Skip to content
SideSeat

MCP Server

SideSeat includes a built-in Model Context Protocol (MCP) server that gives AI coding agents direct access to your agent’s execution history — prompts sent, responses received, tool calls made, costs incurred, and errors encountered. Instead of copy-pasting traces into your coding assistant, the agent reads them directly.

Your coding agent can then:

  • Optimize prompts — compare what produces good results vs. poor ones
  • Debug failures — inspect the full execution trace when something breaks
  • Reduce costs — identify expensive model calls and find cheaper alternatives
  • Refine tool definitions — see how models interact with your tools and where they struggle

Setup

Section titled “Setup”

MCP is enabled by default. Pick your tool and connect:

CLI:

Terminal window
kiro-cli mcp add --name sideseat --url http://localhost:5388/api/v1/projects/default/mcp

Or config file — add to .kiro/settings/mcp.json:

{
  "mcpServers": {
    "sideseat": {
      "url": "http://localhost:5388/api/v1/projects/default/mcp"
    }
  }
}

See Kiro MCP docs for more options.

For any other MCP-compatible client, point it at the Streamable HTTP endpoint:

http://localhost:5388/api/v1/projects/{project_id}/mcp

Available Tools

Section titled “Available Tools”

7 tools covering the full observability workflow:

ToolDescription
list_tracesList recent runs with summaries, tokens, costs, and error status
list_sessionsList multi-turn sessions grouping related runs
list_spansSearch operations by type (LLM call, tool exec, agent step), model, framework, or status
get_traceGet a run’s execution structure: span tree with timing, models, and tool invocations
get_messagesGet the normalized conversation with roles, content blocks, token counts, and costs
get_raw_spanGet raw OTLP span data with all attributes and events
get_statsCost and token analytics by model, framework, and time period

Workflows

Section titled “Workflows”

Prompt Optimization

Section titled “Prompt Optimization”

  1. Find recent runs — list_traces to see recent activity

  2. Read the conversation — get_messages with a trace ID to see exact prompts and responses

  3. Compare runs — look at successful vs. failed runs to identify what works

  4. Check costs — get_stats to see cost breakdown by model

Example prompt to your coding agent:

Look at my last 5 agent runs in SideSeat. Find any that errored or had high token usage. Show me the system prompts and suggest improvements.

Debugging Failures

Section titled “Debugging Failures”

  1. Find errors — list_spans with status_code: "ERROR"

  2. Get the trace — get_trace to see the execution tree and where it broke

  3. Read messages — get_messages to see what the model was asked before failing

  4. Check raw data — get_raw_span for framework-specific attributes

Cost Analysis

Section titled “Cost Analysis”

Use SideSeat to compare my costs over the last week. Break down by model and show which runs are most expensive.

Configuration

Section titled “Configuration”

MCP is enabled by default. To disable it:

{
  "server": {
    "mcp": {
      "enabled": false
    }
  }
}

Verifying the Connection

Section titled “Verifying the Connection”
Terminal window
npx @modelcontextprotocol/inspector http://localhost:5388/api/v1/projects/default/mcp

This opens a web UI where you can list tools, call them interactively, and verify responses.

Next Steps

Section titled “Next Steps”