superagnt

Debug an AI agent from Claude Code: observability over MCP

Read what your deployed agents actually did, from the coding agent you already use: sessions, tool calls and their errors, schedule and webhook runs, escalations and spend, all over the same MCP connection. Your coding agent finds the failing call in three reads.

// platform tool

Read-only. Every tool in this family reads; none of them changes an agent, a session or a setting, which is why it is safe to hand to a debugging agent.

updated Nov 3, 2026

Copy the setup prompt.
Your agent does the rest.

works with any agent+ 8 more
Install superagnt for me. It's the superagnt plugin from the github.com/superagnt/plugins marketplace, plus the superagnt MCP server (on Claude it's a connector I add in my browser; elsewhere I sign in through my browser). The setup guide for each kind of agent is at https://www.superagnt.com/install
or follow the detailed setup instructions

» the observability tools are on every server, on every plan.

01// what is inside

What can your agent do with agent observability?

Find the bad sessions

search_sessions filters by text, status, tool name or problems only, so the sessions that hit an error come first.

An index, then the events

read_session returns a compact per-event index with the errors flagged; pass event_ids for the full payload of only the events you need.

Trigger runs

list_trigger_runs shows every schedule and webhook firing with its status, error and the session it started.

Health report

diagnostics rolls deploy state, dependency checks, recent sessions, trigger runs, escalations and usage into one call.

Spend

session_usage totals what a session cost, split by input, output, cache and runtime.

tool names1 tool // 10 actionsas a new server lists them
  • agnt_observability_readaction: list_agents · list_sessions · search_sessions · read_session · list_trigger_runs · session_usage · list_escalations · diagnostics · check_dependencies · read_config_request
per-operation names10older servers list these
agnt_observability_list_agentsagnt_observability_list_sessionsagnt_observability_search_sessionsagnt_observability_read_sessionagnt_observability_list_trigger_runsagnt_observability_get_session_usageagnt_observability_list_escalationsagnt_observability_get_diagnosticsagnt_observability_check_dependenciesagnt_observability_read_config_request
02// recorded run

What does a real call return?

A run we recorded on superagnt, with the result trimmed to fit. Nothing on it is invented: the date, the wall time and the cost are the run's own.

recorded runrecorded Nov 3, 2026, 17:17 UTC · wall time 3.9 s · 3 calls · no charge

asked One of my agents says it saved something but I think it failed. Find out what went wrong.

  1. → call 1agnt_observability_read
    { "action": "search_sessions", "agent_id": "f624ee4d-…", "problematic_only": true }
    ← resulttrimmed
    { "session": { "id": "92b8e814-…", "title": "memory probe, session 1", "status": "idle" },
      "has_problem": true }
  2. → call 2agnt_observability_read
    { "action": "read_session", "session_id": "92b8e814-…" }
    ← resulttrimmed
    "total_events": 46, "has_problem": true, "error_count": 1,
    "tools": { "agnt_memory_write": { "calls": 5, "errors": 1 } },
    "index": [ ...
      "#5 17:02:31 agent.mcp_tool_use agnt_memory_write 391c",
      "#6 17:02:31 agent.mcp_tool_use agnt_memory_write 462c",
      "#9 17:02:33 agent.mcp_tool_result 626c ERR sevt_0173Hys2…", ... ]
  3. → call 3agnt_observability_read
    { "action": "read_session", "session_id": "92b8e814-…", "event_ids": ["sevt_0173Hys2…"] }
    ← resulttrimmed
    "is_error": true,
    "output": "Unknown action \"commit\" for agnt_memory_write. Valid actions: remember, forget."

» The agent had been given remember and forget, not commit. Its first remember worked, so it told the user the preference was saved; the extra commit call failed. The fix is one line of config: add commit to its tools, or tell it not to commit.

03// answered

What is the fastest way to triage a failing agent?

Three reads, in this order: search_sessions with problematic_only to find the session, read_session for its event index with the errors flagged, then read_session with event_ids for the failing call's full payload. If nothing failed but nothing happened either, start from list_trigger_runs instead: a schedule that never started a session is a different problem from a session that gave up.

04// measured

Does reading a session dump the whole transcript?

No. The default read is an index: one short line per event with its time, type, tool, size and an error flag. In the triage above the index of all 46 events was 5.8 KB, while the full payloads of just 16 of those events came to 16.6 KB, so a coding agent reads the map first and fetches only what it needs.

sample · One 46-event session, Nov 3, 2026

05// answered

What can a coding agent see about a deployed agent?

Its config and whether its dependencies check out, every session and its events, every schedule and webhook firing, escalations waiting on a person, and what each session spent. A schedule that ran but did nothing, a tool call that failed, an approval nobody gave: each shows up in one of these reads.

06// what it gets used for

What do people use it for?

01

Debug a failing agent

Search for problem sessions, read the event index for the worst one, and pull the failing tool call with its arguments and its error rather than guessing from the reply.

02

Find out why a schedule did nothing

List the agent's trigger runs. A firing that was skipped, with its reason, is a different fix from a session that ran and gave up.

03

Watch what a run costs

Pull session usage after changing a prompt or a tool set, and see whether the change made the agent cheaper or only quieter.

07// cost

What does it cost?

The tools carry no per-call price: every plan includes them, and plans differ only in how much they hold. Plans are Starter $49 a month ($20 of data credit included), Growth $149 a month ($90) and Scale $399 a month ($300), and the no-card trial gives $3 of data credit with Growth limits.

08// questions

What do people ask before they wire it up?

01Does reading a session dump the whole transcript?

No. The default read returns a compact per-event index with position, time, type, tool, size and an error flag. You then fetch only the events you need by their ids.

02What is the fastest way to triage?

search_sessions with problematic_only, then read_session for the index, then read_session with the failing event's id. For "it never ran", start from list_trigger_runs. diagnostics gives the whole health report in one call.

03Can these tools change anything?

No. Every tool in the family is read-only and grants nothing the caller doesn't already have, which is why it is safe to hand to a debugging agent.

04Which clients can use it?

Any MCP client connected to your workspace: Claude Code, Codex, Cursor, Claude or ChatGPT. The tools are on every server, so a coding agent can debug your deployed agents from the same connection it builds them with.

09// keep going

What else can your agent reach?