HealthChain
智能体读取临床数据很容易——但正确写回数据才是难点,而通用智能体框架并不会校验这一点。HealthChain 为任何模型或智能体提供类型安全、经过验证的 FHIR 工具,让它们值得信赖:来自正确系统的正确代码,作用于正确的患者,并带有有效的状态。此外还提供实时 EHR 连接和生产部署能力——让你构建的东西在演示之外同样经得起考验。
安装
pip install healthchainRelated MCP server: FhirMCP
快速开始
# Scaffold a FHIR Gateway project
healthchain new my-app -t fhir-gateway
cd my-app
# Run locally
healthchain serve编辑 app.py 以添加你的模型,并编辑 healthchain.yaml 以配置部署设置。
查看 CLI 参考 了解所有命令。
核心功能
为 AI 开发者和研究人员提供最快捷的医疗 AI 交付方式——开箱即用,随你扩展。
为什么选择 HealthChain?
每个严肃的医疗 AI 项目都在从零构建相同的集成基础设施。无论你部署的是逻辑回归、700 亿参数模型还是智能体工作流,训练模型与实时临床系统之间的壁垒都是一样的:真实的 FHIR API、经过验证的写入、多站点部署、可审计的治理。市面上没有现成的解决方案,而同时精通 AI 和医疗协议的工程师既稀缺又难以留住。
HealthChain 帮你处理这些复杂性,让你专注于真正重要的事情:模型和患者。
为实时而优化 - 连接实时 FHIR API 和集成点,而不是过时的数据导出
内置验证 - 类型安全的 FHIR 资源和验证报告,在数据发布前捕获错误数据——包括仅靠类型检查无法发现的规范无效临床编码
不虚构事实 - 辅助工具绝不会添加你没有提供的临床声明:不自动生成时间戳,不猜测状态——进入记录的内容与你的模型产出的完全一致
支持任何模型或智能体 - LLM、智能体或经典机器学习——都能输出经过验证的 FHIR
与你现有的技术栈兼容 - 与 FastAPI、MCP 和 LangChain 集成
生产级基础 - Docker 化部署、可配置的安全机制,以及为 NHS 和 HIPAA 环境设计的架构
🏆 认可与社区
特色与演讲:
入选 TLDR AI 通讯(90 万+开发者)
被 Medplum 推荐用于 Epic 开源集成
在 NHS Python 开源大会 上演讲(观看演讲)
源自 NHS AI 部署经验 – 阅读起源故事
🤝 合作与生产使用
正在为你的产品或组织探索 HealthChain?联系我们 讨论集成、试点或合作,或加入我们的 Discord 与社区交流。
使用示例
创建 Gateway [文档]
from healthchain.gateway import HealthChainAPI, FHIRGateway
from healthchain.fhir.r4b import Patient
# Create healthcare application
app = HealthChainAPI(title="Multi-EHR Patient Data")
# Connect to multiple FHIR sources
fhir = FHIRGateway()
fhir.add_source("epic", "fhir://fhir.epic.com/r4?client_id=epic_client_id")
fhir.add_source("cerner", "fhir://fhir.cerner.com/r4?client_id=cerner_client_id")
@fhir.aggregate(Patient)
def enrich_patient_data(id: str, source: str) -> Patient:
"""Get patient data from any connected EHR and add AI enhancements"""
bundle = fhir.search(
Patient,
{"_id": id},
source,
add_provenance=True,
provenance_tag="ai-enhanced",
)
return bundle
app.register_gateway(fhir)
# Available at: GET /fhir/transform/Patient/123?source=epic
# Available at: GET /fhir/transform/Patient/123?source=cerner
if __name__ == "__main__":
app.run(port=8888)为智能体提供 FHIR 工具 [文档]
from healthchain.tools import FHIRToolkit
# One toolkit: build, validate, read, and code FHIR — as typed agent tools
kit = FHIRToolkit(bundle="patient_bundle.json")
kit.as_mcp().run() # serve to Claude or any MCP client
tools = kit.as_langchain() # or drop into a LangChain agent或者直接在终端中操作,无需编写代码:
healthchain mcp --bundle patient_bundle.json使用 AI 助手进行构建
安装 HealthChain 插件,为 Claude Code 或 Codex 提供最新的 API、CLI 和配方:
# Claude Code
/plugin marketplace add healthchainai/HealthChain
/plugin install healthchain@healthchain
# Codex
codex plugin marketplace add healthchainai/HealthChain
codex plugin add healthchain@healthchain或者让任何助手访问 llms.txt 获取当前 API 文档的索引。
🛣️ 我们的开发路线
🔒 安全基础 — API 密钥认证、审计日志和 TLS,通过
healthchain.yaml配置并由网关中间件强制执行📋 治理即配置 — 将临床安全、数据访问协议以及 NHS/HIPAA 部署的合规标准作为
healthchain.yaml中的一等部署工件🔌 更深入的 EHR 连接 — 更多 FHIR 数据源、实时数据模式,以及来自试点部署的真实集成示例
📊 可观测性 — 面向医疗系统的部署遥测和审计追踪
🤖 临床 AI 智能体工具包 — 内置验证和术语的类型化 FHIR 工具,通过 MCP 和 LangChain 提供
🤝 贡献
HealthChain 由下一代医疗开发者构建并服务于他们——这些研究者将模型从回顾性数据带入实时系统,AI 开发者不想花几个月学习 FHIR 才能交付任何东西。最好的贡献来自那些遇到过真实问题并有具体见解的人。
开始参与:
正在处理医疗或研究数据? 贡献一个 cookbook — 带来你的用例,我会亲自支持你完成
阅读 CONTRIBUTING.md 了解贡献指南
技术问题和想法 → GitHub Discussions
试点和合作 → 邮件
🤗 致谢
本项目基于 fhir.resources 以及由 HL7 和 波士顿儿童医院 制定的 CDS Hooks 标准。
另请参阅 groundeval —— 面向医疗 AI 智能体的开源评估框架。
© 2024–2026 dotimplement ai。HealthChain 是由 dotimplement ai 维护的开源项目。
Available Tools
8 toolsbuild_resourceA
Build a FHIR resource from flat fields and validate it. Call this instead of writing FHIR JSON by hand — it does the nesting (CodeableConcepts, Dosage) and returns {ok, resource} or {ok: false, issues} to correct. Supported resource_type values: Condition, MedicationStatement, Observation, AllergyIntolerance, Patient. Use a code returned by lookup_code, never an invented one.
Inputs: resource_type (required), code, unit, onset, +12 more.
get_coded_entriesA
Read the coded identity of every resource of a type in the loaded bundle as flat records (code/display/system/status/authored_on/value/unit). Call this first for questions about conditions, medications, allergies, or observations — it resolves medicationReference to the actual Medication for you. Optionally filter by status (e.g. 'active').
Inputs: resource_type (required), status.
get_resourcesA
Get all resources of a type from the loaded bundle as full FHIR JSON. Call this when you need fields the flattened get_coded_entries view doesn't carry.
Inputs: resource_type (required).
list_codesA
List every code in the local terminology catalog, optionally filtered to one code system URI. Call this to browse what codes are available (e.g. to build a mention list); use lookup_code to search by name.
Inputs: system.
load_bundleA
Load a FHIR Bundle from a file path or JSON string and make it the working bundle for the read tools. Call this first when given bundle data. Invalid bundles return issues[] locating each problem (e.g. Bundle.entry[2].resource.subject).
Inputs: source (required).
lookup_codeA
Search the terminology catalog for codes matching free text (e.g. 'metoprolol 25', 'type 2 diabetes'). Call this whenever you need a code — use a returned candidate, never invent or recall one. Results are ranked tightest-match-first; optionally restrict to one code system URI.
Inputs: query (required), system.
resolve_referenceA
Resolve a FHIR reference (e.g. 'Patient/123', 'urn:uuid:...') to its target resource within the loaded bundle. For a contained reference ('#med1'), also pass parent_resource_id — the id of the resource the reference appears in.
Inputs: reference (required), parent_resource_id.
validate_resourceA
Validate a FHIR resource (JSON object) against the spec, including required value-set bindings. Call this after constructing or editing FHIR JSON yourself; read issues[] (severity/diagnostics/expression) and correct the resource. Never raises.
Inputs: resource (required), resource_type.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 8 tools
Each tool has a clear, distinct purpose: building, validating, loading, retrieving, resolving references, and terminology lookup. No two tools overlap in functionality.
All tool names follow a consistent verb_noun pattern using lowercase and underscores, e.g., build_resource, load_bundle, get_resources, lookup_code. This makes the API predictable.
With 8 tools, the set is well-scoped for a FHIR utility server, covering essential operations without being bloated or too sparse.
The set covers building, validating, loading, and reading FHIR resources, as well as reference resolution and terminology lookup. Minor gaps like missing update/delete tools or limited build resource types prevent a perfect score.
Maintenance
Related MCP Connectors
Privacy-preserving synthetic health data generation. FHIR R4/R5 compliant.
Securely access and manage FHIR healthcare data stored in Medplum.
Guardrailed FHIR access for AI agents: PHI redaction, audit trail, step-up auth, tenant isolation
300+ deterministic data utilities for AI agents: validate, normalize, parse, match, redact.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables LLM-based agents to interact with FHIR healthcare data through natural language prompts, providing full CRUD operations on FHIR resources, document processing, and semantic search capabilities.13102MIT
- AlicenseNot gradedqualityBmaintenanceEnables LLMs to securely interact with FHIR healthcare servers and HL7 terminology services. Provides comprehensive healthcare data operations with built-in PHI protection, audit logging, and SMART on FHIR authentication.MIT
- AlicenseAqualityBmaintenanceProvides seamless integration with FHIR APIs, enabling AI/LLM tools to search, retrieve, and analyze clinical healthcare data with support for SMART-on-FHIR authentication and multiple transport protocols.71,087 PyPI139Apache 2.0
- FlicenseNot gradedqualityAmaintenanceEnables LLM-based agents and MCP-compatible clients to query and manipulate FHIR resources through full CRUD tools, with document ingestion, semantic search, LOINC terminology lookup, and OAuth2-authenticated FHIR API integration.-