Query AI agent traces via MCP. Fix what GlassFlow finds.

Most agent observability tools only let your agents send traces in. GlassFlow goes the other direction: query AI agent traces via MCP, understand what broke, and open a fix. No dashboard required.

Get started Talk to the founders

Query AI agent traces via MCP, right in your editor

GlassFlow · mcp · claude-codeLIVE
Two OTEL lines to ingest. One MCP server to query. No dashboard required.

AI Agent Observability via MCP

Full visibility on each of your MCP queries

Every query your agent makes is a first-class trace. See what it asked, which tool ran, how long it took, and what it cost.

MCP Tool Calls
Call logTool stats
Total calls (1d)
12
across all agents
Success rate
83.3%
10 / 12
Avg latency
1057ms
p95 3.2s
Most used
web_search
89 calls
Search tool or agent… Agent: All document sales infra research
TimeToolAgentDurationTokensStatusCost
09:14:02web_searchdocument-processor2140ms1,200success$0.018
09:16:38read_filedocument-processor310ms420success$0.003
09:19:55write_filedocument-processor180ms580success$0.004
09:22:11web_searchsales-outreach-agent3200ms1,800error$0.024

Works with every AI coding agent and editor

One MCP server, any client. Query GlassFlow from Claude Code, Codex, Windsurf, or Cursor.

GlassFlow
Claude Code
Codex
Windsurf
Windsurf
Cursor
Cursor

Close the loop automatically

Your agent queries GlassFlow at the start of each run, reads what failed last time, and adjusts before making the same mistake again.

Monitor AI agents without a dashboard

Ask which workflow is slowest, which user spends the most tokens, which tool fails most often. Get answers in natural language.