What is LangSight?
Agents call three types of things:- MCP servers (postgres-mcp, jira-mcp, slack-mcp) — via the MCP protocol
- Non-MCP tools (Stripe API, Sendgrid, Python functions, code execution) — direct calls
- Sub-agents (agent-to-agent handoffs in multi-agent workflows)
server_name field in a ToolCallSpan can be “stripe-api”, “sendgrid”, “openai-api”, or “postgres-mcp”. It is not locked to MCP servers.
MCP servers get extra depth because the MCP protocol is standard and inspectable — proactive health checks, security scanning, schema drift detection, and alerting on DOWN/DEGRADED state. Non-MCP tools appear in every trace but cannot be proactively health-checked (no standard protocol to ping).
What you can do per tool type
| Tool type | Observe calls | Health check | Security scan | Cost tracking |
|---|---|---|---|---|
| MCP servers | Yes | Yes | Yes | Yes |
| HTTP APIs (Stripe, Sendgrid, etc.) | Yes | No | No | Yes |
| Python functions | Yes | No | No | Yes |
| Sub-agents | Yes | No | No | Yes |
Example: mixed agent session
Agent Session Traces
Full ordered trace for every agent session — MCP calls, HTTP API calls, function calls, sub-agent handoffs — all in one view with cost per session.
Multi-Agent Support
When Agent A hands off to Agent B which calls Agent C, trace the full tree. Parent-child span relationships (same model as OpenTelemetry).
MCP Health Monitoring
Proactive availability checks, latency tracking, and schema drift detection for MCP servers. Unique: synthetic probes, not just passive recording.
MCP Security Scanning
CVE detection, OWASP MCP Top 10 audit, tool poisoning detection, and auth gap analysis — for MCP servers specifically.
LangSight and Langfuse
LangSight is not a Langfuse replacement. They answer different questions:| Langfuse | LangSight | |
|---|---|---|
| What it traces | LLM calls — prompts, completions, token costs | Tool calls — MCP spans, latency, errors, cost |
| What it monitors | Nothing (passive only) | MCP server health — synthetic probes, schema drift |
| Security | None | CVE, OWASP MCP Top 10, tool poisoning, auth audit |
| Multi-agent | Agent reasoning chains | Agent-to-agent handoffs via parent_span_id |
postgres-mcp/query. LangSight shows that call took 4 seconds and failed because the connection pool was exhausted.
How it works
Installation
Quickstart →
Get from install to your first agent trace in 5 minutes.
Integrations
LangSight works with every major MCP client and agent framework:| Framework | Integration |
|---|---|
| Claude Desktop | Auto-discovered by langsight init |
| Cursor | Auto-discovered by langsight init |
| VS Code | Auto-discovered by langsight init |
| LibreChat | Native plugin |
| LangChain | LangSightLangChainCallback |
| Langflow | LangSightLangChainCallback |
| LangGraph | LangSightLangChainCallback |
| LangServe | LangSightLangChainCallback |
| CrewAI | LangSightCrewAICallback |
| Pydantic AI | @langsight_tool decorator |
| Any OTEL framework | OTLP endpoint |