Give your agent one connection to hosted tools for web research, SEO, public data, media jobs and supported account actions. ToolRouter exposes MCP, a CLI and an installable skill so the assistant can discover the operation it needs, run it and return the finished result.
Connect your agent · Install the skill · Browse the live catalog
The catalog contains 266 tools and 1,380 operations. Availability, price and account requirements come from the live operation schema, not a provider logo or an installed skill.
Which agents can use ToolRouter?
These 13 clients document a remote MCP setup path. Open the guide for exact configuration, authentication, a first task, examples and troubleshooting. Finish the advertised sign-in or account flow and check a real result to verify your connection.
| Agent and guide | Made by | Connection | Authentication |
|---|---|---|---|
| Claude Code | Anthropic | Terminal MCP + skill | OAuth or returned account flow |
| Claude Cowork | Anthropic | Custom connector | OAuth or returned account flow |
| Claude | Anthropic | Custom connector | OAuth or returned account flow |
| ChatGPT | OpenAI | Developer-mode remote MCP | OAuth or returned account flow |
| Codex | OpenAI | Terminal MCP + skill | OAuth or returned account flow |
| Cursor | Cursor | Project or user MCP config | OAuth or returned account flow |
| Gemini CLI | Terminal MCP | OAuth or returned account flow | |
| Grok Build | xAI | Terminal MCP | OAuth or returned account flow |
| OpenClaw | Open source | Remote MCP + skill | OAuth or returned account flow |
| Kimi Code | Moonshot AI | Remote MCP config + skill | OAuth or returned account flow |
| Hermes Agent | Nous Research | Remote MCP config + skill | OAuth or returned account flow |
| Paperclip | Open source | Custom MCP connector | OAuth or returned account flow |
| Perplexity Computer | Perplexity | Custom remote connector | OAuth or returned account flow |
Seven more clients: current compatibility
These clients have different plugin, connector or container requirements. The links below lead to official documentation; a product marketplace listing or universal OAuth setup is not confirmed here.
| Client | Current setup boundary |
|---|---|
| OpenAI dots | Dots uses plugins from the connected ChatGPT account. Use the ChatGPT or Codex guide for its supported tools; a separate product plugin for dots is not confirmed. |
| Grok Bot | Grok Bot uses its own plugin marketplace and saved workflows. Use the Grok Build guide for terminal MCP; Bot plugin availability needs to be checked in your account. |
| DeepSeek Harness | HTTP MCP and local skills are documented. Browser OAuth is not documented here; operators need a supported credential flow or a configured CLI runtime. |
| NanoClaw | Setup is managed by the operator of the group and container. Follow the runtime documentation for skill mounts, MCP plugins and credentials; there is no one-command connection guide here. |
| Meta Muse | Muse supports custom API connectors. Use the service API documentation and supported account credentials; generic remote MCP and local skill import are not documented here. |
| Manus | Manus supports custom API or MCP connectors and single-skill imports. A skill package needs SKILL.md at its root. Check the authentication options available in your account. |
| Cue | Cue has separate connector and skill settings. Skills imported into Manus do not transfer to Cue; check the options in your Cue account before connecting. |
Already using another compatible client? Start with the MCP server reference, or use the CLI from a documented shell runtime.
How do agents connect?
| Surface | What it supplies | Authentication | Guide |
|---|---|---|---|
| Remote MCP | Tool discovery and execution in a compatible client | Client OAuth or advertised account flow | MCP server |
| Skill | SKILL.md workflow plus provider resources | Does not authenticate the account | Install the skill |
| CLI | Terminal discovery, execution and JSON output | Configured hosted account or the documented local mode | CLI guide |
| REST | Direct integration for your application | Supported API credentials | API reference |
1. Add the remote server
Paste the endpoint into a compatible remote HTTP MCP client, or use its exact command from the guide:
https://api.toolrouter.com/mcpFor example, Codex uses:
codex mcp add toolrouter --url https://api.toolrouter.com/mcp
codex mcp login toolrouterClaude chat and Cowork guides use the Anthropic-specific profile at https://api.toolrouter.com/mcp/anthropic, whose catalog excludes standalone AI image, video and synthetic audio generation. Inspect the tools returned to your client.
2. Add the workflow skill
For a supported skill-capable runtime:
pnpm dlx skills add Humanleap/agent-skills --skill toolrouterSelect the supported agent and read the installed SKILL.md. A local file does not automatically become a cloud connector or approved marketplace plugin. Authentication and provider-account access remain separate.
3. Start with a real, read-only result
Use ToolRouter to discover a free operation for current weather in London, run it and show the tool, skill and every field returned. Do not buy credits or change an account.
A saved connector proves configuration. A successful provider result proves that operation worked for this account. If a setup or claim link is returned, use it to retain the account and unlock the access you need.
What can the agent do?
| Job | The useful result | Workflow |
|---|---|---|
| Web and company research | Dated source-linked records, with coverage gaps | Research and extraction |
| SEO and competitor research | Provider, market, units, page evidence and original sources | SEO workflow |
| Image, video and audio work | Completed asset URLs or transcripts, after checking job status | Media jobs |
| X and LinkedIn publishing | The authorised post's provider result, ID or URL | Social publishing |
| Economic and public data | Original registry or statistical source, dates, units and geography | Provider index |
| A local file or pipeline | Actual output saved in the selected runtime | CLI · Files and media |
Research example
Research the company names I provide. Use ToolRouter discovery for public company data and original web sources, inspect requirements and price, then return legal identity, finding, source URL and available date. Flag ambiguous identities and failed sources. Do not contact anyone.
SEO example
Discover the available page-audit, keyword and backlink operations for my site and named competitors. Show provider, target market, schema and price before paid calls. Separate measured results from recommendations, and do not infer traffic or revenue from backlink coverage.
Publishing example
Draft an X or LinkedIn post from the announcement I provide. Discover the exact posting operation, verify the connected account and show the final text and any price before I approve publication.
The current catalog includes X text and image posting and LinkedIn text, link and image posting. Instagram, TikTok and YouTube reading or scraping operations do not establish publishing support. An MCP connection does not create a social scheduler.
The workflow: discover, inspect, execute, finish
| Operation | Purpose |
|---|---|
discover | Find a tool and skill for the requested outcome, with schema and requirements |
use_tool | Run the selected operation with schema-matching input |
credits_balance | Inspect available credits when paid work needs them |
job_get | Resolve asynchronous work when the response advertises a job handle |
Built-in tools are called directly. Inspect the current tools/list before relying on a built-in name. The common pattern is: connect, discover, inspect schema and cost, execute, resolve queued work, then return sources or completed outputs.
Do not repeat an uncertain write before inspecting its status. A queued job is unfinished; a provider response and a completed downstream business outcome are separate evidence.
Which providers and public sources are covered?
The public provider index currently lists 55 entries, covering search and extraction, SEO intelligence, AI media, company registries and economic data. These pages describe integrations; the live catalog determines which operations are available to the connected account.
Bureau of Labor Statistics · European Central Bank · Bank of England · Bank of Canada · Federal Reserve Bank of New York · US Energy Information Administration · GLEIF · SEC EDGAR · World Bank · Open-Meteo · Open Food Facts · Companies House · NSW FuelCheck · WA FuelWatch · Service Victoria Servo Saver · PetrolSpy · fal.ai · ElevenLabs · Serper · Firecrawl · Exa · Parallel AI · ScrapeCreators · Prodia · Higgsfield · Runway · Perplexity · Photalabs · Google · Mapillary · OpenRouter · DataForSEO · HubSpot · Sentry · Zendesk · Cloudflare · PostHog · Microsoft 365 · Klaviyo · OpenAI · Anthropic · TypeSafe · DeepL · RapidAPI · Duffel · Foursquare · Google Maps · Ticketmaster · VirusTotal · FRED · TMDb · Discogs · Brandfetch · RentCast · Stripe
Explore provider coverage · Check the live tool catalog
Can an agent repeat a task automatically?
The connection supplies tools, not a background scheduler. Use the client's documented automation capability or a persistent runtime you control. Define the schedule, inputs, chosen provider, connected account, approved budget and expected output before the first run.
For example: repeat a bounded public-page extraction each weekday, retain the previous result, report changed fields with source URLs, and stop on a provider or credit error. Check a previous queued job before starting a replacement. Do not infer automatic posting permission from permission to draft.
For source-linked news monitoring after the chat ends, SentryDock's agent guides describe a separate monitor account and signed callback workflow.
Frequently asked questions
Do I need a separate connection for every provider?
ToolRouter provides one discovery and execution connection across its catalog. Some operations use hosted provider access; account actions can still require your own upstream account. One connection does not bypass provider permissions.
Is the skill enough to run tools?
No. The skill teaches the workflow. MCP, CLI or REST supplies the execution path and account access. Install the skill in a client that documents skill support.
Does every agent see every operation?
No. Client-specific profiles and account access can limit the catalog. The Claude connector profile filters standalone AI media generation. Use the returned schema as the source of truth.
What does it cost?
Free operations do not require purchasing credits. Paid operations have their own price and charging unit. Inspect the discovered price, credits and approved budget before a batch. See pricing and billing.
How do I know the assistant actually called a tool?
Ask for the operation name and returned fields, original source links or completed asset URL. A plausible answer or visible connector label does not establish a provider call.
Can it publish or message from my accounts?
Only through an available operation, the correctly connected upstream account and authority for the actual action. A request to research or draft does not authorise publication, messaging or account changes.
Are these approved marketplace installs?
The setup guides here use documented custom MCP or runtime methods. Marketplace acceptance is a separate distribution step; an installable public skill does not establish approval in a client's catalog.
Where do I troubleshoot?
Open your client guide for login and configuration checks. Then inspect the selected operation's requirements, billing or files and media. Report provider errors rather than hiding them behind a generic retry.
Start with one useful task
Choose your client · Install ToolRouter's skill · Browse tools
Setup methods and catalog operation names reviewed on 3 October 2026. Examples are task templates; they do not claim completed authenticated runs in every client.