# AgentMuxer

> AgentMuxer is a tool marketplace for agents: search a catalog of tools and
> call them within your price limit through one managed MCP connection.
> Buyers hold closed-loop USD credits, tracked in integer micros. One account, one balance,
> every provider through the same connection.

Availability note (2026-09-03): live: https://mcp.agentmuxer.com/mcp serves production (bound 2026-09-01; public sign-up is open).

## What AgentMuxer is

- The AgentMuxer MCP server at https://mcp.agentmuxer.com/mcp exposing 5 buyer tools.
- A catalog with an honest trust ladder: Benchmarked > Verified > Indexed.
  - Benchmarked: Has a benchmark designation. See the benchmark evidence.
  - Verified (API value `verified`): Passed request checks. This does not guarantee successful calls or payments.
  - Indexed: No current verification.
- Launch categories: web search, scraping, extraction, enrichment (verified and indexed tiers).
- Managed fulfilment: the gateway makes the provider call and pays the provider
  itself. Provider credentials never reach you; you only ever hold USD credits.
- no minimum listing price; sub-cent listings exist.

## Discover, inspect, call, or recover

1. `search_offerings` - Find API operations when you do not know which one to use.
2. `inspect_offering` - Read request documentation and current pricing when inputs are unclear.
3. `call_offering` - Call provider and path directly within a maximum charge ($1.00 unless set).
4. `get_invocation` - Read an invocation by ID or original key without calling the provider again.
5. `report_outcome` - Optionally report how useful the result was.

Search when you do not know which operation fits. Inspect when its inputs are
unclear. If you know the provider, path and inputs (plus the method when a path
serves several), call directly.
Use get_invocation to recover a lost response or read progress without invoking
the provider again. Reporting feedback is optional. Search orders results by relevance, not price.

## Billability - when you are and are not charged

You are charged:

- At most once per invocation, at the authorized price.
- Only for a 2xx provider response; nothing else bills.

You are never charged for:

- Marketplace failures (anything AgentMuxer itself gets wrong).
- Every other provider response: any non-2xx status - not-found (404),
  malformed requests (400, 422), redirects, every 5xx - plus network errors
  and timeouts.
- Idempotent replays: the stored response returns without re-execution.

Refusals are machine-actionable: `refusedBy` is one of auth, policy, balance,
budget, cap, resolution, request, listing. An insufficient-credits refusal
carries `nextAction {kind: "top_up", url}`; relay that link to
the owner, who adds credits there. One transport-level exception:
any MCP request with no credential answers as a standard HTTP 401 challenge that starts the host OAuth flow - not a refusal envelope.

## Discovery pricing and payment authorization

- Discovery describes fixed prices, unit rates, conditional prices, estimates, or unknown pricing in buyer credits. Estimates are not spending ceilings.
- Fixed-price calls retain the accepted price. Native-payment calls obtain actual
  terms from the provider and enforce the call's maximum (`maxCostMicros` or
  `maxCost` in USD; $1.00 when unset) before payment. Discovery
  estimates and unknown prices are descriptive, never permission to spend.
  Account-required confirmation still applies. Accepted invocation terms survive
  later catalog changes; supported native payment amounts pass through at par; discovery rates and estimates are advisory.
- Credits are closed-loop USD credits, tracked in integer micros. There is a
  $10 signup credit for all new accounts while the launch pool lasts. Card top-ups ($10 minimum card top-up)
  carry a 5% funding fee ($0.80 minimum) added on top at checkout — the selected amount is always credited in full, and credit only after payment
  confirmation.

## Quickstart

Set up AgentMuxer from https://agentmuxer.com/SKILL.md: it has the setup steps for every host
and the owner handoff. For Codex, Claude Code, and OpenCode with a local
shell, run this installer exactly:

    npx -y @agentmuxer/cli@latest install --no-cli-login

Allow it at least six minutes; it can wait for the owner's browser approval.
Do not substitute native MCP-add commands: they omit the managed guidance and
verified removal. That guidance tells the agent to call directly when it knows
the operation and to search, which is free, when it does not; the agent still
chooses native versus specialized tooling by fit, simplicity, and cost.

The installer configures every detected host, registers the hosted MCP
globally, adds one reversible guidance block, and starts sign-in for the
agent's own host when it can. Relay each line it prints that starts with
`Owner action:` to the owner verbatim. Hosts without a local shell add https://mcp.agentmuxer.com/mcp as a connector.
SKILL.md has the owner steps that follow the installer and the recipe for
each other host; the owner's manual recipes are at https://agentmuxer.com/setup.md.

Authentication: sign-in is required for free and paid MCP use. Your MCP host
starts the OAuth flow against the remote endpoint and the owner approves it;
never complete a sign-in or consent screen yourself. Every request requires an
account; the first exchange from a fresh install answers the 401 challenge
that starts that flow.
Before paid calls, the owner should add credits as needed. A spending limit for
each agent is recommended but optional; without one, an agent can spend up to
the available balance.

## After setup

After setup, relay these next steps and the starter prompt to the owner.
Do not run the starter prompt as part of installation.

Follow the steps your agent relays, in order: they keep this conversation and include approving AgentMuxer sign-in. Then paste this prompt. After a terminal setup, start your coding agent and paste it there. For connector hosts, start a new conversation.

> Use AgentMuxer to find tools that could help with this project. Recommend one, explain what it does and its price, and don't execute a paid call.

## Docs

- [Setup](https://agentmuxer.com/setup.md): copy-paste install recipes for the owner.
- [Agent skill](https://agentmuxer.com/SKILL.md): self-setup skill for agents, with steps for every host.
- [HTTP reference](https://agentmuxer.com/http.md): call https://api.agentmuxer.com directly from your own code.
- [Terms](https://agentmuxer.com/terms.md)
- [Privacy](https://agentmuxer.com/privacy.md)
- [Security](https://agentmuxer.com/security.md)
- [Acceptable use](https://agentmuxer.com/acceptable-use.md)