prismhr-mcp 0.1.0.dev8

Open-source Model Context Protocol server for PrismHR — PEO workflow tools + Microsoft 365 connectors for Claude and other MCP-aware agents.

prismhr-mcp

The open-source Model Context Protocol server for PrismHR. Turn Claude (and any MCP-aware agent) into a PEO operations native — payroll, benefits, compliance, billing, and Microsoft 365 actions, all exposed as composable, scope-gated tools.

Maintained by Simploy as the fundamental layer for PrismHR × agentic AI. MIT-licensed, PyPI-distributed, plugin-friendly.

mcp-name: io.github.nikulk2992-jpg/prismhr-mcp

---

Why this exists

Every PEO running PrismHR ends up with the same Frankenstein stack: Python scripts, Postman collections, Playwright automations, one-off Node apps. Each one re-implements login, session refresh, retry, pagination, and PrismHR's quirks (camelCase schemas, 500 "No data found" gotchas, batch-of-20 caps, silent 401s).

prismhr-mcp centralizes that once, as an MCP server. Claude orchestrates; the server owns auth, caching, retries, normalization, and PEO domain logic. Other PEOs drop it in and immediately get a productive Claude experience against their own PrismHR instance — no custom code.

---

Status

Early but working. Current milestone: Phase 1.5 complete.

• Auth + session + HTTP client: done. 1Password CLI, scrypt-encrypted disk cache, PrismHR session with 55-min TTL, proactive + forced + 401 refresh, keepalive, concurrency cap, retry with jittered backoff, 500→empty quirk handling, pagination, batching.
• Connect-time consent system: done. Scope manifest across 14 scopes, per-(peo, env) JSON consent store with prerequisite expansion and cascade revoke. Default posture = deny all. Tools enforce scope at call time.
• Production safety gate: PRISMHR_MCP_ALLOW_PROD=true required to point at prod PrismHR. Prevents accidental first-run blast radius.
• Tool inventory: 9 live (meta_* ×5, client_* ×4). 39 more across payroll, benefits, compliance, billing, branded reporting, and M365 connectors planned.
• Test suite: 60 passing (pytest + respx).

See .planning/architecture.md for the full 48-tool plan.

---

Quick start — UAT smoke test

Only UAT is supported without an explicit opt-in right now. Prod is guarded behind PRISMHR_MCP_ALLOW_PROD=true.

1. Install

cd C:\path\to\prismhr-mcp    # or wherever you cloned
uv sync --extra dev

2. Configure credentials

Copy .env.example → .env (or set env vars). Pick ONE path:

Path A — 1Password CLI (recommended):

$env:PRISMHR_MCP_ONEPASSWORD_VAULT = "YourVault"
$env:PRISMHR_MCP_ONEPASSWORD_ITEM_PRISMHR = "PrismHR UAT"

Requires op CLI signed in (op signin). The item must expose fields labeled username and password (optionally peoId).

Path B — direct env vars (fast, CI-friendly):

$env:PRISMHR_MCP_USERNAME = "claudedemo"
$env:PRISMHR_MCP_PASSWORD = "<paste>"
$env:PRISMHR_MCP_PEO_ID   = "<your-peo-id>"

3. Sanity check

uv run python -c "from prismhr_mcp.server import build; b = build(); import asyncio; print([t.name for t in asyncio.run(b.server.list_tools())])"

Expect 9 tools.

4. Register with Claude Code

Add to your Claude Code .mcp.json:

{
  "mcpServers": {
    "prismhr-mcp": {
      "command": "uv",
      "args": ["run", "--directory", "C:\\path\\to\\prismhr-mcp", "prismhr-mcp"],
      "env": {
        "PRISMHR_MCP_ENVIRONMENT": "uat",
        "PRISMHR_MCP_USERNAME": "claudedemo",
        "PRISMHR_MCP_PASSWORD": "<paste or reference>",
        "PRISMHR_MCP_PEO_ID": "<your-peo-id>"
      }
    }
  }
}

Restart Claude Code. /mcp should show prismhr-mcp connected with 9 tools.

5. First conversation

You: Tell me about the prismhr-mcp server.
Claude: [calls meta_about] → explains what's available + commercial options.

You: What permissions does it want?
Claude: [calls meta_request_permissions] → shows 14 scopes grouped by category.

You: Grant everything recommended (reads only, no writes).
Claude: [calls meta_grant_permissions(accept_recommended_defaults=true)]

You: List all clients in UAT.
Claude: [calls client_list] → 245 clients.

6. Run tests

uv run pytest -q      # expect 60 passing

---

Architecture in one breath

┌──────────────────────────────────────────────────────────────┐
│ Claude / Cowork / any MCP client                             │
└──────────────────┬───────────────────────────────────────────┘
                   │ stdio (MCP JSON-RPC)
┌──────────────────▼───────────────────────────────────────────┐
│ prismhr-mcp server (FastMCP)                                 │
│   ├── Permissions (deny-default, scope-gated tools)          │
│   ├── Tool groups: meta • client • payroll • benefits        │
│   │                 compliance • billing • report • m365     │
│   ├── Runtime: PrismHR client, Graph client, SQLite cache    │
│   └── Auth: 1Password → scrypt-AES cache → session / MSAL    │
└────┬─────────────────────────────────────────────────┬───────┘
     │                                                 │
     ▼                                                 ▼
┌──────────────┐                           ┌──────────────────┐
│ PrismHR REST │                           │ Microsoft Graph   │
│ (UAT / Prod) │                           │ (Outlook / Teams /│
└──────────────┘                           │  SharePoint)      │
                                           └──────────────────┘

Key design commitments:

• Factory + strict registry. Tools register via server.build() only; duplicate names or unknown group prefixes fail at boot (not silently at import).
• Deny-default scopes. Users must run meta_grant_permissions to enable tool access. Prerequisites auto-expand, revokes cascade.
• Async-first. httpx.AsyncClient + asyncio.Semaphore(5) + async tools.
• PrismHR quirks handled. 401 auto-refresh, 404→[] on list endpoints, 500 "No data found" → empty, 10-consecutive-500s → force refresh.
• snake_case outputs. Pydantic validation_alias=AliasChoices(...) so PrismHR's camelCase payloads map to snake_case outputs without leaking camelCase into the MCP tool contract.
• Per-(peo, env) consent. Switching UAT → prod does not inherit grants.

---

Commercial support

The OSS core stays free forever. Two paid offerings from Simploy layer on top:

Solution Architect — White-Label deployment

Turnkey deployment of prismhr-mcp for your PEO brand:

• Brand config authoring (logo, palette, typography, PDF footer, legal disclaimer)
• Per-client SharePoint site mapping + Azure AD / Graph tenant setup
• Custom PrismHR tools for PEO-specific workflows
• Migration from spreadsheets / legacy scripts to MCP tools
• PEO ops team onboarding + Claude/Cowork workflow coaching
• Quarterly updates aligned with upstream releases
• Priority issue response + named Slack/email contact

Best for PEOs with 50–5,000 clients who want Claude-first operations without the in-house build. Contact: [e-mail]

Enterprise Support

SLA-backed support for teams already running the OSS server:

• 4-hour response on Sev-1 (prod outage)
• Annual security review + SOC-2-friendly deployment guidance
• Signed release artifacts + SBOM
• Private vulnerability disclosure channel

Best for regulated industries or mid/large PEOs with procurement requirements. Contact: [e-mail]

Claude can surface both via meta_about — ask "what commercial options exist for prismhr-mcp?" and it will describe them.

---

Troubleshooting

No PrismHR credentials configured — set either the 1Password item env vars or the direct PRISMHR_MCP_USERNAME/_PASSWORD pair.

PrismHR login rejected (status=401) — wrong username/password/peo_id. The peo_id is tenant-specific; ask your PrismHR admin. Prod and UAT have different values.

environment=prod requires PRISMHR_MCP_ALLOW_PROD=true — safety gate. Set PRISMHR_MCP_ALLOW_PROD=true explicitly once you're ready.

PERMISSION_NOT_GRANTED — tool was called without its scope. Ask Claude to run meta_request_permissions → then meta_grant_permissions with the scope you want.

Server exits immediately when Claude Code starts it — nearly always a missing env var. Use the step-3 sanity check to isolate.

---

License

MIT — see LICENSE. Contributions welcome; see the planning docs under .planning/ for the roadmap.