How it works
Every agent uses the same hosted MCP server. Only two things differ: how the agent signs in, and whose credits it spends. On each call, Locus checks that the tool is allowed and within your limits, runs it, and charges credits.
All three use this Streamable HTTP server:
MCP server URL
Connect an MCP client
Interactive OAuth (recommended)
Add the MCP server URL to your client and approve the connection in your browser. Your client stores and refreshes the tokens. Connect → Set up in your workspace shows the steps for each client, and Connect an MCP client covers Claude, ChatGPT, Codex, Cursor, Cline, and others.Install the Locus plugin
The Locus plugin bundles the OAuth connection with three skills that teach your agent to set up Locus, use tools, and build Workflows. Install the plugin or a manual connection, not both.Give your coding agent the Locus skills
Send your agent paywithlocus.com/SKILL.md. Or, if you don’t use the plugin, install the skills yourself:Install the skills
locus, locus-setup, and locus-workflows when prompted. Skills still need an authenticated MCP connection.
Agent keys
An agent key (lcac_…) lets one agent runtime call tools within limits you set, without you there. It can’t manage your workspace. The API and SDK call it an Agent Connection. Agent keys has the full guide.
Create an agent key
Open Connect → API keys, select Create key, and choose An agent. Pick its tools, account, spending limits, and expiry. Locus shows the key once, with a ready-to-use MCP configuration. See Create an agent key.Agent key configuration
Store the key asLOCUS_AGENT_CONNECTION and pass it as a bearer token in your client’s MCP settings. See client configuration examples.
Agent key scope and budgets
Each key has a per-call limit and an optional per-run budget shared by every call with the same loop ID. See Agent key scope and budgets.Rotate or revoke an agent key
Rotation issues a new key with the same scope and stops the old one at once. Revocation is immediate and permanent. See Rotate or revoke an agent key.Use the MCP tools
Ask your agent for an outcome. It finds the right tool, checks the price, and uses the result. Price caps, approvals, and safe retries keep each paid call under control; see Paid tool calls.Turn structured results into final answers
Store a paid result before the model summarizes it, so a failed answer never repeats the purchase. See Turn structured results into final answers.Sign in without a browser
Agent-owned headless OAuth
An agent that created its own account can approve its own MCP client sign-in, with the credential it received at signup or with AgentID. See Agent-owned headless OAuth.Headless device authorization (RFC 8628)
A client with no redirect listener can use the device-code grant: it shows a verification link, a person approves, and the client receives its tokens. See Headless device authorization.Troubleshoot a connection
Most connection errors come down to sign-in, scope, balance, or budget. Troubleshoot an agent connection explains each error and its fix.Next steps
Connect an MCP client
Set up Claude, ChatGPT, Codex, Cursor, and other clients.
Agent keys
Give an unattended agent a scoped key and budget.
Framework integrations
Connect your model SDK or agent framework.
Locus CLI
Find, price, and call tools from a terminal.