Latitude

Latitude is an MIT-licensed observability and quality platform built specifically for AI agents rather than adapted from LLM call logging. It groups failing production traces into tracked issues, hands them to a coding agent such as Claude Code or Cursor with full context, and replays the resulting fix against the original failing traces to prove it worked.

Latitude is an open-source observability and quality platform for AI agents, distributed under the MIT licence from the latitude-dev/latitude-llm repository, which carries roughly 4,700 GitHub stars, 392 forks and 91 open issues; the TypeScript project was created in June 2024 and was still receiving commits on 20 September 2026, with contributions from more than thirty developers. Its design premise is a deliberate break from the previous generation of LLM tooling: Langfuse, LangSmith, Braintrust and W&B Weave were built to monitor LLM calls and later extended to cover agents, treating an agent as a sequence of calls, whereas Latitude takes the agent session as its unit of analysis. The product is organised as a closed loop of Observe, Understand, Fix and Verify. A single line of telemetry captures complete execution traces including multi-turn sessions and tool calls. Failing traces are automatically grouped into tracked signals that carry state through active, resolved and regressed rather than sitting inert in a log. A coding agent such as Claude Code or Cursor is then dispatched with the full failure context through an MCP server and opens a pull request. The proposed fix is finally replayed against the real failing traces, so a regression is caught before merge rather than in production. Evaluations can be generated from annotated production failures instead of authored up front, which is the practical difference from eval-first tools that only catch failure modes somebody already anticipated. Deployment options span managed cloud, Docker Compose, a Kubernetes Helm chart and one-click Railway.

Category

Developer Tools

Pricing

Freemium, Subscription, Usage-based, Free

Target Market

Enterprise Developers, AI Engineers, MLOps Engineers, CTOs, Product Engineers

Deployment

Open-source, Self-hosted, Cloud-first, API-based

• ✓
  Single-line trace instrumentation

  Captures complete agent execution traces — multi-turn sessions, tool calls and intermediate state — from one line of telemetry rather than manual span annotation.

• ✓
  Signals: failures as tracked issues

  Auto-groups failing traces into named signals with a lifecycle of active, resolved and regressed, so the same failure mode is counted once and its frequency is visible.

• ✓
  Coding-agent dispatch over MCP

  Sends a failing signal with full context to Claude Code or Cursor through a Model Context Protocol server, and the agent opens a pull request with the proposed fix.

• ✓
  Verification against real traces

  Replays a candidate fix against the actual traces that failed, so a change is proven to resolve the issue and not regress others before it is merged.

• ✓
  Evals generated from production failures

  Builds evaluation cases out of annotated real failures instead of requiring teams to predict the failure surface in advance, which is where eval-first tools miss unknown modes.

• ✓
  Fully self-hostable under MIT

  Ships as Docker Compose for a single host, a Helm chart for Kubernetes, or one-click Railway, with no feature-gated open-core split on the observability core.

• •
  Production agent debugging

  Find which step of a multi-turn customer-support agent silently picked the wrong tool, across thousands of sessions, without reading raw traces.

• •
  Regression prevention on prompt changes

  Replay every previously failing trace against an edited prompt or model swap before shipping, catching regressions that unit tests never model.

• •
  Closing the loop with a coding agent

  Route a recurring failure signal straight to Claude Code or Cursor so the fix arrives as a reviewable pull request rather than a ticket.

• •
  Building an eval suite from real traffic

  Annotate production failures as they occur and let those annotations become the regression suite, so coverage grows with observed reality.

• •
  Self-hosted observability for sensitive prompts

  Run the whole stack inside a private network when prompts and traces contain customer data that cannot go to a managed vendor.

Latitude publishes real list prices, which is unusual for this category. Metering is credit-based rather than per-seat, and seats are explicitly unlimited on every tier — the free Starter plan gives 20,000 credits a month with 30-day retention, and Pro is $99 a month for 100,000 credits with 90-day retention, with overage at $20 per 10,000 credits. Enterprise is quote-only and is where on-premises deployment, fine-grained RBAC, SAML SSO and an uptime SLA live; SOC 2 and ISO 27001 reports are available from Pro upward. The genuinely free path is self-hosting the MIT-licensed distribution, in which case the only cost is the infrastructure it runs on. Product Hunt reviewers mention hitting occasional credit and timeout issues, so model the credit burn of high-volume agent traffic before committing to a tier.

This page was written from 5 sources, 4 on domains other than latitude.so.

1. 1.latitude.so — pricingvendor
2. 2.github.com — latitude llm
3. 3.api.github.com — latitude llm
4. 4.producthunt.com — latitude 4
5. 5.aiprosol.com — llm observability