Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/berntpopp/hpo-link/agents-mdgit clone --depth 1 https://github.com/berntpopp/hpo-linkWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01893 | $0.01893 |
| Opus 5 | $0.00946 | $0.00946 |
| Sonnet 5 | $0.00379 | $0.00379 |
| Haiku 4.5 | $0.00189 | $0.00189 |
Grade A, and why
hpo-link AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 145 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md — hpo-link
Guidance for agents and contributors working in this repository.
What this is
hpo-link is an MCP + REST server that grounds phenotype work in the Human
Phenotype Ontology (HPO). It builds a local SQLite database from the HPO OBO
release and HPOA gene/disease annotation files, then serves read-only tools
for phenotype term lookup, the is_a hierarchy, cross-ontology mapping, and
gene↔phenotype↔disease association queries. It mirrors the sibling mgi-link
stack/architecture.
Two planes (non-negotiable boundary)
- Data plane —
config.py,constants.py,identifiers.py,ingest/,data/,services/. Downloads the HPO release (conditional GET), atomically builds the SQLite database (terms, labels, synonyms, definitions,is_aclosure, cross-references with provenance + predicate, deprecated/replaced_by, HPOA gene and disease annotations), and returns plain dicts. It raises typed exceptions fromhpo_link.exceptions; it never builds error envelopes. - MCP plane —
mcp/. Domain-agnostic scaffolding shared with siblings.run_mcp_tool(inmcp/envelope.py) ownssuccess/_metaand converts exceptions into returned structured errors (never raised to the client).
Architecture: ingest → data DAO → services → MCP
ingest/ download (conditional GET) → lock → parser → builder (schema.sql)
data/ HpoRepository (read-only SQLite DAO)
services/ HpoService, AnnotationService, shaping, pagination, refresh
mcp/ envelope, capabilities, annotations, schemas, next_commands,
metrics, middleware, facade, arg_help, resources, tools/
server.py unified REST + MCP transport (FastAPI + uvicorn)
mcp_server.py stdio transport (Claude Desktop)
Invariants
- Services return plain dicts; the envelope owns
success/_metaand returns structured errors. 7-code error taxonomy:invalid_input,not_found,ambiguous_query,data_unavailable,rate_limited,upstream_unavailable,internal_error. - Every
compact(default) or richer response carries_meta.next_commands(ready-to-call follow-ups);minimalis the explicit opt-out and returns only_meta = {tool, request_id}._metaverbosity is tiered byresponse_mode(_shape_meta):compactkeepsnext_commands+capabilities_versionbut drops theelapsed_msecho;standard/fulladdelapsed_ms. - Every tool declares
output_schema+READ_ONLY_OPEN_WORLDannotations, and its first description sentence is a discovery summary ending withSignature: tool(args...). - Every tool's real output (success + error, all response modes) must validate
against its own
output_schema— enforced bytests/unit/test_output_schemas.py. Grouped-by-prefix payloads (xrefs,mappings) are objects keyed by prefix, not arrays; declare them as objects or the envelope leaks a raw validation error. response_mode∈minimal | compact | standard | full. List tools also carry a pagination block (total/returned/limit/offset/truncated/next_offset); when truncated,_meta.next_commandsoffers a forward-page step (advanceoffset).compact+_metaechoescapabilities_version(a hash of the discovery contract) so warm clients can skip re-fetchingget_server_capabilities(omitted inminimal).- Keep
mcp/capabilities.py::TOOLSin sync with the registered tool set (tests/unit/test_tool_names.pyenforces this). - Identifiers are normalised in
identifiers.py(HP:NNNNNNN; external CURIEs case-folded). - Ground every claim in the local database and cite the HPO id + HPO release
version.
hpo_versionis the per-call citation anchor (echoed on every non-minimal payload). The long-formrecommended_citationis inlined only onstandard/fullpayloads — term and association alike (one rule across both planes);compact/minimalcarryhpo_versionand defer the full citation toget_server_capabilities, which is the canonical source of record per the advertisedprovenance_policy. structloglogs to stderr only — stdout is reserved for the stdio MCP protocol. Neverprintto stdout outside the CLI.- Files stay under 500 lines (hard cap enforced by
scripts/check_file_size.pyin CI).
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 145 lines · 1,893 tokens per session scan A 17e795c3af57
hpo-link AGENTS.md is an instructions file published in the GitHub repository berntpopp/hpo-link (1 stars, last pushed 8d ago), licensed MIT. It adds 1,893 tokens to every session, about $0.0095 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
orphanet-link AGENTS.md
AGENTS.md instructions for berntpopp/orphanet-link, covering agents.md — orphanet-link, what this is, two planes (non-negotiable boundary), the eight orphadata products and where each is parsed and artifact pipeline.
orphanet-link CLAUDE.md
Claude Code instructions for berntpopp/orphanet-link, covering claude.md, essentials and common commands.
mgi-link AGENTS.md
AGENTS.md instructions for berntpopp/mgi-link, covering agents.md — engineering conventions for mgi-link, golden rules, architecture invariants (do not break), data plane and testing.
mgi-link CLAUDE.md
Claude Code instructions for berntpopp/mgi-link, covering claude.md, tl;dr, common commands and layout.
genefoundry-router AGENTS.md
Instructions for berntpopp/genefoundry-router, covering agents.md, what this project is, required check before handoff, coding standards and project-specific guidance.
genefoundry-router CLAUDE.md
Instructions for berntpopp/genefoundry-router, a project described as: MCP gateway federating 21 biomedical MCP servers — gnomAD, ClinVar, HPO, UniProt, Ensembl VEP, PanelApp and more — behind one Streamable-HTTP endpoint, with collision-free namespaced tools and BM25 tool search.