Agent Platforms and Enterprise Control MCP

Connect Agent Platforms and Enterprise Control MCP

Token360 supports two distinct Agent integrations. Use an inference key for model calls. Enterprise owners can additionally use a scoped control key with the hosted Control MCP to manage sub-accounts, their model access, and their inference keys.

Never use a control key as a model provider key. Control keys are rejected by inference endpoints and should be granted only the scopes an automation needs.

Personal accounts: connect the model router

Create a normal API key in API Keys, store it in a secret manager or environment variable, and configure the Agent platform as an OpenAI-compatible provider:

Provider typeOpenAI compatible
Base URLhttps://api.token360.ai/v1
API keyYour inference key (sk-...)
ModelA model name visible in the Token360 model catalog

For an Agent that can edit its own configuration, give it this instruction:

text
1Add an OpenAI-compatible provider named token360.
2Read the API key from TOKEN360_API_KEY; do not write the secret into source control.
3Use https://api.token360.ai/v1 as the base URL.
4Keep the model name configurable and verify the setup by listing models or sending a minimal chat request.

Generic provider configuration:

JSON
1{
2  "provider": "openai-compatible",
3  "baseUrl": "https://api.token360.ai/v1",
4  "apiKeyEnv": "TOKEN360_API_KEY",
5  "model": "your-token360-model-name"
6}

Platforms label these fields differently, but the three required values are always the base URL, inference key, and model name. When a platform offers separate protocol choices, select the endpoint your Agent supports:

  • OpenAI Chat Completions: /v1/chat/completions
  • OpenAI Responses: /v1/responses
  • Anthropic Messages: /v1/messages
  • Gemini native: /v1beta/models/{model}:generateContent

Enterprise accounts: connect the hosted Control MCP

An enterprise owner first creates a Management or Provisioning key with an authenticated user session and selects the minimum scopes required by the automation. The plaintext control key is returned only once:

Shell
1curl https://api.token360.ai/enterprise/control-keys \
2  -H "Authorization: Bearer <enterprise-owner-session-token>" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "name": "provisioning-agent",
6    "keyType": "PROVISIONING",
7    "scopes": [
8      "tenant:members:read", "tenant:members:write",
9      "tenant:keys:read", "tenant:keys:write"
10    ]
11  }'

Then configure a remote Streamable HTTP MCP server:

MCP URLhttps://api.token360.ai/enterprise/mcp
TransportStreamable HTTP
AuthenticationAuthorization: Bearer <control-key>

Generic MCP configuration:

JSON
1{
2  "mcpServers": {
3    "token360-enterprise": {
4      "url": "https://api.token360.ai/enterprise/mcp",
5      "headers": {
6        "Authorization": "Bearer ${TOKEN360_CONTROL_KEY}"
7      }
8    }
9  }
10}

Use the Agent platform's secret store or environment-variable interpolation. If it does not support secret references in remote MCP headers, use its secure credential UI instead of committing the key to a configuration file.

Tools and scopes

list_sub_accountstenant:members:read
create_sub_account, set_sub_account_status, delete_sub_accounttenant:members:write
get_sub_account_model_accesstenant:policies:read
set_sub_account_model_accesstenant:policies:write
list_sub_account_api_keystenant:keys:read
create_sub_account_api_key, update_sub_account_api_key, delete_sub_account_api_keytenant:keys:write

The create-key and create-sub-account tools may return a one-time secret. The Agent must save it directly to the intended secret store and must not repeat it in chat, logs, or tickets. Delete tools are marked destructive; require human confirmation before allowing an autonomous Agent to call them.

Verify your connection

  1. Connect the Agent platform to the hosted MCP using your scoped control key.
  2. Call list_sub_accounts with the tenant:members:read scope to confirm access.
  3. If a tool returns API_KEY_SCOPE_DENIED, review the key's scopes and grant only those needed by your automation.
  4. Use an inference key for a minimal model request to verify model access separately.

Security recommendations

  • Prefer a Provisioning key for account/key lifecycle automation; use a Management key only when policy or read-only billing/audit scopes are needed.
  • Set an expiry and an IP/CIDR allowlist for unattended automation.
  • Separate credentials per Agent and environment so each can be revoked independently.
  • Rotate a suspected leaked key immediately. Token360 never returns an existing plaintext key again.
Was this page helpful?
Agent Platforms and Enterprise Control MCP · Token360