dalao.ai

Latitude

An open-source observability platform that traces AI agents in production and groups recurring failures into trackable signals, self-hostable and MIT-licensedOpen source

About Latitude

Latitude is an open-source observability platform for AI agents: it collects production traces and sessions over OpenTelemetry, groups recurring failures into trackable signals, and through Agent Dispatch hands each problem, with sample traces, to coding agents such as Claude Code and Cursor to fix. The product is organized as Observe, Understand and Refine, and runs self-hosted or as the managed Latitude Cloud, whose Starter plan is free and Pro plan costs $99 a month. It is MIT-licensed, relicensed from LGPL-3.0 to MIT on 2026-05-20 15:47 (UTC+8).

EditorialUpdated 7 min read

What problem it solves

The GitHub description calls Latitude open-source observability for AI agents: it finds the points at which agents fail, dispatches a coding agent to fix it, and verifies the fix against real traces. The README headline frames the goal as self-healing AI agents, with the subline Agents that fix themselves, at scale.

Problem and approach. The docs say AI agents fail in ways that are hard to predict upfront, and logs and dashboards show only individual events; Latitude is organized as Observe, Understand and Refine: collect agent telemetry, group failures into signals, then keep them from coming back with Monitors, Evaluations and Datasets.

How it runs. The docs describe two ways to run Latitude: Self-hosted, deployed by the user with data kept in its own infrastructure, and Latitude Cloud, the managed version. Both run the same open-source platform.

History. The repository was created 2024-06-28 19:16 (UTC+8). Commit 8bef88d on 2026-05-20 15:47 (UTC+8) relicensed it from LGPL-3.0 to MIT. GitHub Releases ship per package, for example python-telemetry-3.8.4 and typescript-telemetry-4.3.0 on 2026-10-08 21:59 (UTC+8), typescript-sdk-5.7.1 on 2026-10-09 16:35 (UTC+8) and cli-4.0.1 on 2026-10-09 17:28 (UTC+8).

Publisher: Latitude Data, S.L. The terms of service and privacy policy both name LATITUDE DATA, S.L. as the operator, and the LICENSE names the copyright holder as Latitude Data SL.

Agents that fix themselves, at scale.

Latitude README subline

Core capabilities

Observe. Each interaction becomes a trace made of spans for LLM calls, tool calls and retrieval steps, and related traces are grouped into sessions to show full multi-turn conversations. The docs list Users, Tools, Memory and Cost views, covering per-user activity, per-tool reliability and latency, how an agent's persistent memory evolves, and how spend breaks down.

Understand. Search finds traces by meaning, exact text or metadata filters; Behaviors clusters sessions by meaning into topics; Signals are the recurring failure patterns Latitude groups from failed scores, each with example traces, affected-user counts, trends and a lifecycle.

Flaggers. The docs call Flaggers built-in automatic annotators that inspect completed sessions for categories such as frustration, jailbreaking, refusal, tool call errors and empty responses, and their annotations feed signal discovery just like human feedback.

Agent dispatch. The docs say that when a new signal is discovered, a signal escalates into an incident, a resolved signal regresses or a monitor opens an incident, Latitude assembles a prompt with the signalId, a deep link and sample trace ids and fires the configured target: Cursor, Claude Code, Linear or Webhooks. The coding agent then investigates over Latitude's MCP connection and implements a fix.

Refine. Monitors watch a signal, a saved search or raw traffic and open an incident that needs attention, with notifications in-app, by email or in Slack; Evaluations generated from a signal keep scoring live traffic; Datasets turn failing traces into cases for local simulation and regression testing.

Interfaces. The README says everything in the UI is also available through the MCP server and CLI, and the homepage describes it as OTEL compatible, with No proprietary format, no lock-in.

Who uses it

As of 2026-10-26 the GitHub API shows 4,734 stars and 404 forks; its open_issues_count (which includes open issues and PRs) is 100.

As of 2026-10-26 the homepage's Running on real production volume panel shows 16,072,336 spans ingested, 343,706 issues detected, 57,262 signals created, 568,493 npm downloads, 50,179 PyPI downloads and 85,976 Docker pulls (as reported by the vendor). Downloads and pulls include repeat and automated fetches and are not a count of installs or users.

Customer story. The Customers page describes Tavi running 75 personalised agent deployments, one per customer; centralizing traces in Latitude replaced per-machine SSH debugging with one place its coding agent can investigate the whole fleet (as reported by the vendor).

How it compares

Against Langfuse: the Langfuse LICENSE puts content under ee/ and related directories under the licence in ee/LICENSE, with the rest under MIT Expat; the Latitude LICENSE is MIT. The Langfuse GitHub description calls it an open-source agent evals and observability platform; the Latitude docs say its difference is a signal-centric workflow that ties observability, semantic search, behaviors, annotations, flaggers, scores and evaluations together automatically (as reported by the vendor).

Against Arize Phoenix: the Phoenix LICENSE is the Elastic License 2.0, and its GitHub description reads AI Observability & Evaluation; Latitude is MIT, self-hostable and open to forking.

Who it's for / Not for

Good fit:

  • Teams with AI agents already in production that need scattered failures turned into trackable problems.
  • Teams fixing code with Claude Code or Cursor that want failure samples and context handed straight to a coding agent.
  • Teams that need an MIT licence and an observability platform on their own infrastructure.

Not a fit:

  • Self-hosted setups with no external AI provider at all: the configuration docs say that without API keys only ingest and trace viewing work, and AI-dependent features stay disabled or fail.
  • Free users who cannot accept public display of their content: the terms of service say User Content on free accounts is public by default unless the account is configured otherwise.
  • Cases that need fixes executed at runtime: the docs say Latitude is the trigger and context provider, not the agent runtime, and does not run the fix.

Getting started

Latitude Cloud. The README says signing up at latitude.so gives an API key and a project slug, with free use of 20K credits a month, 30-day data retention and unlimited seats.

Instrumentation. The manual setup is npm install @latitude-data/telemetry; the README says Python and any OpenTelemetry-compatible runtime are also supported, and lists integrations including OpenAI, Anthropic, Bedrock, Vercel AI SDK and LangChain. Claude Code users can run npx -y @latitude-data/claude-code-telemetry install to capture session transcripts as traces.

Self-hosting. The deployment docs offer Single-host, which runs on one machine with Docker Compose, and Cluster, which deploys on Kubernetes via a Helm chart; the README also lists a Railway template deploy. Images are published on Docker Hub. The platform consists of five application services, web, api, ingest, workers and workflows, running on Postgres with pgvector, Redis, Temporal, an object store and other components.

Self-hosting essentials. The configuration docs require two secrets generated uniquely per deployment, LAT_MASTER_ENCRYPTION_KEY and LAT_BETTER_AUTH_SECRET; production ships without an email service, and login depends on a working email transport, with exactly one of SMTP, Mailgun or SendGrid configured.

Pricing & license

Open-source edition: MIT, free to self-host.

Latitude Cloud:

  • Starter: free, 20K credits a month, 30-day data retention, unlimited seats, general channel Slack support.
  • Pro: $99 a month, 100K credits a month, 90-day data retention, unlimited seats, priority support; extra credits at $20 per 10K; SOC2 and ISO27001 reports.
  • Enterprise: custom, with custom credit volume and data retention, custom on-premises or cloud deployment, RBAC, team trainings, SAML SSO, uptime and support SLA, and dedicated support.

Metering. The pricing FAQ says ingesting a trace costs 1 credit; custom signal runs cost a baseline credit plus the tokens of any AI they use; session intelligence and flaggers also consume credits when they run hosted AI.

Licence: the LICENSE is MIT, with the line Copyright (c) 2026 Latitude Data SL; the GitHub API, the README and the root package.json all mark it MIT. Commit 8bef88d on 2026-05-20 15:47 (UTC+8) moved the LICENSE and README from LGPL-3.0 to MIT and added an MIT declaration to the root package.json. The LICENSE in the packages/cli directory is the Apache License 2.0, and that directory's README says it is built on Fern's fern-cli-sdk, which is under the Apache License 2.0. The @latitude-data/telemetry package on npm is marked MIT.

Risks & limits

Data terms for free accounts. The terms of service say User Content on free accounts is public by default unless the account is configured otherwise, with an additional licence for Latitude to display and share it publicly for promotional purposes; free-account data may be used in aggregate, anonymized form to improve the Services, while paid-account data is not used to train or improve machine learning models.

AI dependencies when self-hosting. The configuration docs say core observability works without API keys while AI-dependent features stay disabled or fail; with nothing set, generation defaults to Amazon Bedrock and embeddings and reranking to Voyage AI, and the matching credentials have to be supplied.

Fixes come from an outside agent. The Agent dispatch docs say Latitude only triggers and provides context, and neither runs the fix nor waits on completion.

Usage billing. The Flaggers docs say the User-side signals and Agent behavior groups use an LLM, so they are sampled and billed by AI usage, while Cost and efficiency and Response validity are deterministic and cost nothing.

Certification claims. The homepage says Latitude is SOC 2 certified and the privacy policy says SOC 2 Type II certified; both say data is encrypted with TLS 1.2+ in transit and AES-256 at rest (as reported by the vendor).

Our verdict

What sets Latitude apart: it puts traces, failure grouping, fix dispatch and regression checks into one loop; it runs no fixes itself and instead hands samples and context to coding agents such as Claude Code and Cursor; it is MIT-licensed and can be self-hosted.

Its limits: self-hosted, the AI-dependent features need credentials for providers such as Bedrock and Voyage AI plus an email transport configured in-house; on the free Cloud plan, User Content is public by default, and only paid-plan data is excluded from model training.

Sources

Official

  1. AI Agent Observability & Monitoring - Latitudelatitude.so · Accessed Oct 26, 2026
  2. Pricing - AI Agent Observability - Latitudelatitude.so · Accessed Oct 26, 2026
  3. Customers | Latitudelatitude.so · Accessed Oct 26, 2026
  4. Latitude - Open-source AI Agent Monitoring (About)latitude.so · Accessed Oct 26, 2026
  5. Terms of Service - Latitudelatitude.so · Accessed Oct 26, 2026
  6. Privacy Policy - Latitudelatitude.so · Accessed Oct 26, 2026
  7. Latitude Overview - Latitude Docsdocs.latitude.so · Accessed Oct 26, 2026
  8. Overview - Latitude Docsdocs.latitude.so · Accessed Oct 26, 2026

Updates & reading