Docs

Talaria MCP server

Hosted Talaria MCP server: docs tools, investigation, integration scopes, and authentication.

mcpagents

Hosted MCP at https://api.newtalaria.com/mcp. Transport is Streamable HTTP. Auth is OAuth (the client discovers /.well-known/oauth-protected-resource, registers at /oauth/register, and you approve consent) or a personal token from Workspace › Integrations. Install scopes are mcp:read mcp:write mcp:keys. A personal token lasts 30 days, 90 days, or until you revoke it. It is opaque base64url JSON (one segment, not a three-part JWT).

Warning

MCP uses OAuth or a personal token. API keys, publishable (tal_pk_…) or secret (tal_sk_…), never work on /mcp.

Never send a monitor token to MCP.

tools/list requires a grant. get_doc, list_docs, and search_docs tools/call still work without auth when called by name. After connect, call get_connection before mutating.

The dashboard button Copy Cursor mcp.json is the Cursor file. Other clients use a different key for the same URL. Pasting the Cursor file into Claude Code fails, because an entry with url and no type is skipped.

Clients

Client Config key Page
Cursor mcpServers → url Cursor
Grok Bot Custom Add MCP chat prompt or marketplace handle talaria Grok Bot
VS Code servers → type: http + url VS Code
Copilot CLI copilot plugin install No guide yet
Claude Code mcpServers → type: http + url Claude Code
Gemini CLI mcpServers → httpUrl Gemini CLI
Codex [mcp_servers.Talaria] → url Codex
Windsurf mcpServers → serverUrl Windsurf

Agent skills

The Talaria plugin for Cursor and Grok Bot includes five closed-loop skills that guide agents from install through fix verification:

Skill When to use
talaria-get-started Connect, confirm org/project, first smoke read (get_doc, list_docs, and search_docs or search_issues)
talaria-instrument SDK install, create project/key (setup_project), send first test event
talaria-observe-performance Health (get_project_stats), traces, spans, queries, sessions, monitors
talaria-triage-issue Investigate one issue (pack → trace/replay/suspects → optional GitHub issue)
talaria-verify-fix After release: verify_fix verdict, optional update_issue_status

Closed loop: get-started → instrument → observe → triage → verify-fix.

Skills are available from the Talaria plugin.

Agents can also follow the Markdown playbooks via get_doc: Add Talaria with an agent (≈ instrument), Triage issues with an agent (≈ triage). See Agent skills overview for the skills → playbooks map.

Tool catalogue

Public by name (no org auth on tools/call)

Tool Purpose
list_docs Manifest entries; filter by sdk / prefix
search_docs Ranked search over title, path, tags, headings, body
get_doc Full Markdown for a path/id

For install tasks, start with get_doc on agents/add-talaria.

Authenticated (mcp:read and above)

Tool Purpose
get_connection Announce signed-in user, org, and capabilities
switch_organization Move the grant to another org you belong to
list_projects List org projects
get_project Safe project settings (tracing/analytics/replay/heatmaps/sample rates) — no webhook secrets
search_issues / get_issue Grouped issues and detail, including comments on get_issue. get_issue also returns investigation: issue, latest event, trace summary, replay summary, source slice, and breadcrumbs. search_issues takes environment
get_replay Replay summary (url, duration, status, dashboard URL). No recording bytes
get_source_context Commit-exact GitHub source for one stack frame (projectId, eventId, frameIndex)
search_events / get_event Event instances. search_events takes environment and issueId. get_event returns the rewritten JavaScript frame when a source map matches
get_source_map Original source window for one stored frame (eventId, frameIndex). mapped: false includes the fileName still to upload
search_traces / get_trace Transactions / waterfalls. search_traces takes environment. Set grouped: true and sort (count, p95, impact, errorRate) for one row per transaction name with count, error count, p50, and p95
list_suspect_spans Child spans with the highest p95 in the window. Takes environment
list_database_queries Database statements grouped by text: calls, total time, p95, errors, and N+1 trace count. nPlusOneOnly keeps N+1 statements. Takes environment
search_sessions Product sessions. Takes environment
get_project_stats Health snapshot for one environment, plus per-environment counts. Takes environment
Feature flag tools List / get / kill-switch / percent rollout. One flag definition per project; evaluation uses the API key's environment

Elevated scopes

Tool Scope Purpose
update_project_settings / create_project / send_test_event mcp:write Safe settings, create project, ingest a test event
update_issue_status / add_issue_comment mcp:write Resolve, ignore, or reopen one issue (a note is required to resolve or ignore, and is stored as a comment), or add a comment without changing status
create_api_key / setup_project mcp:keys Mint a key once; never log or commit. Omit scopes for the publishable key (every publishable scope). scopes: ["releases:write"] is a release key for source-map upload. setup_project stays on the publishable key

Integration workflow

See Add Talaria with an agent, Triage issues with an agent, and the client pages above.

Environment on read tools

search_issues, search_events, search_traces, list_suspect_spans, list_database_queries, search_sessions, and get_project_stats take environment: development, test, staging, or production. The default is production when the project has a production key, otherwise development. An agent that just installed should query development. get_project_stats countsByEnvironment is analytics volume for each environment and stays 0 when analytics is off. Install success is search_issues or search_events.

setup_project mints a development key. Production shipping is a second key.

Investigation workflow

  1. list_projects → real projectId
  2. search_issues (pass environment; use development right after install, production for a live failure) → get_issue
  3. A JavaScript frame that is still a hashed *.js file needs a source map for that release. Follow Upload JavaScript source maps, then get_event or get_source_map.
  4. Optional get_trace / search_sessions in the same environment
  5. Fix locally → verify with search_events (issueId set) / search_issues on development
  6. When the issue has stopped firing, update_issue_status to resolved with a note. Noise is ignored with a note. A later event with the same fingerprint reopens a resolved issue. An ignored issue reopens when the project's reopen-ignored setting is on (the default). resolved and ignored can only return to open.