ToolRouter exposes a remote MCP server for discovery and execution across the hosted catalog. Instead of setting up a separate MCP server for each provider, your agent connects once and discovers the operation it needs.
The endpoint and setup
https://api.toolrouter.com/mcpUse the client-specific setup guide for the right command or configuration file. The client must support remote HTTP MCP and the advertised authentication flow. Complete OAuth or the account setup flow shown by the client and tool response; keep tokens in the client credential store.
Claude chat and Cowork use https://api.toolrouter.com/mcp/anthropic in their guides. That profile filters standalone AI image, video and synthetic audio generation. Inspect the catalog returned to your client instead of assuming every profile exposes the same operations.
Discovery keeps the initial tool surface small
The connection supplies gateway operations such as discover and use_tool. Discovery returns the specific tool and skill, schema and requirements for the job. You do not need to paste hundreds of provider definitions into the conversation before choosing an operation.
Example discovery input:
{
"query": "extract the title, headings and links from a public website"
}Use the returned tool and skill with their exact input schema. Tool names and fields can change; a provider logo or documentation page does not prove that the account has an active operation.
The core workflow
| Step | Tool or evidence | What to inspect |
|---|---|---|
| Discover | discover | Outcome, returned tool and skill, input schema |
| Check access and price | Returned requirements; credits_balance when available | Account connection, approved budget, charging unit |
| Execute | use_tool | Exact tool, skill and schema-matching input |
| Resolve queued work | Returned status handle, such as job_get | Completion, errors, output assets |
| Return the result | Sources, completed files or provider action result | Dates, coverage gaps, action identity |
Built-in MCP tools are called directly. Inspect the current tools/list before relying on a built-in name. Discovery is not a provider call, and a queued job is not finished output.
Accounts, pricing and provider access
Free operations can be used without buying credits. Retaining the account, paid work and connected-account actions require the access advertised by the operation. Follow the returned setup or claim link when needed. Read billing and pricing before high-volume work.
A supported X operation does not grant access to a particular X account. A public-data integration and a social write have different credential and permission requirements. Check the chosen provider and actual action scope before execution.
Does one connection run recurring work?
The MCP connection exposes tools; it does not keep an assistant or scheduler running. Use the client's documented automation capability or your own runtime for repeated jobs, with an explicit schedule, scope, account and budget. Reconcile a previous job before repeating an uncertain write.
Troubleshooting
- No tools: check remote HTTP transport, per-conversation enablement and the team allowlist.
- Auth failure: reconnect with the advertised OAuth flow or account setup link; do not paste browser access tokens into config.
- Paid call refused: inspect the price and credit requirement; a free result does not prove paid access.
- Provider action refused: check the exact upstream account and operation requirement.
- Timeout after a write: inspect status or provider readback before retrying.