Priority REST API MCP Server
An MCP server that exposes Priority ERP's OData operations (query, create, update, delete, batch, attachments, text fields) as 19 tools for AI assistants like Claude. It supports Streamable HTTP transport with OAuth 2.1 and two-layer authentication (bearer token plus Priority credentials).
README
Priority REST API MCP Server
An MCP server that connects AI assistants — Claude and others — directly to a Priority ERP system. Every OData operation (query, create, update, delete, batch, attachments, text fields) is exposed as an MCP tool, so AI agents can read and write live business data without custom integration code.
Version: 0.2.0 · Transport: Streamable HTTP (SSE optional) · Runtime: Node.js 18 · Tools: 19
Quick Start
1. Clone and install
git clone https://github.com/priority-mcp/priority-odata-mcp priority-mcp
cd priority-mcp
npm install
2. Create .env from the example
cp .env.example .env
At minimum, set these four variables:
PRIORITY_BASE_URL=https://<host>/odata/Priority/<tabula.ini>/<company>/
PRIORITY_AUTH_TYPE=basic
PRIORITY_USERNAME=myuser
PRIORITY_PASSWORD=mypassword
3. Start the server
# Development (from source)
node src/index.js
# Production (bundled)
npm run build
node dist/index.js
On first run, if ODATA_MCP_TOKEN is not set, a random Bearer token is generated and printed to stdout. Copy it for the next step.
4. Connect from Claude Code
Add to your MCP config:
{
"mcpServers": {
"priority": {
"type": "http",
"url": "http://localhost:3000/mcp",
"headers": {
"Authorization": "Bearer <ODATA_MCP_TOKEN>"
}
}
}
}
Transport
The server uses Streamable HTTP as its primary transport — each POST /mcp request is fully stateless. A new McpServer and StreamableHTTPServerTransport are created per request and torn down after.
| Endpoint | Method | Purpose |
|---|---|---|
/mcp |
POST | Primary MCP endpoint (Streamable HTTP) |
/sse |
GET | SSE stream — requires SSE_ENABLED=true |
/sse |
POST | JSON-RPC messages for SSE clients |
/health |
GET | Health check — returns version and status |
/.well-known/oauth-authorization-server |
GET | OAuth 2.1 discovery (required by Claude Code ≥2.1.92) |
/authorize, /token, /register |
GET/POST | OAuth 2.1 PKCE flow — auto-approves |
Note: The OAuth 2.1 endpoints exist to satisfy Claude Code's Streamable HTTP connection handshake. They auto-approve all requests and are not intended for real access control — that is handled by
ODATA_MCP_TOKEN.
Authentication
Authentication operates at two independent layers.
Layer 1 — Protecting this server
All routes (except /health and OAuth endpoints) require:
Authorization: Bearer <ODATA_MCP_TOKEN>
Set ODATA_MCP_TOKEN in .env. If absent, a random UUID is generated at startup and printed to stdout.
Layer 2 — Calling Priority ERP
Controlled by PRIORITY_AUTH_TYPE:
basic— HTTP Basic auth usingPRIORITY_USERNAME+PRIORITY_PASSWORDpat— Bearer token viaPRIORITY_PAToauth2— same aspat(pass PAT as Bearer token)none— no auth header (local testing only)
Write operations (POST/PATCH/DELETE) automatically fetch and retry with an X-CSRF-Token header if the initial request is rejected, following Priority's CSRF protection pattern.
Optional per-application license headers are sent with every Priority request when PRIORITY_APP_ID and PRIORITY_APP_KEY are set (X-App-Id / X-App-Key).
Configuration
Copy .env.example to .env. The server searches for .env in order: ENV_FILE_PATH → ./mcp-servers/Priority-REST-API-MCP-Server/.env → ./.env.
Required
| Variable | Description |
|---|---|
PRIORITY_BASE_URL |
OData root URL — format: https://<host>/odata/Priority/<tabula.ini>/<company>/ |
PRIORITY_AUTH_TYPE |
basic | pat | oauth2 | none |
PRIORITY_USERNAME |
Username — required when AUTH_TYPE=basic |
PRIORITY_PASSWORD |
Password — required when AUTH_TYPE=basic |
Priority Auth (optional)
| Variable | Description |
|---|---|
ODATA_MCP_TOKEN |
Bearer token protecting /mcp. Random UUID used if not set. |
PRIORITY_PAT |
Personal Access Token (when AUTH_TYPE=pat or oauth2) |
PRIORITY_APP_ID |
Application license ID — sent as X-App-Id header |
PRIORITY_APP_KEY |
Application license key — sent as X-App-Key header |
PRIORITY_LANGUAGE |
Overrides Accept-Language header (e.g. en) |
HTTP Server
| Variable | Default | Description |
|---|---|---|
HTTP_HOST |
0.0.0.0 |
Bind address |
HTTP_PORT |
3000 |
Listen port |
SSE_ENABLED |
false |
Enable /sse endpoint |
Timeouts & TLS
| Variable | Default | Description |
|---|---|---|
PRIORITY_HTTP_TIMEOUT_MS |
30000 |
Read timeout for Priority API calls (ms) |
MCP_WRITE_TIMEOUT |
15000 |
Timeout for POST/PATCH/DELETE operations (ms) |
MCP_PROC_TIMEOUT |
45000 |
Timeout for batch operations (ms) |
TLS_REJECT_UNAUTHORIZED |
false |
Set true in production to reject self-signed certs |
Debugging
| Variable | Default | Description |
|---|---|---|
LOG_LEVEL |
INFO |
DEBUG logs every request and response |
MCP_DEBUG |
false |
Print full OData URLs, params, result counts |
PRIORITY_ENABLE_TRACE |
false |
Adds X-App-Trace: 1 to every Priority request |
STRICT_DATA_INTEGRITY |
true |
Throws on empty/mock API responses — disable only for testing |
ENV_FILE_PATH |
— | Override path to .env file (useful for submodule deployments) |
Tools
All 19 tools are defined in src/tools/ and registered in src/tools/priorityTools.js.
System & Metadata
| Tool | Description | Parameters |
|---|---|---|
version_get |
Fetch the Priority service version and response headers | — |
metadata_entities_list |
List all OData entity sets; filter to REST-enabled forms only | apiOnly?, includeMetadata? |
metadata_schema_get |
Get field schema for an entity by fetching a sample record. Auto-redirects subform names to parent + $expand |
entity, sample?, top? |
metadata_refresh |
Clear and refresh server-side metadata cache. Always does a full flush (see Known Limitations) | entity? |
Querying
| Tool | Description | Parameters |
|---|---|---|
entity_get |
Fetch a single record by key or lookup, with optional $expand and $select |
entity, key, lookup, select?, expand? |
query_run |
Run an OData query with full filter/select/top/skip/orderby/expand/count support. Validates date filter results post-fetch | entity, filter?, select?, top?, skip?, orderby?, expand?, count?, deltaToken? |
safe_query_run |
Like query_run but auto-discovers valid fields first and validates $select field names before executing — prevents 400 errors from invalid column names |
entity, filter?, select?, top?, skip?, expand?, count? |
query_sum |
Sum a numeric field across an entity with an optional filter. Tries $apply=aggregate first; falls back to a full paging scan |
entity, field?, filter? |
Create / Update / Delete
| Tool | Description | Parameters |
|---|---|---|
entity_create |
Create a new record. Supports subform creation via parentEntity + parentKey + subform |
entity, data, parentEntity?, parentKey?, parentLookup?, subform? |
entity_update |
Update a record via PATCH with If-Match: *. Supports composite keys |
entity, key, data, parentEntity?, parentKey?, subform? |
entity_delete |
Delete a record via DELETE with If-Match: *. Supports subform deletes |
entity, key, parentEntity?, parentKey?, subform? |
batch_operations |
Execute multiple POST/PATCH/DELETE in one $batch request with dependency chaining |
requests[] (id, method, url, body?, dependsOn?) |
Text Fields
| Tool | Description | Parameters |
|---|---|---|
entity_text_get |
Fetch the rich-text content of a record's /Text subresource |
entity, key |
entity_text_create |
POST new text content to /Entity(Key)/Text |
entity, key, textData |
entity_text_update |
PATCH existing text content on /Entity(Key)/Text |
entity, key, textData |
Attachments
| Tool | Description | Parameters |
|---|---|---|
entity_attachments_get |
List attachments on a record | entity, key |
entity_attachments_upload |
Upload a file to a record's /Attachments subresource as multipart/form-data. fileData must be base64-encoded |
entity, key, fileData, fileName, contentType? |
Configuration & Help
| Tool | Description | Parameters |
|---|---|---|
instructions_get |
Returns the full operational guide: OData syntax, subform patterns, throttle limits, date handling rules, known failure patterns, and architecture examples. Call this first when exploring an unfamiliar entity | — |
config_restflag_update |
Set RESTFLAG=Y or N in the FORMLIMITED table to enable or disable REST API access for a Priority form |
formName, restFlag, formType? |
Prompts & Resources
The server registers MCP prompts (reusable instruction templates) and resources (live data endpoints).
Prompts (src/prompts/)
| Name | Purpose |
|---|---|
query_priority_entity |
Guide for constructing OData queries against an entity |
explore_entity_relationships |
Explains the subform hierarchy for a given entity |
modify_priority_data |
Guides create, update, and delete operations |
date_handling_guide |
Critical rules for date filters — ISO format, operator validation |
known_failure_patterns |
Documented 404/501/400 patterns and their workarounds |
pagination_guide |
Explains $top/$skip and count patterns |
Resources (src/resources/)
| URI | Purpose |
|---|---|
priority://entities/list |
Live list of all REST-enabled entities (RESTFLAG=Y) |
priority://entity-schema/{entity} |
Schema for a specific entity (template URI) |
priority://queries/common |
Library of ready-to-use query examples |
priority://subforms/reference |
Reference guide for subform patterns and operations |
Example Tool Call
Query the three most recent sales orders for customer 1011 — sent as JSON-RPC 2.0 to POST /mcp:
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "query_run",
"arguments": {
"entity": "ORDERS",
"filter": "CUSTNAME eq '1011'",
"select": ["ORDNAME", "CUSTNAME", "CURDATE", "TOTPRICE"],
"top": 3,
"orderby": "CURDATE desc"
}
}
}
The server issues:
GET /odata/Priority/.../ORDERS?$format=json&$filter=CUSTNAME+eq+'1011'
&$select=ORDNAME,CUSTNAME,CURDATE,TOTPRICE&$top=3&$orderby=CURDATE+desc
Response:
{
"jsonrpc": "2.0",
"id": 1,
"result": {
"content": [{
"type": "text",
"text": "{\"value\":[{\"ORDNAME\":\"SO25000001\",\"CUSTNAME\":\"1011\",\"CURDATE\":\"2025-07-15T00:00:00+03:00\",\"TOTPRICE\":15000.0},...],\"_mcp_metadata\":{\"entity\":\"ORDERS\",\"resultCount\":2,\"filterApplied\":true}}"
}],
"isError": false
}
}
Date format: Priority returns dates as ISO 8601 with a timezone offset (e.g.
2025-07-15T00:00:00+03:00), not UTCZ. UseCURDATE ge 2025-01-01syntax in date filters — not ISO-Z format.
Deployment
Docker
# Build
docker build -t priority-mcp .
# Run
docker run --env-file .env -p 3000:3000 priority-mcp
The Dockerfile uses node:18-slim, runs npm run build to bundle src/ → dist/ via esbuild, then starts dist/index.js. A Docker Compose setup and local TLS certificate generator are in deployment/local/.
Production checklist
- Set
ODATA_MCP_TOKENexplicitly — do not rely on the auto-generated one - Set
TLS_REJECT_UNAUTHORIZED=true - Set
STRICT_DATA_INTEGRITY=true(default) - Set
LOG_LEVEL=INFO(default — suppresses housekeeping noise) - Pin
HTTP_HOSTto a specific interface if not exposing publicly
Known Limitations
Priority ERP-specific behaviors worth knowing before you build.
Rate limiting — 100 calls/minute per user Priority Cloud throttles to 100 API calls/minute per user, maximum 10 parallel requests, 3-minute timeout per call. Design agents to batch operations where possible.
Response cap — MAXFORMLINES
Priority silently truncates responses at the MAXFORMLINES system constant regardless of $top. Use $skip-based pagination if you need all records.
Subforms are not standalone entities
Querying PORDERITEMS_SUBFORM directly returns HTTP 404. Subforms must be accessed via the parent entity with $expand=PORDERITEMS_SUBFORM. metadata_schema_get auto-detects this and redirects.
$apply=aggregate not supported
query_sum always falls back to a full paging scan because $apply=aggregate(...) is not supported on this Priority version.
GET /ENTITY/$count returns 500
Use ?$top=0&$count=true instead. Internally, tryEstimateCount() tries /$count first then pages in 500-record batches (capped at 10,000).
contains()/startswith() unsupported on some fields
EPROG.ENAME and EREP.ENAME only support eq exact match — string functions return HTTP 501.
Entity-level metadata refresh returns 400
metadata_refresh ignores the entity argument and always does a full cache flush, because Priority rejects entity-scoped cache-clear requests.
Batch URL encoding
URLs inside batch_operations requests are never auto-encoded. Spaces and special characters must be manually percent-encoded (spaces → %20).
Composite keys
Some entities use composite keys, e.g. FORMLIMITED: ENAME='X',TYPE='F'; AINVOICES: IVNUM='T9696',IVTYPE='A',DEBIT='D'. Pass the full composite key string to entity_update and entity_delete.
Project Structure
/
├── src/
│ ├── index.js Entry point — creates and starts PriorityMCPServer
│ ├── server.js Express app, all routes, auth guard, OAuth 2.1 PKCE
│ ├── sseServer.js SSE connection manager
│ ├── config.js Reads all env vars, resolves .env path
│ ├── version.js SERVER_VERSION, KNOWN_ISSUES list
│ │
│ ├── priority/
│ │ └── client.js PriorityClient — axios instance, auth headers,
│ │ all API methods (runQuery, createEntity, …)
│ │
│ ├── mcp/
│ │ ├── handler.js JSON-RPC 2.0 dispatcher (SSE path)
│ │ ├── registry.js ToolRegistry — registerTool, callTool, listTools
│ │ ├── prompt-registry.js
│ │ ├── resource-registry.js
│ │ ├── priority-mcp-sdk-server.js Wires registries into McpServer (SDK path)
│ │ ├── tool-call-runner.js Executes tool, wraps result for MCP response
│ │ └── json-schema-to-zod.js JSON Schema → Zod conversion
│ │
│ ├── tools/ One file per tool + priorityTools.js (registration)
│ ├── prompts/ One file per prompt + priorityPrompts.js
│ ├── resources/ One file per resource + priorityResources.js
│ └── utils/
│ ├── data-integrity.js ensureNoMockData(), validateApiResponse()
│ ├── date-handling.js Date parsing and validation helpers
│ ├── errors.js createPriorityApiError(), FilterNotAppliedError
│ ├── filter-resolver.js OData filter string building
│ ├── expand-resolver.js $expand normalization
│ ├── entity-resolver.js Entity name / subform name resolution
│ ├── resolve-query-args.js
│ └── subform-query-resolver.js
│
├── data/
│ └── entity-relationships.json Hardcoded subform map (PORDERS, ORDERS, …)
│
├── tests/
│ ├── scripts/ Manual test scripts
│ └── results/ Saved JSON/Markdown test output
│
├── docs/ Design docs (DATA_INTEGRITY_POLICY, DATE_HANDLING_RULES, …)
├── postman/ Postman collection for manual API testing
├── deployment/local/ Docker Compose + TLS cert generator
├── build.js esbuild bundler: src/ → dist/
└── .env.example All env vars documented with descriptions
Tests
There is no automated test runner. Tests are manual scripts that require a live Priority connection:
# Read operations
node tests/scripts/test-priority-operations.js
# Write operations (interactive — asks for confirmation)
node tests/scripts/test-write-operations.js
# Test all 19 MCP tools via the running server
node tests/scripts/test-all-mcp-tools-via-server.js
# Standalone resolver smoke tests
node test-keyresolver.js
node test-resolver.js
Warning: Write tests will create, update, and delete real records. Run against a development company only.
Tech Stack
- Runtime: Node.js 18, ES Modules (
"type": "module") - MCP SDK:
@modelcontextprotocol/sdk ^1.29.0 - HTTP server:
express ^4.21.1 - HTTP client:
axios ^1.7.7 - Schema validation:
zod ^4.3.6 - Bundler:
esbuild ^0.25.0(vianpm run build) - Other:
cors,dotenv,form-data,uuid,http-errors
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。