AI agents

Connect Claude Code

Let Claude Code, or any other AI agent that speaks MCP, work on your disputes: read them, gather evidence, answer requests and take decisions you ask for. You connect it with one command and approve it in your browser. It acts as you, never beyond your role, and every action shows up in your workspace's activity log.

Set up in one command

Terminal
claude mcp add --transport http dispute-markets https://mcp.dispute.markets/mcp
  1. Run the command above.
  2. In Claude Code, type /mcp, pick dispute-markets and choose Authenticate. (Without the interactive menu: claude mcp login dispute-markets.)
  3. Your browser opens dispute.markets. Sign in as usual (with your second factor if you use one).
  4. On the consent screen, tick the workspaces the agent may use and the access it gets, then press Allow. The browser hands back to Claude Code.

The consent screen shows Claude Code as verified: it identifies itself with a client metadata document published by Anthropic at claude.ai, which our sign-in service checks. Any other agent that registers itself is shown with a warning that nobody has verified who made it; allow it only if you started the connection yourself.

Then just ask, for example: “Which disputes need a response this week?” or “Draft a response to dsp_… and attach the delivery confirmation as evidence.”

What the agent can do

Every tool takes an optional workspace_id; without it the agent works in the connection's default workspace (the one you were in when you allowed it).

ToolWhat it doesNeeds
list_workspacesYour workspaces this connection may use, their environment (Production, Staging…), your role and what the agent may do there.(any)
list_disputesDisputes, newest first; filter by status, search text, unanswered, dates.disputes:read
get_disputeOne dispute with its transaction, evidence, deadlines and timeline.disputes:read
get_evidence_download_urlA short-lived download link for an evidence file (audited as a sensitive read).disputes:read
search_transactionsList or search payments.disputes:read
get_transactionOne payment and any dispute on it.disputes:read
add_dispute_noteAdd a note to a dispute’s timeline.disputes:write
submit_evidenceAttach a file (text or base64, up to 2 MB) as evidence; the server uploads it.disputes:write
respond_to_rfiAnswer a request for information.disputes:write
challenge_disputeSubmit the response and contest the dispute.disputes:write
accept_disputeConcede the dispute. Cannot be undone; marked destructive, so your client asks you first.disputes:write
list_audit_eventsThe workspace’s activity log, with filters.audit:read
get_audit_eventOne audit event.audit:read

Never available to an agent: escalating a dispute to arbitration (it costs a fee and cannot be undone, so it stays a decision you make in the portal or through the API), bulk actions, playbooks, reports, webhooks, flows and the dev panel, people and invites, synthetic accounts, API keys, workspace settings and deletion, and CSV or audit exports. The server refuses those routes for an agent even if a client tried to call them.

Access, roles and workspaces

Access you allowMeans
disputes:readRead disputes, transactions and evidence. Always included.
disputes:writeAdd notes, upload evidence, respond to requests, accept or challenge.
audit:readRead the activity log (owners only).

What the agent may do is the access you allowed andyour current role in that workspace, checked on every call. A read-only member's agent never writes, even if write access was ticked. If you are demoted or removed from a workspace, your agent loses it at the same moment. A workspace you join later is not included: connect again to add it.

Synthetic accounts cannot connect agents. The consent screen refuses them, they have no AI agents page, and the MCP server refuses any connection that belongs to one. Connect your agent as yourself; it then acts with your role.

Everything is in the activity log

Each tool call writes an agent.tool_call event, and each action writes its usual event (dispute.accept, evidence.create, …). They carry source: "mcp" and actor.via, the agent and connection, so the log reads “dana@acme.com via Claude Code”. Connecting and disconnecting are recorded as agent_connection.create and agent_connection.revoke. Filter the audit events with source=mcp.

Disconnect

In the portal, Settings > AI agents lists your connected agents with their workspaces, access and last use. Disconnect stops one within seconds. Connections also end after 90 days; then just authenticate again.

Workspace ownersalso see, on the same page, every member's agents that can act in the workspace: who connected it, which agent, the access allowed and when it was last used. An owner can disconnect any of them. Disconnecting ends that connection in every workspace it covers, and it is recorded as agent_connection.revoke with reason: "owner".

Details for other clients and the curious

Endpoints

MCP server (streamable HTTP)https://mcp.dispute.markets/mcp
Protected resource metadatahttps://mcp.dispute.markets/.well-known/oauth-protected-resource/mcp
Authorization server (issuer)https://api.auth.productcraft.co/dispute-markets

Any client that implements the MCP authorization spec (2025-06-18 or later) works: OAuth 2.1 with PKCE (S256), client ID metadata documents or dynamic client registration, resource indicators, and refresh tokens that rotate on every use. Redirect URIs must be https or http on localhost. Tool calls are limited to 120 reads and 30 writes a minute per connection.