sep 28 to Oct 4, 2026), GitHub was the most cited software domain at 1.2% of citations, followed by learn.microsoft.com at 0.98% and G2 at 0.73%. Docs and help centers earned the largest share by source type at 2.64%, and vendor sites plus docs together earned about 3x what review sites and analysts did. That report doesn't measure Claude Code, Codex or OpenCode, so don't read it as their picks. It does suggest that crawlable docs and a public GitHub presence are the sources an agent is most likely to find.

Which agent for which job
Since recommendation behavior is only partly documented, the practical choice comes down to workflow. The reviews broadly agree on this:
| Need | Better fit |
|---|---|
| Fast interactive loop, deep customization (29 hook events, Skills, plugins) | Claude Code |
| Long unattended backend work, strongest sandboxing, lower entry price | Codex |
| Any model, local-only for NDA work, open source | OpenCode |
Benchmark numbers cited in the Medium comparison: Claude Code (Opus 4.7) led SWE-bench Pro at 64.3% vs 58.6%, while Codex (GPT-5.5) led SWE-bench Verified (88.7% vs 87.6%) and Terminal-Bench 2.0 (82.7% vs 69.4%). Those test coding ability, not recommendation quality, so they won't tell you which agent picks a better auth library.
A common pattern among heavy users is routing: Claude Code for interactive work, Codex Cloud for overnight runs, OpenCode for cheap or local-model tasks. If you do that, expect your project's default stack to depend on whichever agent happened to scaffold it first. Pin your choices in AGENTS.md early.

What tool makers should do
If you sell a developer product, you're now optimizing for two audiences. Some practical steps, based on the evidence above rather than any magic trick:
- Make the default path work with no configuration. Agents favor tools that slot into a greenfield repo with a one-line install.
- Publish clear, crawlable docs and keep a public GitHub repo with a good README. Those are the sources web-enabled agents reach for.
- Ship framework-specific guides. The Amplifying study showed picks depend heavily on ecosystem, so being the obvious choice for Next.js or FastAPI beats being generically good.
- Compete with DIY. In 12 of 20 categories the top outcome was the agent writing it itself. Your real rival may be a 40-line function.
- Publish an AGENTS.md or Skill that users can drop into their repo so their agent uses your product correctly.
- Measure it. Run your own open-ended prompts against each agent, with no tool names, and track picks over time.
Tracking how AI systems mention you is a solved problem for chat answers. Promptwatch monitors how ChatGPT, Gemini, Perplexity and others cite your brand, and it also tracks how the coding agents Codex, Claude Code and OpenCode recommend tools, SDKs and docs, which is relevant if you ship an API or SDK. Its AI crawler logs show whether AI bots can read your docs at all.

For a wider look at tools in this space, the GEO software directory lists alternatives.
Caveats worth keeping in mind
- The Amplifying study is a single snapshot from February 2026 on older model versions (Sonnet 4.5, Opus 4.5, Opus 4.6). Picks drift as models update.
- Extraction used LLM subagents, and about 15% of responses had no identifiable primary tool.
- Greenfield repos aren't representative of existing codebases, where the agent mostly follows what's already installed.
- Our research did not re-verify pricing, and we didn't test how CLAUDE.md or AGENTS.md steer choices.
Bottom line
Claude Code tends to build things itself and, when it does buy, reaches for a tight default stack: Vercel, Postgres, Stripe, Tailwind, shadcn/ui, GitHub Actions. Codex and OpenCode are open questions, and OpenCode's answers likely depend on the model you connect. Until someone runs the same 2,430-prompt study on them, the safest move is to write your stack into AGENTS.md or CLAUDE.md and stop relying on defaults.