Openrouter Admin Mcp — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Openrouter Admin Mcp (Agent Skill) and scored it 45/100 (orange). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 1 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
A base64 string of 128+ characters appears in a documentation file. Encoded prompt injection hides the hostile instruction in base64 — invisible to keyword filters — and relies on the agent's ability to decode it at runtime. There is no normal authoring reason to embed a multi-hundred-byte base64 blob in skill docs.
*.sig, SIGNATURES) outside the documentation.Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
MCP server for the OpenRouter management API — programmatic control of credits, inference keys, guardrails, organization members, and usage analytics from inside Claude Code, Claude Desktop, Cursor, or any MCP-compatible client.
demo
Read-only by default. Destructive write operations are gated behind an opt-in env flag.
This is not an inference proxy. It uses an OpenRouter Provisioning API key, which can manage your account but cannot make completion calls.
OpenRouter's UI is solid for one-off tweaks, but checking spend across keys, drilling into per-model/per-day usage, configuring account-wide guardrails, or rotating limits requires a lot of clicking. This server exposes those operations as MCP tools — Claude can answer "what burned my credits this week?", "create a temporary key with a $0.30 daily cap", or "bind the new prod key to our daily-$50 guardrail" in a single turn.
26 tools wrapping the OpenRouter management API:
limit_reset (daily/weekly/monthly) and expires_at.by_model, by_day, by_provider, or by_key. Token-level breakdown (prompt / completion / reasoning) per row.gen-… id (cost, tokens, latency, finish reason).Plumbing: HTTP retry on 429/5xx with Retry-After support, in-session GET cache, write invalidation. Returns raw fields — interpretation is left to the agent.
Replace the placeholder sk-or-v1-... with your real key after install.Run via npx (no install needed) or install globally:
npx -y openrouter-admin-mcp
# or
npm install -g openrouter-admin-mcpCreate a Provisioning key at <https://openrouter.ai/settings/provisioning>, then add the server to your MCP client config.
Claude Code (~/.claude.json) or Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"openrouter-admin": {
"command": "npx",
"args": ["-y", "openrouter-admin-mcp"],
"env": {
"OPENROUTER_PROVISIONING_KEY": "sk-or-v1-..."
}
}
}
}Restart your client to load the server. By default this gives read-only access — to enable key creation, mutation, and deletion, see Enabling write tools below.
Destructive operations (or_key_create, or_key_update, or_key_delete) are disabled by default to prevent accidental key mutation through prompt-injected tool calls. Add OPENROUTER_ADMIN_ALLOW_WRITE to your env block to enable them:
"env": {
"OPENROUTER_PROVISIONING_KEY": "sk-or-v1-...",
"OPENROUTER_ADMIN_ALLOW_WRITE": "1"
}The server logs which mode it started in:
openrouter-admin-mcp running on stdio (write tools: ENABLED)Read tools (or_overview, or_credits, or_current_key, or_keys_list, or_key_get, or_activity) are always available regardless of this flag.
| Tool | Purpose |
|---|---|
or_overview | One-shot dashboard: credits + active keys + today's burn by model. |
or_credits | Account balance: total purchased, total used, remaining. |
or_current_key | Metadata of the Provisioning key the server is using. |
or_keys_list | All inference keys with usage, limits, reset cadence, expiration. |
or_key_get | Detailed view of one key by hash. |
or_activity | Usage for the last 30 UTC days. Filters: date, api_key_hash, user_id. Aggregations: none, by_model, by_day, by_provider, by_key. |
or_generation | Fetch a single inference call by its gen-… id (cost, tokens, latency, finish reason). |
or_models | Model catalog with per-1M pricing, context length, modalities. Filterable. |
or_model_get | Full details for one model id. |
or_model_endpoints | Per-provider endpoints for a model: pricing, uptime, status. |
or_models_user | Models filtered by your account's privacy/guardrail settings. |
or_zdr_endpoints | ZDR-compliant endpoints for privacy-constrained workflows. |
or_guardrails_list | All guardrails: limit_usd, reset, allow/block lists, ZDR. |
or_guardrail_get | Full details for one guardrail. |
or_guardrails_assignments | Composite — which keys/members are bound to which guardrail. |
or_org_members | Members of the OpenRouter organization (Management-key + org account). |
OPENROUTER_ADMIN_ALLOW_WRITE=1)| Tool | Purpose |
|---|---|
or_key_create | Create a new inference key. Returns the one-time secret. |
or_key_update | Update name / disabled / limit / limit_reset / expires_at. |
or_key_delete | Permanently delete a key. |
or_guardrail_create | Create a guardrail. |
or_guardrail_update | Update a guardrail. |
or_guardrail_delete | Delete a guardrail. |
or_guardrail_assign_keys / _unassign_keys | Bulk-(un)assign a guardrail to inference keys. |
or_guardrail_assign_members / _unassign_members | Bulk-(un)assign a guardrail to org members. |
> Create a key called "scratch" with a $0.50 daily limit that expires in 30 days.
[Claude calls or_key_create with:
name="scratch",
limit=0.50,
limit_reset="daily",
expires_at="2026-05-25T00:00:00Z"]
Created key **scratch**
Hash: f7a3...
Limit: $0.5 (resets daily)
Expires: 2026-05-25T00:00:00Z
Secret: `sk-or-v1-...`
⚠️ This secret cannot be retrieved later — store it now.> Why did my OpenRouter spend jump yesterday?
[Claude calls or_activity with date="2026-04-24", aggregate="by_model"]
Activity: 47 rows | total $0.8570 | 305 req
By model (top 5):
- xiaomi/mimo-v2-flash $0.5240 | 178 req | 24 rows
- anthropic/claude-haiku-4.5 $0.2105 | 39 req | 5 rows
- ...> Cap our prod-bot key at the daily-$50 guardrail.
[Claude calls or_guardrails_list, finds gr_prod_daily_50,
then or_guardrail_assign_keys with id="gr_prod_daily_50", hashes=["…"]]
Assigned guardrail `gr_prod_daily_50` to 1 key(s).The server returns raw API fields. It does not flag what's "near limit", what's an "anomaly", or what the agent should do — that's the agent's job. Composite tools (or_overview, or_guardrails_assignments) merge multiple endpoints into one workflow response, but they don't interpret. Rationale: production MCP servers (Stripe, GitHub, Linear) follow the same pattern.
OPENROUTER_PROVISIONING_KEY environment variable and forwarded only to https://openrouter.ai/api/v1/* over HTTPS./credits, /key, /keys, /keys/{hash}) is held in memory only and dropped on shutdown.git clone https://github.com/oviron/openrouter-admin-mcp.git
cd openrouter-admin-mcp
npm install
npm run build # compile TypeScript + chmod +x build/index.js
npm test # vitest, 84 unit tests, mocked fetchProject layout:
src/
client.ts # OpenRouterClient — fetch wrapper, retry, in-session cache
index.ts # MCP server entry — registers tool groups
tools/
credits.ts # or_credits, or_current_key
keys.ts # or_keys_list, _get, _create, _update, _delete
activity.ts # or_activity (5 modes incl. by_key)
overview.ts # or_overview composite
generation.ts # or_generation
models.ts # or_models, _model_get, _model_endpoints, _models_user, _zdr_endpoints
guardrails.ts # or_guardrails_list/_get/CRUD/bulk-assign + composite assignments
organization.ts # or_org_members
test/
client.test.ts, credits.test.ts, keys.test.ts,
activity.test.ts, overview.test.ts, generation.test.ts,
models.test.ts, guardrails.test.ts, organization.test.ts,
registration.test.ts, helpers.ts, smoke.test.tsMIT — see LICENSE.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.