Agent Trace Check in Claude

Questions it answers

Agent Trace Check reads a tool-call log from an AI agent run and says what went wrong: which tools failed and how often, where the agent retried the same call in a loop, which calls it repeated with the same arguments, and which calls were slow.

Paste the log as JSON lines or one JSON value. It reads OpenAI chat messages with tool_calls, Responses API function_call items, Anthropic tool_use and tool_result blocks, Claude Code session JSONL files, MCP tools/call requests and their answers, LangSmith tool runs, OpenTelemetry tool spans, and flat rows with a tool name, arguments, a status or error, and a duration or timestamps.

Calls are paired with their results by id, or in order when the log has no ids. A call is slow at 10 seconds or more, or at 5 times its tool's median. A retry loop is 3 or more back-to-back calls to one tool where each repeats the last arguments or follows a failure. Both thresholds can be changed in one call.

The answer has findings ranked worst first, a table per tool with failures, median and slowest time and the most common error, and the lists of failures, retry loops, repeated calls and slow calls.

There is no account and nothing is stored. The log is read for the call and is not kept or cached, and no tool in it is called again. The only lasting record is a daily count of calls per tool.

What it can do

• Reads OpenAI, Anthropic, Claude Code, MCP, LangSmith, OpenTelemetry and flat JSON tool-call logs
• Pairs each tool call with its result by id, or in order when there are no ids
• Counts failures per tool and names the most common error
• Finds retry loops: back-to-back calls that repeat or follow a failure
• Finds the same call made again with the same arguments
• Lists slow calls by a fixed limit and by each tool's median time

Tools

• analyze_trace, Analyze an agent tool-call log. Use this when the user pastes a log of an AI agent's tool calls and wants to know which calls failed, where the agent retried in a loop, which calls it repeated, or which were slow. Pass the log as trace: JSON lines or one JSON value from OpenAI messages, the Responses API, Anthropic messages, Claude Code session JSONL, MCP JSON-RPC, LangSmith, OpenTelemetry spans, or flat rows with tool, args, status or error and duration_ms. Returns findings worst first, per-tool stats, failures, retry loops, repeated calls and slow calls. Nothing is stored. It does not run the agent or any tool.

Add Agent Trace Check in Claude

Claude connects to a remote MCP server as a custom connector. The connection is made from Anthropic's cloud, not from your laptop, so the server must be on the public internet. Agent Trace Check is. These steps were checked against Claude's custom-connector article (updated 2026-08-11) and the Claude Code MCP docs on 2026-10-01.

On a Pro or Max plan:

1. Open Customize, then Connectors.
2. Select the plus button, then Add custom connector.
3. Paste https://agenttrace.openkrill.app/mcp. Leave Advanced settings empty: this server does not need an OAuth client id or secret.
4. Select Add. The connector is yours. Anthropic has not reviewed it, and it is not a Directory listing.
5. In a conversation, open the plus button, then Connectors, and enable Agent Trace Check. Ask: Analyze this agent tool-call log and tell me why the run failed.

On Team and Enterprise, an Owner adds the connector first at Organization settings, then Connectors, then Add, then Custom, then Web, and pastes the same URL. Members then connect it from Customize. A Free plan can add one custom connector.

In Claude Code, add it from the project directory:

claude mcp add --transport http agent-trace https://agenttrace.openkrill.app/mcp

That writes a local or user config. A JSON entry needs "type": "http" (or streamable-http) plus the url. An entry with a url and no type is skipped. Confirm with claude mcp get agent-trace. The tools you should see are analyze_trace.

If Claude tries to fetch https://agenttrace.openkrill.app/mcp from its code sandbox and reports HTTP 403, the block is Claude's sandbox network allowlist, not this server. The sandbox proxy answers 403 with x-deny-reason: host_not_allowed for every host outside the allowlist. Use the connector above instead: connector calls come from Anthropic's cloud and do not pass the sandbox proxy. To let the sandbox reach it as well, an Owner adds agenttrace.openkrill.app under Settings, then Capabilities, then Additional allowed domains.

Enable the connector only in conversations that need it. Claude can call the tools while you work, including during research, so turn off any tool you do not want invoked. Agent Trace Check does not write to your accounts. It still receives the arguments of each call. Read the privacy page before you paste anything that is not public.

If it does not work

A call checked on 2026-10-11

Four made-up log rows: one fetch that timed out 3 times in a row, then a search that worked. The tool reads only this text and calls nothing. The request below was posted to https://agenttrace.openkrill.app/mcp as tools/call. HTTP 200. Source: the Agent Trace Check server, read 2026-10-11.

{
  "method": "tools/call",
  "params": {
    "name": "analyze_trace",
    "arguments": {
      "trace": "{\"tool\":\"web_fetch\",\"args\":{\"url\":\"https://example.com/a\"},\"status\":\"error\",\"error\":\"timeout after 30000 ms\",\"duration_ms\":30000}\n{\"tool\":\"web_fetch\",\"args\":{\"url\":\"https://example.com/a\"},\"status\":\"error\",\"error\":\"timeout after 30000 ms\",\"duration_ms\":30000}\n{\"tool\":\"web_fetch\",\"args\":{\"url\":\"https://example.com/a\"},\"status\":\"error\",\"error\":\"timeout after 30000 ms\",\"duration_ms\":30000}\n{\"tool\":\"search\",\"args\":{\"q\":\"example\"},\"status\":\"ok\",\"duration_ms\":800}"
    }
  }
}

• notice. Computed from the pasted log alone. Nothing was stored, and no tool in the log was called again. Calls are matched by id, or in order when the log has no ids.
• findings. 3 returned
• by tool. 2 returned
• failures. 3 returned
• retry loops. 1 returned
• repeated calls. 1 returned

Repeat the call yourself if you need a newer reading. Cached answers expire. A rate limit is not a result: wait and try again. Nothing in the call is a ranking, a filing, or advice.

Using it in production

Nothing to install and nothing to run: no package, no Python or Node.js version and no environment variable. Agent Trace Check runs on our servers, and your client only needs the URL https://agenttrace.openkrill.app/mcp.

Use it in other apps

• Agent Trace Check in ChatGPT
• Agent Trace Check in Cursor
• Agent Trace Check in VS Code
• Agent Trace Check in Pi

Frequently asked questions

Words on this page

• MCP (Model Context Protocol) definition
• Streamable HTTP transport definition
• Tool primitive definition
• Rate limit policy definition

Source: plugin listing for agent-trace in this repository, and a live tools/call on 2026-10-11. As of 2026-10-11.