There is a new surface to optimise for, and it does not read citations — it picks tools. When a developer asks Claude Code what library to use, or when the agent autonomously reaches for a dependency mid-task, something decides the answer, and it is not the same something that decides a ChatGPT citation. Claude Code blends training-data priors, live documentation retrieval, callable MCP servers, and project-level instructions. The tool authors who understand that are starting to engineer for it. This is the playbook for getting recommended by Claude Code, sitting under the broader pillar on generative engine optimization, extended here from citation to agent tool-selection.
Key takeaways
- The strongest lever — Agent-native packaging: ship an MCP server and register docs with Context7. This injects current, authoritative information into the model's context at query time, overriding a stale training cutoff.
- The baseline lever — Training-data prevalence. Strong GitHub activity, package-registry downloads, and well-indexed public docs before the cutoff create the priors the agent reaches for by default.
- The override lever — Project CLAUDE.md / AGENTS.md files can name preferred libraries, and that takes immediate precedence — making framework maintainers and platform teams high-value targets.
- The honest caveat — Anthropic publishes no formal ranking spec for tool recommendations, so these tactics are inferred from documented behaviour. Treat them as hypotheses to measure.
How Claude Code chooses what to recommend
Claude Code is Anthropic's official CLI coding agent, running Claude models with a knowledge cutoff in mid-2025. When it recommends or reaches for a tool, four inputs are in play, layered from slow-moving baseline to real-time override.
- Prior — training-data prevalence sets the default, driven by GitHub, package registries, and documentation indexed before the cutoff.
- Retrieve — WebSearch and WebFetch pull current information when the agent needs it.
- Call — MCP servers expose tools and documentation the agent can use directly.
- Read — the project's CLAUDE.md or AGENTS.md can name preferred libraries, taking precedence.
- Recommend — the agent synthesises across these and suggests or uses a tool.
Two implications follow. First, the training cutoff is not destiny: the retrieval and MCP layers let a tool author override a stale or absent prior with current, authoritative context, which is why agent-native packaging is the highest-leverage move. Second, project-level instruction beats everything else for a given codebase, so the path to being a default recommendation often runs through framework maintainers and platform teams who set the CLAUDE.md, not through individual end users.
The playbook
Tactics ordered by leverage, calibrated for Claude Code. The top two are agent-native and specific; the rest are the durable fundamentals of being a well-represented, well-documented tool.
- Ship an MCP server. The Model Context Protocol lets Claude Code call your tool directly as a first-class capability rather than merely knowing about it. A maintained MCP server turns your tool from something the agent might mention into something it can actually use mid-task, which is the strongest form of recommendation there is.
- Register your docs with Context7. Context7 injects current, authoritative documentation into the model's context when a developer asks about your domain, overriding the mid-2025 training cutoff. For any tool whose docs have changed since the cutoff, this is the single most effective way to ensure Claude Code reasons about the current version of you, not a stale snapshot.
- Maintain excellent public docs with an llms.txt. When Claude Code retrieves live, the quality of what it finds decides the quality of the recommendation. A clean, text-friendly documentation site, a clear README, and an llms.txt that indexes your key pages all improve live-retrieval accuracy — and the same assets help the text answer engines, so the work compounds.
- Build strong GitHub and package-registry signal. Training-data prevalence is the baseline prior, and it tracks GitHub activity, npm or PyPI downloads, and how widely your tool is discussed. This is the slow lever — it compounds over releases and cannot be shortcut — but it is what makes you a default suggestion rather than a retrieval-only one.
- Target framework maintainers who set CLAUDE.md defaults. A project's CLAUDE.md or AGENTS.md can name your tool as the preferred choice, and that overrides general priors for everyone using the project. Getting your tool into a popular starter template or framework's agent instructions is a force multiplier no amount of individual outreach matches.
- Use conventional naming and clear capability boundaries. Tools that describe what they do in the vocabulary developers and models already use are easier to retrieve and recommend than cleverly-named ones whose purpose is opaque. Make your tool's job legible from its name, its README's first paragraph, and its MCP server's tool descriptions.
What's different from Codex, Hermes Agent, and OpenClaw
Each coding agent has its own first-class extension format, and that is where the recommendation levers diverge. The shared cross-cutting principles are covered in the related pages below; the per-agent packaging is what changes.
- OpenAI Codex reads AGENTS.md from the repo root rather than centring on MCP, so the lever there is an AGENTS.md and llms.txt with explicit usage guidance. The Codex playbook is at get recommended by Codex.
- Hermes Agent uses a SKILL.md skills system and a community registry, so packaging your tool as a Hermes Skill is the direct channel. The Hermes playbook is at get recommended by Hermes Agent.
- OpenClaw extends through ClawHub skills, also SKILL.md-based, so the lever overlaps with Hermes. The OpenClaw playbook is at get recommended by OpenClaw.
- Claude Code's MCP-plus-Context7 combination is its distinctive pairing, and it happens to be the most mature agent-native packaging path of the four, which is why it leads this list.
Measurement
Agent tool-selection is young and under-instrumented, so the loop is mostly proxies and direct tests. Build it in three layers:
- Direct testing. Ask Claude Code the domain questions your customers would ask and record whether, and how, it recommends you. Re-run after each intervention — shipping an MCP server, registering Context7 — to see the change. This is the most direct signal available.
- MCP call volume and install correlation. If you ship an MCP server, track its call volume; watch installs and sign-ups for the agent-driven discovery pattern of developers arriving already knowing your tool by name.
- Cross-reference the text engines. The same docs-and-llms.txt work that helps Claude Code also helps the answer engines, so an LLM-visibility tracker gives you a related signal. The Radar's shortlist is at 6 GEO Tools the Radar Actually Recommends; CTAIO Labs tested ten in the visibility tools test.
Related reads
Frequently asked questions
How does Claude Code decide which tool to recommend?
It blends four inputs. The largest is training-data prevalence: libraries well-represented across GitHub, package registries, and documentation by its mid-2025 knowledge cutoff form strong default priors. On top of that, Claude Code can retrieve live information through its WebSearch and WebFetch tools, call MCP servers that expose tools and docs directly, and read the project's CLAUDE.md or AGENTS.md, which can name preferred libraries outright. Anthropic does not publish a formal ranking algorithm for recommendations, so this is a description of the documented inputs rather than a disclosed scoring formula.
What is the single highest-leverage thing I can do?
Ship an MCP server for your tool and register your documentation with Context7. An MCP server lets Claude Code call your tool directly as a first-class capability, and Context7 injects your current, authoritative docs into the model's context when a developer asks about your domain. Together they bypass the training-cutoff problem: even if the model's baseline knowledge of you is stale, the live context it receives is current. This is the clearest lever available to a tool author today.
Does my GitHub and package-registry presence still matter?
Yes, as the baseline. Training-data prevalence sets the prior Claude Code starts from, and that prevalence is driven by GitHub stars and activity, npm or PyPI download counts, and how often your tool appears in well-indexed documentation and discussion before the cutoff. Strong public signal makes you a default suggestion; weak signal means you are only recommended when live retrieval or an MCP server surfaces you. Both layers matter — the baseline and the live override.
How do CLAUDE.md and AGENTS.md affect recommendations?
Directly and immediately. When a project includes a CLAUDE.md or AGENTS.md that names a preferred library or tool, Claude Code treats that as instruction and follows it, overriding its general priors for that project. This makes framework maintainers, internal platform teams, and template authors unusually high-value: if your tool is the default in a popular starter's CLAUDE.md, every developer using that starter gets your tool recommended by default.
Does an llms.txt help with Claude Code?
It helps the live-retrieval layer. When Claude Code fetches your documentation through WebFetch, a clean llms.txt and a well-structured README give it accurate, current information to work from, which improves both the quality and the likelihood of a recommendation. It is not a ranking signal in the search-engine sense, but as input to live retrieval it is worth maintaining — and it costs little, since the same file helps several agents and the text answer engines.
How is getting recommended by Claude Code different from ranking in an AI search engine?
The mechanism is tool selection, not citation. A search engine cites a page in an answer; Claude Code chooses a library or tool to use or suggest, often inside an autonomous workflow with no user reading citations. That shifts the levers from on-page GEO tactics to agent-native packaging — MCP servers, Context7, project instructions — and from being quotable to being callable. The shared ground is that excellent public documentation helps with both.
How do I measure whether this is working?
Indirectly, for now. Track installs and traffic that correlate with agent-driven discovery, watch your MCP server's call volume if you ship one, and test directly by asking Claude Code domain questions and seeing whether it recommends you. Because the field is young and Anthropic publishes no recommendation analytics, treat measurement as a set of proxies and direct tests rather than a dashboard, and run interventions as experiments.
How does Claude Code decide which tool to recommend?
It blends four inputs. The largest is training-data prevalence: libraries well-represented across GitHub, package registries, and documentation by its mid-2025 knowledge cutoff form strong default priors. On top of that, Claude Code can retrieve live information through its WebSearch and WebFetch tools, call MCP servers that expose tools and docs directly, and read the project's CLAUDE.md or AGENTS.md, which can name preferred libraries outright. Anthropic does not publish a formal ranking algorithm for recommendations, so this is a description of the documented inputs rather than a disclosed scoring formula.
What is the single highest-leverage thing I can do?
Ship an MCP server for your tool and register your documentation with Context7. An MCP server lets Claude Code call your tool directly as a first-class capability, and Context7 injects your current, authoritative docs into the model's context when a developer asks about your domain. Together they bypass the training-cutoff problem: even if the model's baseline knowledge of you is stale, the live context it receives is current. This is the clearest lever available to a tool author today.
Does my GitHub and package-registry presence still matter?
Yes, as the baseline. Training-data prevalence sets the prior Claude Code starts from, and that prevalence is driven by GitHub stars and activity, npm or PyPI download counts, and how often your tool appears in well-indexed documentation and discussion before the cutoff. Strong public signal makes you a default suggestion; weak signal means you are only recommended when live retrieval or an MCP server surfaces you. Both layers matter — the baseline and the live override.
How do CLAUDE.md and AGENTS.md affect recommendations?
Directly and immediately. When a project includes a CLAUDE.md or AGENTS.md that names a preferred library or tool, Claude Code treats that as instruction and follows it, overriding its general priors for that project. This makes framework maintainers, internal platform teams, and template authors unusually high-value: if your tool is the default in a popular starter's CLAUDE.md, every developer using that starter gets your tool recommended by default.
Does an llms.txt help with Claude Code?
It helps the live-retrieval layer. When Claude Code fetches your documentation through WebFetch, a clean llms.txt and a well-structured README give it accurate, current information to work from, which improves both the quality and the likelihood of a recommendation. It is not a ranking signal in the search-engine sense, but as input to live retrieval it is worth maintaining — and it costs little, since the same file helps several agents and the text answer engines.
How is getting recommended by Claude Code different from ranking in an AI search engine?
The mechanism is tool selection, not citation. A search engine cites a page in an answer; Claude Code chooses a library or tool to use or suggest, often inside an autonomous workflow with no user reading citations. That shifts the levers from on-page GEO tactics to agent-native packaging — MCP servers, Context7, project instructions — and from being quotable to being callable. The shared ground is that excellent public documentation helps with both.
How do I measure whether this is working?
Indirectly, for now. Track installs and traffic that correlate with agent-driven discovery, watch your MCP server's call volume if you ship one, and test directly by asking Claude Code domain questions and seeing whether it recommends you. Because the field is young and Anthropic publishes no recommendation analytics, treat measurement as a set of proxies and direct tests rather than a dashboard, and run interventions as experiments.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.