Changelog

​

SDK 0.6.1

• An aborted call that used an idempotency key (run, execution commands, and Admin API key issuance) tells you to recover with the original key, and keeps the response status and requestId.
• A ConnectionError raised after response headers arrive keeps status and requestId.

​

CLI 0.2.1

• Requests name the CLI and its version in the X-AgentMuxer-Client header, so support can find your calls.
• The update check runs at most once a day, including when you are offline, and still shows a newer release it already knows about.
• Uses SDK 0.6.1.

​

SDK 0.6.0

• The client is a class. Replace createClient() with new AgentMuxer() from @agentmuxer/sdk, and createAdminClient() with new AgentMuxerAdmin() from @agentmuxer/sdk/admin. The @agentmuxer/sdk/client entry is removed.
• Methods take (input, options?). Pass an abort signal as { signal } instead of a positional argument. A caller abort throws AgentMuxerError with kind "aborted".
• Refusal reasons, refusedBy, credentialKind, and balance status are open string types. Add a default branch to exhaustive switch statements.
• The framework adapters’ AgentMuxerOptions types are renamed to OpenAIAgentsOptions, ClaudeAgentOptions, and AISDKOptions.

• run needs only a provider, path, and inputs. method is optional when the path serves one method, and the maximum cost defaults to $1.00. Pass maxCost: "0.25" as a USD string instead of maxCostMicros.
• search, inspect, run, getInvocation, accountStatus, and execution handle reads retry lost connections, timeouts, rate limits, and gateway errors twice by default, honoring Retry-After. run resends the same idempotency key, so a retry never charges twice.
• Set timeout (100 seconds by default) and maxRetries per client or per call, and pass a custom fetch to either client.
• New ConnectionError, ConnectionTimeoutError, RateLimitError, and RefusalError classes, and throwOnRefusal to throw for any result other than responded.
• accountStatus(), inspect with view: "call_ready", requestId on results, and require() support for the root, /contracts, and /admin entries.
• Validation errors name the failing field.

​

CLI 0.2.0

• install run by an agent or without a terminal no longer signs in the CLI. It waits up to 330 seconds for browser approval, exits 0, and prints each remaining step as an Owner action: line. Run agentmuxer login before using API commands there, and status to confirm setup.
• With no coding agent detected, install exits 1 and points to SKILL.md.
• status exits 0 when a coding agent is set up and the CLI is not signed in, printing CLI: not signed in. Scripts that call the API should check for that line or run agentmuxer login.
• A flag given to a command it doesn’t apply to now fails with an error that names it.

• Agents can run setup themselves. install asks no questions, configures every detected coding agent, starts sign-in for the agent it runs in, and hands the remaining steps to you. --no-cli-login skips only the CLI sign-in.
• run PROVIDER PATH needs only the provider, path, and inputs. --method is optional when the path serves one method, and the maximum cost defaults to $1.00. run --query takes a JSON object of query parameters.
• inspect --view call_ready shows the compact request view.
• API commands retry lost connections, timeouts, rate limits, and gateway errors up to twice with the same idempotency key, so a retry never charges twice.
• agentmuxer help, fuller help for every API command, and a once-a-day notice when a newer release is available. Set AGENTMUXER_NO_UPDATE_CHECK=1 to turn the notice off.
• Licensed under MIT. Uses SDK 0.6.0.

​

One-line agent setup

• Paste Set up AgentMuxer from https://agentmuxer.com/SKILL.md into your coding agent. It runs setup, then relays the sign-in steps for you to finish in the same conversation.

​

Simpler calls

• maxCostMicros is optional on call_offering. Each call’s maximum defaults to $1.00, and maxCost takes a USD string such as "0.25" as an alternative to micros.
• method is optional on call_offering and inspect_offering when the path serves one method. When a path serves several, the response lists the methods to choose from.
• Invalid arguments name the failing field.

​

SDK 0.5.0

• New @agentmuxer/sdk/admin entry for the Admin API, with typed methods for the account balance, applications (create or get by your own external id, list, update, cap, disable), application keys (create, rotate with a grace period, revoke), and usage. SDK 0.6.0 renames createAdminClient to new AgentMuxerAdmin().
• The package includes the Admin API’s OpenAPI description (admin-api.openapi.json) and a runnable platform integration example.
• Results for application keys omit the account-wide balance (billing.balanceMicros), and account status reports balance: null. The CLI sign-in still shows your balance.
• Responses from search, inspect, run, getInvocation, and execution handles accept fields added by newer servers.
• An invalid baseUrl throws AgentMuxerError with kind invalid_request.

​

CLI 0.1.3

• agentmuxer status accepts account fields added by newer servers. Uses SDK 0.5.0.

​

Team accounts

• Invite teammates from the dashboard’s Members page. Members use the account’s credits with their own agents, and admins also manage applications and members. Invitation links work for the invited email address and stay valid for 7 days.

​

Call-ready inspection

• inspect_offering returns a compact, call-ready view of each request: the fields, types, and locations your agent needs for a valid first call.

​

Admin API

• Manage your account from your servers at https://api.agentmuxer.com/v1/admin with an admin key created on the dashboard’s Applications page. Create one application per customer company, cap its spending, issue and rotate its keys, and read usage.
• Admin keys manage applications, and application keys make calls. All applications draw on your account’s balance.

​

SDK 0.4.0

• Execution handles (client.execution(...)) attach to an existing call to wait for its result, stream its progress, follow a provider job that continues after its first response, resume a pending approval, and stop following.
• The partition option separates calls and results per end user within one application key. Derive it from your own authenticated users.
• The Vercel AI SDK connection exposes AgentMuxer’s instructions, validates purchase arguments before the tool runs so the model can correct them, and accepts a trusted server-side mcpUrl.
• The initialization signal and deadline also cover tool discovery.

​

SDK 0.3.1

• Type-checking fix for the Vercel AI SDK adapter. Runtime behavior and public interfaces are unchanged.

​

CLI 0.1.1 and 0.1.2

• Simpler installed guidance: agents call search_offerings without extra options.
• Uses SDK 0.3.1 and 0.4.0.

​

CLI 0.1.0

• install, status, and remove set up AgentMuxer’s hosted MCP server and guidance in Codex, Claude Code, and OpenCode, using each client’s own sign-in.
• login and logout sign the CLI in through your browser and keep credentials in the OS keyring.
• search, inspect, run, and get call AgentMuxer from a terminal or script.
• To move from setup, run npx -y @agentmuxer/cli@latest install.

​

SDK 0.3.0

• accessToken is accepted instead of apiKey for an already-issued AgentMuxer access token.
• New @agentmuxer/sdk/contracts entry with the request types and validators.

​

Call APIs by provider and path

• search_offerings returns operations as provider, method, and path addresses. inspect_offering and call_offering take that address, with path values in pathParams and inputs in query, headers, cookies, or body.
• Call a known operation directly, without searching first.
• New get_invocation tool recovers a lost response or checks a pending call by invocationId or the original idempotencyKey, at no charge.

​

Search ranking

• search_offerings ranks results by how closely each operation fits your query.

​

SDK 0.2.1

• Vercel AI SDK transport errors use the MCP client’s error classes.
• Manual purchase confirmation can pause and resume on a later authenticated request, keeping the original purchase and tool call.
• Vercel AI SDK tool calls allow 100 seconds.

​

SDK 0.2.0

• A direct API client with no framework dependency, addressing operations by provider, method, and path.
• Automatic recovery keys on every call, and explicit original keys for accepting changed terms.
• Typed discovery, inspection, invocation results, and read-only invocation lookup.
• An agentmuxer command with JSON, file, and stdin input. It moved to @agentmuxer/cli on September 23.
• Framework adapter imports and authentication are compatible with 0.1.0.

​

Payment confirmation per application

• Choose on the dashboard’s Applications page whether each application’s paid calls follow the account default, require confirmation, or run unattended within your spending limits. See billing and approvals.

​

Structured pricing

• Search results show structured pricing for each operation: a fixed total, a unit rate, an observed estimate, or an unknown price.

​

SDK 0.1.0

• Connect OpenAI Agents, Claude Agent SDK, and Vercel AI SDK to AgentMuxer.
• Authenticate with an application key through AGENTMUXER_API_KEY or apiKey.
• Give agents discovery and purchase tools, subject to balance, spending limits, and payment confirmation settings.
• Use TypeScript and ESM on Node.js 22.18 or later.

​

Setup 0.1.4 to 0.1.6

• Setup detects which supported clients are installed and adds AgentMuxer guidance to each one’s global instructions.
• Concurrent setup runs wait their turn, and each configuration file is updated atomically.
• remove cleans up configuration left behind after you uninstall a client.

​

Price confirmation through your agent

• When a call’s price is above your maximum, call_offering returns the exact terms. To accept them, repeat the call with the same arguments plus acceptTerms from nextAction. See the MCP tool reference.

​

Setup 0.1.2 and 0.1.3

• Setup keeps your existing sign-ins and client registrations, including OpenCode credentials after an access token expires.
• status recognizes Claude Code’s connected status.

​

Sign-in on connect

• https://mcp.agentmuxer.com/mcp asks for sign-in when a client first connects, so discovery and paid calls both run under your account. Clients that support MCP authorization start the browser sign-in for you.

​

Setup 0.1.1

• Setup treats sign-in as a required step and guides you through it.

​

Setup 0.1.0

• install, status, and remove connect Codex, Claude Code, and OpenCode to AgentMuxer’s hosted MCP server. See connect your agent.
• Setup preserves your existing client settings and rolls back cleanly if a step fails.