Connect Claude & AI tools

Marsad ships a remote MCP server at https://www.marsad.space/api/mcp. Connect it once and your AI assistant can list issues, read symbolicated stack traces, inspect performance traces and even resolve or assign issues — using your account, limited by the scopes and organization you approve.

The Marsad MCP server lets an AI assistant read your monitoring data directly. MCP — the Model Context Protocol — is an open standard for connecting assistants to real systems, and Marsad implements it as a remote server rather than a local process, so there is nothing to install. You add one HTTPS endpoint in Claude's connector settings, or run a single command in Claude Code, and authorise it over OAuth 2.1 with PKCE. On the consent screen you choose which organization the assistant can see and what it may do: read-only by default, with triage actions opt-in. It then has fourteen tools — listing issues, reading a symbolicated stack trace, opening a trace waterfall, querying logs, looking up affected users, and resolving, assigning or commenting on an issue. In practice that means you can ask what broke in production and get an answer grounded in your own events, instead of pasting stack traces into a chat window.

On this page

What you get

Fourteen tools covering errors, performance and analytics. Ask things like “what broke in the last release?”, “show me the stack trace for MAR-K7P2QX9M” or “resolve every issue from the checkout page”.

list_projectsOrganizations and projects the token can access
list_issuesFilter and sort grouped errors — status, level, release, environment, tag
get_issueOne issue with its symbolicated stack trace, recent events, affected users and comments
list_transactionsp50/p75/p95 duration, throughput and error counts per transaction
find_traces / get_traceLocate recent traces and read the full span waterfall
get_project_overviewHeadline health: events, active issues, crash-free rate, time series
list_releases / list_logsReleases seen in events; captured console logs
search_users / get_user_eventsFind an affected end-user and read their event timeline
update_issue_status / assign_issue / add_issue_commentTriage from chat (requires the write scope)

Connect Claude (web & desktop)

Claude connects via OAuth — no keys to copy.

  1. In Claude, open Settings → Connectors → Add custom connector.
  2. Enter https://www.marsad.space/api/mcp and save.
  3. Claude redirects you to Marsad: sign in, pick the organization to share, review the scopes and approve.

That's it — Marsad's tools appear in Claude's tool menu. You can revoke access any time from Profile → Connected apps.

Connect Claude Code

One command; Claude Code walks you through the same OAuth flow:

Terminal
bash
claude mcp add --transport http marsad https://www.marsad.space/api/mcp

Prefer a long-lived token (CI, shared machines)? Create one under Profile → API tokens and pass it as a header instead:

Terminal
bash
claude mcp add --transport http marsad https://www.marsad.space/api/mcp \
  --header "Authorization: Bearer mst_your_token_here"

Other MCP clients

Any client that speaks streamable-HTTP MCP works. OAuth-capable clients discover everything from the server metadata automatically; for the rest, send a personal access token as a Bearer header:

mcp.json
json
{
  "mcpServers": {
    "marsad": {
      "type": "http",
      "url": "https://www.marsad.space/api/mcp",
      "headers": { "Authorization": "Bearer mst_your_token_here" }
    }
  }
}

API tokens & scopes

Access is scoped. OAuth clients get the scopes you approve on the consent screen; personal access tokens get the scopes you pick at creation:

  • read:issues — issues, events, stack traces, logs
  • read:performance — transactions, traces, spans
  • read:analytics — overview stats, releases, sessions, users
  • write:issues — resolve, ignore, assign and comment (off by default)

Tokens are shown once at creation and stored hashed — treat them like passwords.

Security model

  • Standard OAuth 2.1: PKCE, dynamic client registration, short-lived access tokens with rotating refresh tokens.
  • An OAuth grant is pinned to the single organization you approve; projects outside it don't exist for that connection.
  • Every query runs under your user with row-level security — an AI client can never see more than you can.
  • Revoke any connection from Profile — it takes effect immediately.