servicenow-ppm-governance
This MCP server enables AI agents to interact with ServiceNow instances, providing typed tools to query, create, update, aggregate, and run background scripts, with read-only mode and a --doctor connection check.
README
servicenow-ppm-governance
Stage-gate approvals, change-request control and demand intake, built on ServiceNow SPM — plus the MCP server used to build and verify it against a live instance.
Built on a personal ServiceNow developer instance (PDI) with synthetic test data. No employer data, configuration or credentials are present in this repository. The specification it implements is my own.
What this is
Two things that grew together:
1. A PPM governance model (ppm-build/) — a stage-gate lifecycle with a driven
approval matrix, snapshot-based change control on project financials, and an M0 qualification gate
on demands that converts an approved demand into a project. Roughly 3,000 lines of business rules,
client scripts, script includes and the deployment tooling that installs them idempotently.
2. A ServiceNow MCP server (server.js) — the tool that built it. Exposes the
instance to an AI agent as typed tools (query, describe, create, update, aggregate, background
script), with a read-only mode, response truncation and a --doctor connection check. Everything in
ppm-build/ was written, deployed and verified through it. See
docs/mcp-server.md.
What got built
| Area | Detail |
|---|---|
| Stage gates | u_pm_gates table, 26 custom fields, 4 business rules, form sections, numbering, ACLs, navigation |
| Change requests | 4 business rules plus snapshot fields on project_change_request — financials are captured at request time so the approved delta is provable |
| M0 demand intake | 20 fields and 5 business rules on dmn_demand; on full approval the demand converts to a project |
| Approver matrix | u_approver_matrix, keyed on project type + portfolio + phase, driving approver generation across all three record types |
| Approval rollup | One rule on sysapproval_approver covering gates, change requests and demands |
Everything is captured in a single update set, so the whole build is portable to another instance.
The engineering log
ppm-build/README.md is the substance of this repository — the run
order, every verification result, and the gotchas that cost real time. It records what did not work
alongside what did.
A few worth surfacing, because they are the kind of thing you only learn by hitting them:
sys_script.nameisVARCHAR(40). Longer names truncate silently, so an upsert matching onname=never finds the record and creates a duplicate on every deploy.deploy.jsnow fails fast on this rather than quietly accumulating rules.- Creating a field auto-adds it to the default form section (the platform rule "Add New Column To Form Section"), so fields appear twice after a build unless you re-run the duplicate check.
GlideAggregateSUM disagrees with agetValue()loop on the cost plan breakdown columns. The loop is correct. Do not "optimise" it into an aggregate.cost_plan_breakdown.estimate_to_completionis virtual with a calculation script — it cannot be written.actualisread_onlybut is script-writable.- Approvals on an inactive record are voided by the out-of-box "Moot Approvals Upon Cancellation"
rule, which filters on
active=false. M0 validation therefore blocks requests on a closed demand.
Verification
The build is verified by scripts in the repository, not by inspection — 11_verify.js covers the
gate and change-request lifecycle (21 of 22 checks), 12_verify_ajax.js the client-callable path,
and 13_verify_m0.js the demand process end to end (20 checks, 0 failures). Results, including the
check that did not pass and why, are recorded in the engineering log rather than summarised away.
One example of that honesty being useful: the source specification states that a gate closes when all
approvals are granted, but no rule in it actually sets the parent approval field when the
sysapproval_approver records all approve. That was confirmed empirically on both record types
before a rollup rule was written to close the gap — the alternative would have been assuming the
spec was complete and shipping a lifecycle that silently never closed.
Quickstart
Requires Node 18+ and a ServiceNow PDI.
cp .env.example .env
# fill in your instance URL, username and password
node server.js --doctor
Deploy the build:
node server.js --run ppm-build/01_fields.js # prerequisite fields
node ppm-build/deploy.js # business rules + client scripts
node ppm-build/deploy_si.js # script includes
node server.js --run ppm-build/03_form_layout.js # form sections
node server.js --run ppm-build/10_seed.js # synthetic test data
node server.js --run ppm-build/11_verify.js # verification
The full 13-step run order is in ppm-build/README.md.
Both deploy scripts are idempotent — re-running them updates in place rather than duplicating.
Run any script against the instance with filtered output:
node server.js --run ppm-build/<file>.js
Repository layout
server.js MCP server + `--run` CLI for background scripts
ppm-build/
README.md the engineering log — start here
br/ 15 business rules
cs/ 6 client scripts
si/ 2 script includes (approver matrix, gate financials AJAX)
deploy.js idempotent business rule + client script deployment
deploy_si.js idempotent script include deployment
fetch_si.js pull an out-of-box script include down for reference
01–13_*.js the build and verification scripts, in run order
docs/mcp-server.md MCP server setup, tools and configuration
ppm-build/oob/ is gitignored. fetch_si.js writes ServiceNow's own out-of-box source there for
reference while working; that is the vendor's code, not mine, so it stays local.
Deliberately not built
Two decisions were left open rather than invented, because the specification did not cover them and guessing would have produced governance nobody asked for:
- Gate sequencing. Nothing auto-creates the P0 → P1 → P2a → P2b → P3 set for a project, or opens the next gate when one closes. The specification describes a single gate's lifecycle only.
- The out-of-box demand
Approvebutton. ServiceNow's ownApprove/Reject/Qualify/Screenbuttons ondmn_demandonly flipstate— they do not touchsysapproval_approverand are not wired to M0. MakingApprovedrive M0 would mean overriding out-of-box behaviour, which is a decision for whoever owns the instance.
Licence
MIT — see LICENSE. ServiceNow is a trademark of ServiceNow, Inc.; this project is not affiliated with or endorsed by ServiceNow.
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。
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。
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。
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。