HCM MCP Server for Oracle Fusion Cloud (unofficial)

Not affiliated with or endorsed by Oracle. Oracle and Fusion are trademarks of Oracle Corporation.

An MCP server that lets AI models interact with Oracle Fusion Cloud HCM across its entire REST surface — workers, org structures, compensation, absence, payroll, recruiting, talent, learning and more — without hand-coding a tool per endpoint.

Built to be packaged once and reused across many Fusion HCM customers, regardless of which modules they license, how their flexfields are configured, or which Oracle release they run.

Status: Feature-complete. All 16 tools implemented — discovery, generic read, 7 curated HR workflows, ATOM change feeds, and gated writes — with an inescapable client-layer PII-redaction + audit floor. Built on ADF REST behavior verified against a live Fusion pod (DESIGN.md §13). 60 unit tests; Docker image and stdio server boot verified. After installing, run the smoke checklist in docs/TEST_CASES.md against your pod (every pod's licensing, roles, and flexfields differ).

Why this exists

Oracle Fusion HCM exposes ~600 REST resources built on Oracle's uniform ADF REST framework. A naive MCP server would create one tool per resource and blow up the model's context window — and would advertise tools that don't exist for customers who haven't licensed the matching module.

This server takes the opposite approach: a small set of generic, schema-aware tools that introspect any resource at runtime via Oracle's /describe endpoint, plus a handful of curated workflow tools for the most common HR tasks. It discovers each customer's licensed footprint automatically and only exposes what actually works on their pod.

Related MCP server: OpenAPI MCP Server

Key design principles

• Generic over hardcoded — 6 generic tools + ~15 curated workflows cover all ~600 resources. No per-resource code.

• Self-documenting — the model reads live schemas via describe_resource; no bundled schema files to maintain.

• Capability-aware — startup probing detects which Oracle modules are licensed/provisioned and lights up only those tool groups (no Recruiting license → no Recruiting tools).

• Safe by default — read-only out of the box; writes are off by default, gated, dry-run-first, and audited. PII (national IDs, salary, DOB) is redacted unless explicitly enabled.

• Distributable — one configurable artifact, deployed single-tenant per customer pod. No customer specifics in code.

Architecture at a glance

auth/     Basic + OAuth2/JWT (OCI IAM / IDCS)
core/     ADF REST client · resource catalog · /describe cache · q= filter builder
tools/    discovery · query · workflows · mutate · atom · bip
safety/   PII redaction · audit log · dry-run gates (writes off by default, schema-validated, fail closed)
config.py base URL · pinned REST version · scopes · feature & module flags

Tools (16)

The generic query_resource / get_record / describe_resource work against any of the ~600 HCM resources; the workflows are curated shortcuts for common HR questions.

Roadmap

Stack

Python · FastMCP · httpx · pydantic Packaged as a Docker/OCI image (primary) built from a hatchling wheel.

Getting started

Install from PyPI

pipx install fusion-hcm-mcp-server    # or: uvx fusion-hcm-mcp-server

pypi.org/project/fusion-hcm-mcp-server — published from this repo's CI via PyPI Trusted Publishing with attestations.

Want to use this from Claude Desktop? Follow the step-by-step guide: docs/INSTALL_CLAUDE_DESKTOP.md — covers both install modes (Python venv + OS credential store, or Docker), the Desktop config file, a test sequence, and the common traps.

Configure

cp config.example.toml config.toml   # then edit; supply secrets via HCM_* env vars

Required: server.base_url (or HCM_BASE_URL). Credentials should come from environment variables (HCM_USERNAME/HCM_PASSWORD, or HCM_CLIENT_ID/HCM_CLIENT_SECRET/HCM_TOKEN_URL), never the committed file. See config.example.toml.

Run with Docker (primary)

Prebuilt images are published to GitHub Container Registry on every release:

docker pull ghcr.io/mohantyajitesh/fusion-hcm-mcp-server:latest
docker run --rm -i \
  -e HCM_BASE_URL="https://your-pod.fa.ocs.oraclecloud.com" \
  -e HCM_USERNAME="INTEGRATION_USER" -e HCM_PASSWORD="..." \
  -v "$PWD/config.toml:/app/config.toml:ro" \
  ghcr.io/mohantyajitesh/fusion-hcm-mcp-server:latest

Or build from source: docker build -t fusion-hcm-mcp-server .

For hosted HTTP transport, set transport.type = "http" (or HCM_TRANSPORT=http) and publish -p 8000:8000.

Local development

Requires Python 3.11+.

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
fusion-hcm-mcp-server          # runs the MCP server over stdio

Documentation

• docs/INSTALL_CLAUDE_DESKTOP.md — install & run with Claude Desktop (venv or Docker), verification sequence, troubleshooting traps.

• docs/TEST_CASES.md — tool matrix, automated coverage, the live-pod checklist, production-readiness assessment.

• DESIGN.md — full technical design: auth, the ADF REST client, exact tool signatures, the q= filter grammar, the safety model, ADF ground truth, and licensing/module alignment.

Status & contributions

This is an actively developing project — issues and PRs from people testing against their own pods are genuinely wanted. See CONTRIBUTING.md; for vulnerabilities, SECURITY.md. The design doc is the source of truth.

License

Apache-2.0. See also NOTICE.

Not affiliated with or endorsed by Oracle. Oracle and Fusion are trademarks of Oracle Corporation.