EdgeGate MCP Server
Set up edge-AI regression gates from Claude Code, Cursor, or Claude Desktop. Zero clicks. Just a prompt.
What it does
10 tools, 4 skills — one stdio MCP server that plugs into any MCP-capable AI client
Authenticates with your workspace API key (egk_live_*) generated in Settings
Drive the whole flow from a Claude prompt: setup_workspace → create_pipeline → run_gate → check_status → setup_github_action
Prerequisites
- workspace_premiumAn EdgeGate workspace on Pro tier or above. The Free tier cannot create API keys. API keys are workspace-scoped — one key per integration.
- terminalNode ≥ 20 — required to run
npx edgegate-mcp-installand the MCP server itself. - computerA supported MCP client: Claude Code, Cursor, or Claude Desktop.
Quick Start
From zero to a working MCP gate in 4 steps.
Generate an API key
- Log in to your EdgeGate Dashboard
- Open your Workspace → Settings → API Keys
- Click Generate new key
- Give it a name (e.g.
claude-code-local) and confirm
One-time secret
egk_live_…) appears exactly once. Copy it immediately — you cannot retrieve it again. If you lose it, revoke and regenerate.Install the MCP server
Claude Code (recommended): install the plugin in one command. This registers the MCP server plus all 6 slash commands (/edgegate-init, /edgegate-status, /edgegate-compare, etc.):
/plugin install https://github.com/frozo-ai/edgegate-mcpType the command in Claude Code, then enter your EDGEGATE_API_KEY when prompted. Restart Claude Code to activate.
Cursor or Claude Desktop: use the interactive installer — it auto-detects your client, prompts for the API key, and writes the config file:
npx edgegate-mcp-installRestart your client when the installer finishes.
Manual config (optional)
If you prefer to edit config files yourself, add the following JSON block to the appropriate file for your client:
{
"mcpServers": {
"edgegate": {
"type": "stdio",
"command": "npx",
"args": ["-y", "edgegate-mcp"],
"env": {
"EDGEGATE_API_KEY": "egk_live_...",
"EDGEGATE_API_URL": "https://edgegateapi.frozo.ai"
}
}
}
}{
"mcpServers": {
"edgegate": {
"command": "npx",
"args": ["-y", "edgegate-mcp"],
"env": {
"EDGEGATE_API_KEY": "egk_live_...",
"EDGEGATE_API_URL": "https://edgegateapi.frozo.ai"
}
}
}
}{
"mcpServers": {
"edgegate": {
"command": "npx",
"args": ["-y", "edgegate-mcp"],
"env": {
"EDGEGATE_API_KEY": "egk_live_...",
"EDGEGATE_API_URL": "https://edgegateapi.frozo.ai"
}
}
}
}egk_live_... with the key you copied in Step 1. The Cursor and Claude Desktop configs omit the type field; Claude Code requires it set to "stdio".Restart your MCP client
Fully quit and relaunch your MCP client (Claude Code, Cursor, or Claude Desktop). The edgegate server must start at launch time — a hot-reload is not enough. After restart, type /tools (Claude Code) or open the tool picker (Cursor) to confirm you see the 7 edgegate tools listed.
Try the killer prompt
Paste this into Claude and hit enter:
Use the edgegate MCP to set up a CI gate for my MobileNet ONNX model.
Gates: inference_time_ms ≤ 10, peak_memory_mb ≤ 150.
Devices: Galaxy S24, Galaxy S23.Claude will call the 7 tools in sequence: edgegate_setup_workspace to confirm your workspace → ask you for the model artifact ID (or have you upload the .onnx file first) → edgegate_create_pipeline with your gate definitions → edgegate_run_gate to trigger a run on Snapdragon hardware → poll with edgegate_check_status and, when complete, offer to call edgegate_setup_github_action to wire the whole thing into your GitHub workflow.
The 10 tools — quick reference
All tools require EDGEGATE_API_KEY to be set in the MCP server's environment.
| Tool | What it does | Typical use |
|---|---|---|
| edgegate_setup_workspace | Confirm or list workspaces | First call in any conversation |
| edgegate_create_pipeline | Define a new regression pipeline with devices, gates, and models | One-time setup per model family |
| edgegate_run_gate | Trigger a run against an existing pipeline | Per-PR or ad-hoc testing |
| edgegate_check_status | Poll a run for status, per-device metrics, and gate pass/fail | Waiting for a run to complete |
| edgegate_get_report | List recent runs in a workspace (up to 50) | Audit, triage, trend analysis |
| edgegate_get_audit_report | Fetch the signed audit PDF URL for a completed run | Compliance evidence bundles |
| edgegate_setup_github_action | Generate GitHub Actions YAML + gh secret set commands | Wiring CI on the first gate |
| edgegate_connect_huggingface | Store a personal HuggingFace token on a workspace (validated via whoami first) | Importing private / gated / Qualcomm-org repos |
| edgegate_get_huggingface_integration | Check whether a workspace has a connected HF token and its status | Verifying connection before an import |
| edgegate_disconnect_huggingface | Remove the stored HuggingFace token from a workspace | Rotating accounts or clearing access |
Detailed input/output schema for every tool is in the full tool reference.
The 4 skills
Slash commands that bundle the tools into high-level workflows. Type them directly in Claude Code or Cursor.
/edgegate-init— Full onboardingZero → CI gate in one flow. Confirms your workspace, creates the pipeline, triggers the first run, and offers to generate the GitHub Action — no hand-holding required.
/edgegate-gate— Trigger a runTrigger a run on an existing pipeline. Accepts a pipeline ID or lets Claude select one interactively. Polls until terminal status and prints the per-device metric table.
/edgegate-status— Check run statusCheck a run's current status and metrics. Pass a run ID or let Claude find the most recent run in your workspace.
/edgegate-audit— Fetch audit PDFFetch the signed audit PDF URL for a completed run. The PDF contains the Ed25519-signed evidence bundle — sha256 manifest, metrics, device list, and gate verdicts.
Common problems
warningTools don't appear in Claude Code
- Fully quit and relaunch Claude Code (not just a tab close)
- Open
~/.claude.jsonand confirm theedgegateblock is present undermcpServers - Verify
EDGEGATE_API_KEYis set to a non-empty value inside theenvblock (not as a shell export — it must be in the JSON) - Run
node --version— must be 20 or above
error401 — Invalid or revoked API key
Your key is either incorrect or has been revoked. Go to Settings → API Keys, revoke the old key, and generate a new one. Update the EDGEGATE_API_KEY value in your MCP client config and restart.
error402 — Pro tier required
API key creation and MCP usage require a Pro workspace. The Free tier is limited to the dashboard UI only. Upgrade at /pricing.
error409 — Concurrent run already in flight
Each workspace allows exactly one active run at a time (workspace_concurrency=1). Wait for the current run to reach a terminal state (passed, failed, or error) before triggering another. Use edgegate_check_status or /edgegate-status to poll.
Links & references
npm package
npmjs.com/package/edgegate-mcp
GitHub repository
github.com/frozo-ai/edgegate-mcp
Full tool reference
Detailed input/output schema for every tool
CI/CD Integration Guide
HMAC auth + GitHub Actions workflow
Ready to gate your models?
Create your EdgeGate workspace, generate an API key, and be running on-device regression tests in under 5 minutes.