Connects your MCP host to Alephant's BYO-KEY platform for AI cost control across 50+ providers and 320+ models. Runs in two modes: Virtual Key mode exposes read-only cockpit usage, budgets, and recent requests (8 tools), while Manager mode with PAT credentials unlocks workspace-wide analytics, virtual key lifecycle operations, agent/member/department drill-downs, and cost anomaly detection (27 tools). Both modes surface daily costs, model breakdowns, and usage summaries. Reach for this when you need LLM spend visibility and guardrails inside Cursor, Claude Desktop, or any stdio-compatible client. Built for teams running distributed AI workloads who want FinOps metrics without leaving their editor.
Model Context Protocol server for Alephant BYO-KEY: FinOps metrics, virtual keys, and workspace analytics from Cursor, Claude Desktop, or any MCP host.
| Mode | Environment | Tools |
|---|---|---|
| VK | ALEPHANT_VIRTUAL_KEY | Cockpit-scoped usage + 3 VK tools (8 tools total incl. shared) |
| Manager | ALEPHANT_PAT + ALEPHANT_WORKSPACE_ID | Workspace-wide management (27 tools total incl. shared) |
PAT takes precedence when ALEPHANT_PAT is non-empty. If neither VK nor PAT is set, the process exits with an error (no mock data).
Required (both modes): ALEPHANT_API_BASE_URL
Optional: ALEPHANT_RATE_LIMIT_RPM (default 60, use 0 to disable client-side throttling)
npx troubleshootingFrom this repo’s root (alephant-mcp/): do not use npx -y @alephantai/mcp to “smoke test” the published package. npm treats the current directory as the local @alephantai/mcp project and does not link the root package’s bin into node_modules/.bin, so Windows then fails with 'alephant-mcp' is not recognized (or the Chinese CMD equivalent).
Use one of these instead while developing in the clone:
npm start
# or
node .\bin\alephant-mcp.js
To verify npx the same way end users do, run it from any other directory (e.g. the parent folder):
cd ..
npx -y @alephantai/mcp
From a normal project folder (after npm install @alephantai/mcp), you can also run:
node .\node_modules\@alephantai\mcp\bin\alephant-mcp.js
If npx -y @alephantai/mcp still fails outside the clone, try:
npx --yes --package=@alephantai/mcp alephant-mcp
Published packages 0.0.2+ include the bin/alephant-mcp.js shim for reliable npm bin resolution when installed from the registry.
Alephant MCP is a local stdio server. Use one server entry per credential scope. For multiple workspaces, create multiple entries with different names and environment variables.
Virtual Key (read-only / scoped cockpit):
{
"mcpServers": {
"alephant": {
"command": "npx",
"args": ["-y", "@alephantai/mcp"],
"env": {
"ALEPHANT_API_BASE_URL": "https://alephant.io",
"ALEPHANT_VIRTUAL_KEY": "vk-..."
}
}
}
}
Personal Access Token (manager):
{
"mcpServers": {
"alephant-workspace-a": {
"command": "npx",
"args": ["-y", "@alephantai/mcp"],
"env": {
"ALEPHANT_API_BASE_URL": "https://alephant.io",
"ALEPHANT_PAT": "pat_...",
"ALEPHANT_WORKSPACE_ID": "00000000-0000-0000-0000-000000000000"
}
}
}
}
Use separate mcpServers entries per workspace when you have multiple PATs.
Add a stdio MCP server in ~/.codex/config.toml:
[mcp_servers.alephant]
command = "npx"
args = ["-y", "@alephantai/mcp"]
env = {
ALEPHANT_API_BASE_URL = "https://alephant.io",
ALEPHANT_VIRTUAL_KEY = "vk-..."
}
startup_timeout_sec = 20
tool_timeout_sec = 120
For Manager mode, replace the VK env with:
env = {
ALEPHANT_API_BASE_URL = "https://alephant.io",
ALEPHANT_PAT = "pat_...",
ALEPHANT_WORKSPACE_ID = "00000000-0000-0000-0000-000000000000"
}
Verify with:
codex mcp list
codex mcp get alephant
Add a local MCP server under mcp in your OpenCode config, for example opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"alephant": {
"type": "local",
"command": ["npx", "-y", "@alephantai/mcp"],
"enabled": true,
"environment": {
"ALEPHANT_API_BASE_URL": "https://alephant.io",
"ALEPHANT_VIRTUAL_KEY": "vk-..."
}
}
}
}
For Manager mode, use ALEPHANT_PAT and ALEPHANT_WORKSPACE_ID in environment instead of ALEPHANT_VIRTUAL_KEY.
Use the CLI to add Alephant as a local stdio server:
claude mcp add-json alephant '{"type":"stdio","command":"npx","args":["-y","@alephantai/mcp"],"env":{"ALEPHANT_API_BASE_URL":"https://alephant.io","ALEPHANT_VIRTUAL_KEY":"vk-..."}}' --scope user
claude mcp list
claude mcp get alephant
For a shared project config, create .mcp.json in the project root:
{
"mcpServers": {
"alephant": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@alephantai/mcp"],
"env": {
"ALEPHANT_API_BASE_URL": "https://alephant.io",
"ALEPHANT_VIRTUAL_KEY": "vk-..."
}
}
}
}
For Manager mode, replace the VK env with ALEPHANT_PAT and ALEPHANT_WORKSPACE_ID.
npx --yes --package=@alephantai/mcp alephant-mcp --audit
scope + usage-summary (billing cycle).GET /api/v1/analytics/overview.Shared (both modes): check_alephant_connection, get_usage_summary, get_daily_costs, get_cost_by_model, list_available_models
VK only: get_my_scope, get_my_budget, get_my_recent_requests
Manager only: get_workspace_overview, get_workspace_budget_status, list_virtual_keys, create_virtual_key, update_key_budget, revoke_virtual_key, list_agents, get_agent_analytics, list_members, get_member_analytics, list_departments, get_department_analytics, get_subscription_info, set_budget_policy, get_live_24h, get_usage_timeseries, get_sparklines, diagnose_cost_anomaly, get_executive_dashboard, drill_down_spend, find_idle_resources, compare_entity_periods
get_request_logs is not included (JWT-only backend route).
/alephant-… 快捷模板、工具对照表docs/marketplace-submission.md for official MCP Registry, Smithery, and Glama submission notes.npm install
npm test
npm run build
Published as @alephantai/mcp (alephant-mcp binary).
ALEPHANT_API_BASE_URL*Alephant SaaS API base URL, for example https://alephant.io.
ALEPHANT_VIRTUAL_KEYsecretVirtual Key for read-only, cockpit-scoped VK mode. Use this mode for developer-level cost and usage visibility.
ALEPHANT_PATsecretPersonal Access Token for Manager mode. When set, Manager mode takes precedence over ALEPHANT_VIRTUAL_KEY.
ALEPHANT_WORKSPACE_IDWorkspace UUID required with ALEPHANT_PAT for Manager mode.
ALEPHANT_RATE_LIMIT_RPMOptional local HTTP tool-call limit per minute. Defaults to 60; set 0 to disable local throttling.