omium-mcp 0.1.1

Official Model Context Protocol (MCP) server for the Omium platform.

Omium MCP Server

Official Model Context Protocol server for the Omium platform. Packaged as omium-mcp on PyPI and also runnable as a Docker container. Routes all tool calls through Kong using the caller's Omium API key for auth and tenant scoping.

Two transports from one package:

• stdio (default) — for end-users installing locally via pip/uvx. Zero-config; the API key is read from $OMIUM_API_KEY.
• Streamable HTTP (omium-mcp serve) — for self-hosted / team deployments. Per-request Authorization: Bearer ... header.

Naming note: the mcp-control-plane component inside Omium-platform/ is a Managed Control Plane (a Go service), unrelated to the Model Context Protocol. This project is the MCP-protocol flavor and lives outside the platform repo.

Quickstart

Get from pip install to a working integration in under 5 minutes.

Prerequisites

• Python 3.10 or newer (python3 --version on macOS/Linux, python --version on Windows)
• An Omium API key — copy it from your Omium dashboard

1. Install

macOS / Linux:

python3 -m venv ~/omium
~/omium/bin/pip install --upgrade pip
~/omium/bin/pip install omium-mcp
export PATH="$HOME/omium/bin:$PATH"

Windows (PowerShell):

python -m venv $HOME\omium
& "$HOME\omium\Scripts\pip" install --upgrade pip
& "$HOME\omium\Scripts\pip" install omium-mcp
$env:PATH = "$HOME\omium\Scripts;$env:PATH"

Tip: pip install omium (the SDK) bundles omium-mcp automatically — install either one.

2. Set your API key

macOS / Linux:

read -rs OMIUM_API_KEY && export OMIUM_API_KEY
echo "key length: ${#OMIUM_API_KEY}"   # sanity check; doesn't print the key

Windows (PowerShell):

$secure = Read-Host "OMIUM_API_KEY" -AsSecureString
$env:OMIUM_API_KEY = [System.Net.NetworkCredential]::new("", $secure).Password
"key length: $($env:OMIUM_API_KEY.Length)"

3. Wire it into your AI client

Claude Code

macOS / Linux:

claude mcp add omium omium-mcp --env OMIUM_API_KEY="$OMIUM_API_KEY"
claude mcp list   # should show "omium"

Windows (PowerShell):

claude mcp add omium omium-mcp --env "OMIUM_API_KEY=$env:OMIUM_API_KEY"
claude mcp list

Inside Claude Code, ask: "Show me my Omium workflows and dashboard metrics." Claude will pick the right tool and render the result.

Claude Desktop

Edit your Claude Desktop config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "omium": {
      "command": "omium-mcp",
      "env": {
        "OMIUM_API_KEY": "omium_..."
      }
    }
  }
}

Replace omium_... with your actual key, then fully restart Claude Desktop.

Cursor

Add an MCP server in Cursor's settings: command omium-mcp, env var OMIUM_API_KEY. Exact UI varies by Cursor version — see Cursor's MCP docs.

Troubleshooting

• pip install fails with Requires-Python >=3.10 — Python 3.10+ is required because the upstream Anthropic mcp package needs 3.10+. Install via pyenv / uv / your OS package manager.
• All tool calls return 401 — OMIUM_API_KEY is wrong or unset. Confirm with echo "len: ${#OMIUM_API_KEY}" (macOS/Linux) — should be > 30.
• Pointing at a non-prod backend — set OMIUM_API_BASE (e.g. http://localhost:8000) before starting the MCP.
• AI client doesn't see Omium tools — fully restart the client after editing config (not just close the window).

Self-hosted (Streamable HTTP / Docker)

For team deployments behind a load balancer:

cd /path/to/omium-MCP
docker compose up -d --build

Server listens on http://localhost:9100/mcp. Wire into Claude Code:

claude mcp add --transport http omium http://localhost:9100/mcp \
  --header "Authorization: Bearer omium_YOUR_KEY_HERE"

---

Tools

The MCP exposes every Kong-reachable Omium endpoint that accepts X-API-Key (~84 tools). JWT-only endpoints (dashboard login/signup, invitations, Slack OAuth, Stripe webhooks) are intentionally excluded — an MCP client only holds an API key. All tools forward the caller's bearer token to Kong as X-API-Key; tenant scope is derived server-side so no tenant_id argument is ever accepted.

The complete client-facing endpoint list is maintained in ../Kong-Client-Facing-APIs.md.

Categories

Category	Example tools	Upstream service
Identity	verify_api_key	auth-service
Workflows	list_workflows, get_workflows, list_workflow_versions	auth-service
Executions	list_executions, create_execution, execute_execution, replay_execution, rollback_execution, apply_fix_to_execution, compare_executions, delete_execution, update_execution_status, get_apply_to_repo_payload	execution-engine
Checkpoints	list_checkpoints, list_all_checkpoints, get_checkpoint, create_checkpoint	execution-engine
Failures	list_failures, get_failures_stats, get_failures_time_series, create_failure_event	execution-engine
Scores	create_score, list_scores, get_scores_stats	auth-service
Traces	ingest_trace, list_traces, get_trace, list_trace_failures, list_trace_projects	auth-service
Projects	create_project, list_projects, connect_project_git, list_project_files, save_project_file, commit_project_git	auth-service
GitHub	github_status, github_setup, github_update_repo, github_disconnect, github_create_fix_pr	auth-service
Recovery	list_recovery_failures, trigger_recovery, create_recovery_command, list_recovery_commands, get_recovery_command, update_recovery_command_status, redeliver_recovery_command	recovery-orchestrator
Replay	get_replay_state, get_replay_step, get_replay_consensus, get_replay_diff, restart_replay	recovery-orchestrator
Analytics	get_usage_summary, get_dashboard_metrics, get_recent_activity, get_performance_metrics, get_performance_time_series, get_performance_agents, get_workflow_performance, get_workflow_cost, get_system_metrics	analytics-engine
Audit	list_audit_logs, search_audit_logs, get_audit_log	audit-logger
Billing	get_billing_balance, get_billing_usage, create_billing_topup, list_billing_transactions, get_subscription_status, estimate_execution_cost, get_cost_breakdown, get_quotas, get_billing_forecast, get_billing_recommendations, get_cost_analytics, list_billing_alerts, …	billing-service

Notable conveniences

• create_execution auto-fills agent_id as mcp-default-<tenant-slug> when omitted.
• execute_execution auto-resolves workflow_type and workflow_definition from the execution's workflow_id.
• Tools that take complex bodies (replay_execution, rollback_execution, apply_fix_to_execution, compare_executions, project/recovery writes, billing checkouts, ingest_trace, audit creation, etc.) accept a pass-through body: dict — the docstring enumerates typical fields. This keeps the MCP layer useful while upstream schemas stabilize.
• Non-JSON upstream responses are wrapped as {"ok": true, "text": "..."}.
• On non-2xx upstream responses, tools raise Omium API <METHOD> <path> -> <status>: <body> so the upstream error is visible to the LLM.

Auth model

Both transports populate the same _api_key ContextVar; tools read it and forward the value as X-API-Key to Kong.

Transport	How the key is bound
stdio	$OMIUM_API_KEY read once at process start (init_from_env)
HTTP	Authorization: Bearer ... extracted per request by BearerAuthMiddleware

Missing / non-bearer auth on HTTP returns 401. Missing $OMIUM_API_KEY for stdio raises a clear startup error.

Architecture

  Claude Code / Desktop
        │
        │  stdio (subprocess)  OR  HTTP (Bearer header)
        ▼
  omium-mcp  (package: omium_mcp)
        │  HTTP  (X-API-Key: omium_...)
        ▼
    Kong  (api.omium.ai  or  http://kong:8000 in Docker)
        │
        ▼
  workflow-manager / execution-engine / auth-service / ...

Stop with:

docker compose down

Smoke-test

Reject missing auth:

curl -i http://localhost:9100/mcp
# HTTP/1.1 401 Unauthorized

Full MCP round-trip (uses the Python MCP client from .venv):

python - <<'PY'
import asyncio, json
from mcp.client.session import ClientSession
from mcp.client.streamable_http import streamablehttp_client

KEY = "omium_..."  # any seeded key from .mcp-test-keys.txt

async def main():
    headers = {"Authorization": f"Bearer {KEY}"}
    async with streamablehttp_client("http://localhost:9100/mcp", headers=headers) as (r, w, _):
        async with ClientSession(r, w) as s:
            await s.initialize()
            print([t.name for t in (await s.list_tools()).tools])
            res = await s.call_tool("list_workflows", {})
            print(res.content[0].text[:200])

asyncio.run(main())
PY

Wire into Claude Code

claude mcp add --transport http omium http://localhost:9100/mcp \
  --header "Authorization: Bearer omium_YOUR_KEY_HERE"

Then /mcp in Claude Code lists the server and the two tools are callable.

Wire into Claude Desktop

Edit ~/.config/Claude/claude_desktop_config.json (Linux):

{
  "mcpServers": {
    "omium": {
      "type": "http",
      "url": "http://localhost:9100/mcp",
      "headers": {
        "Authorization": "Bearer omium_YOUR_KEY_HERE"
      }
    }
  }
}

Restart Claude Desktop.

Configuration

Env var	Default	Purpose
OMIUM_API_KEY	—	Required for stdio transport. Ignored by HTTP (per-request bearer).
OMIUM_API_BASE	https://api.omium.ai	Upstream Kong base URL. Overridden to http://kong:8000 in Docker.
MCP_HOST	0.0.0.0	Bind address for HTTP transport.
MCP_PORT	9100	Bind port for HTTP transport.

Package layout

omium_mcp/
├── __init__.py
├── config.py          # env vars
├── auth.py            # ContextVar, BearerAuthMiddleware, init_from_env()
├── http.py            # omium_get / omium_post / omium_patch / omium_delete + _parse()
├── tenant.py          # tenant-slug cache (for agent_id defaulting)
├── mcp_instance.py    # FastMCP("omium-mcp") singleton
├── cli.py             # entry point — stdio (default) or `serve` for HTTP
└── tools/             # 14 modules, one per API category
    ├── identity.py
    ├── workflows.py
    ├── executions.py
    ├── checkpoints.py
    ├── failures.py
    ├── scores.py
    ├── traces.py
    ├── projects.py
    ├── github.py
    ├── recovery.py
    ├── replay.py
    ├── analytics.py
    ├── audit.py
    └── billing.py

Adding a new tool

1. Pick the right omium_mcp/tools/<category>.py (or add a new one and import it from tools/__init__.py).
2. Write an async def decorated with @mcp.tool(). The docstring becomes the tool description the LLM reads — keep it accurate.
3. Call the appropriate helper:
   • omium_get(path, params=...) for GET
   • omium_post(path, json_body=..., params=...) for POST
   • omium_patch(path, json_body=..., params=...) for PATCH
   • omium_delete(path) for DELETE (handles 204 No Content)
4. Rebuild:
   • Docker (HTTP): docker compose up -d --build
   • Local stdio test: .venv/bin/python -m omium_mcp.cli (reads $OMIUM_API_KEY)
5. Test via the harness: .venv/bin/python scripts/test_all_tools.py.

Publishing to PyPI

.venv/bin/python -m build          # writes dist/omium_mcp-X.Y.Z-py3-none-any.whl + tar.gz
.venv/bin/twine upload dist/*      # needs PyPI credentials

Users can then install with pip install omium-mcp or run without installing via uvx omium-mcp.

Docs

• docs/transport.md — why this MCP uses HTTP instead of stdio, and what that implies for configuration.