What is OE MCP Server?
A standalone binary that implements the Model Context Protocol (MCP) and exposes your enterprise data sources as tools that AI apps can use directly. No code. Define connectors in a single JSON file.
- 45+ connector categories — PostgreSQL, MongoDB, S3, GitHub, Slack, Gmail, SSH, REST API, and more
- Two transport modes —
--stdio for Claude Code / Cursor / Windsurf; --serve for cloud or team deployments
- Persistent memory —
memory_set / memory_get / memory_list / memory_delete survive across sessions
- Action log — every connector call logged automatically with timestamp, tool, input, and result
- Run AI agents —
run_agent executes any OE Runtime SKILL.md agent directly from Claude Code, Cursor, or any MCP client. Manual skills pause for approval via approve_chain.
- Self-hosted — runs on your own machine, no cloud dependency, no call-home
Quick Start via npm (Recommended)
1. Create oe-mcp.json:
{
"connectors": [
{
"name": "my-postgres",
"type": "postgresql",
"host": "localhost",
"port": 5432,
"database": "mydb",
"user": "postgres",
"password": "secret"
},
{
"name": "my-codebase",
"type": "filesystem",
"basePath": "/home/user/projects/myapp"
}
],
"memory": [
{ "key": "project_context", "value": "This is our main application." }
]
}
2. Add to your AI app's MCP config:
macOS / Linux:
{
"mcpServers": {
"oe-mcp": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@openenthrium/oe-mcp", "--stdio", "/path/to/oe-mcp.json"]
}
}
}
Windows:
{
"mcpServers": {
"oe-mcp": {
"type": "stdio",
"command": "npx.cmd",
"args": ["-y", "@openenthrium/oe-mcp", "--stdio", "C:\\path\\to\\oe-mcp.json"]
}
}
}
-y is required — without it npx blocks waiting for keyboard input and the MCP connection never opens.
3. Reload your AI app — connectors appear as tools automatically.
Ask Claude: "What connectors do you have access to?" to verify.
Download (Standalone Binary)
Memory
Persistent memory that survives restarts — stored in oe-mcp-memory.json:
| Tool | Description |
|---|
memory_set | Store a key-value pair across sessions |
memory_get | Retrieve a stored value by key |
memory_list | List all stored key-value pairs |
memory_delete | Remove a stored key |
"Remember that our production database is on prod-db.company.com" → Claude calls memory_set
Action Log
Every connector call is logged automatically to oe-mcp-log.json:
| Tool | Description |
|---|
log_list | List recent connector calls (newest first, supports limit) |
log_clear | Clear all log entries |
Run AI Agents
Execute OE Runtime SKILL.md agents or YAML agents directly from Claude Code, Cursor, or any MCP client — no terminal required:
| Tool | Description |
|---|
run_agent | Run an agent by file path. Auto skills execute immediately; manual skills pause and return pending_skill_chain. |
list_pending_skills | List all manual skills currently paused and waiting for approval |
approve_chain | Approve, skip, or abort a paused manual skill by chain_id |
run_agent parameters:
| Parameter | Required | Description |
|---|
file | ✅ | Absolute path to agent.yaml |
params | ❌ | Key-value pairs substituted via {{key}} in the agent |
input | ❌ | Optional initial message passed to the agent |
approve_chain parameters:
| Parameter | Required | Description |
|---|
chain_id | ✅ | From pending_skill_chain.chain_id in a run_agent response |
approved | ❌ | true to run the skill (default), false to skip it and continue |
abort | ❌ | true to stop the entire pipeline immediately |
OE MCP looks for oe-config.json in the agent's directory first, then falls back to oe-mcp.json.
Agent Skill Approval Flow
When an agent's skill pipeline includes manual skills, Claude handles the approval loop automatically:
- Claude calls
run_agent → response shows ⏸ Skill awaiting approval with chain_id and skill_name
- Claude decides — based on your instructions — whether to approve, skip, or abort
- Claude calls
approve_chain → next skill runs or the next manual skill pauses again
- Repeat until
Pipeline complete or Claude aborts
Example instruction to Claude Code: "Run the OE Skills orchestrator and send a Slack message — skip anything you can't do, abort if it asks for credentials."
Transport Modes
| Mode | Flag | Best for |
|---|
| stdio | --stdio | Claude Code, Cursor, Windsurf, Codex, Claude Desktop — launched as child process |
| HTTP | --serve --port 4040 | Cloud deployments, sharing one server across a team |
HTTP mode — start the server, then add the URL to Cursor / Windsurf / Claude Desktop:
oe-mcp-linux --serve --port 4040 /path/to/oe-mcp.json
Sample Configs
Download oe-mcp-samples.zip — ready-to-use oe-mcp.json for common connectors:
postgres · mysql · mongodb · github · slack · gdrive · ssh · filesystem · oracle · salesforce · servicenow · telegram · notion · confluence · graphql · zoho-mail · sftp · dropbox · multi-connector
Part of Open Enthrium
Contributing
→ See CONTRIBUTING.md for how to add sample configs and connector adapters.
License
Apache-2.0 — free to use, modify, and deploy for any purpose, including commercial use.
No usage limits. No telemetry. No call-home.
