Braintrust

What it does

Braintrust takes in production agent traces, including prompts, model responses, and tool calls, along with datasets and scoring inputs from LLM judges, code, or humans. It produces trace views, latency, cost and quality tracking, evaluation experiments that compare prompts and models, and automated pattern discovery that flags recurring failures.

Teams use it to inspect traces, turn real failures into regression datasets, and run experiments before release. The platform's distinguishing features are the Loop agent, which proposes improved prompts, scorers and datasets, and Patterns, an automation that surfaces recurring issues across production traces. An MCP server connects coding agents to logs and evals, and SDKs exist for Python, TypeScript, Go, Ruby and C#.

It suits teams running agents in production that need shared evaluation workflows across engineering and product. Starter is free without a credit card and includes usage allowances and 14-day retention. Pro costs $249 per month and includes basic role-based access controls and 30-day retention. Enterprise adds custom governance and deployment terms. Model usage beyond included credits and data-processing overages are billed separately.

What it can do

• Trace inspection of prompts, responses and tool calls in real time
• Evals that score outputs with LLMs, code or humans
• Versioned datasets built from production traces
• Pattern discovery and natural language trace investigation
• Loop agent for generating prompts, scorers and datasets
• Custom trace views, dashboards and environments

Works with

• MCP server for connecting coding agents
• Native SDKs for Python, TypeScript, Go, Ruby and C#
• S3 data export

Example workflows

Investigate production failures

Input: production traces from an AI application. Steps: instrument the app with an SDK, use Patterns or natural language search to find recurring failures, annotate examples. Output: a set of flagged traces and recurring issue patterns.

Build a regression eval

Input: real failing traces. Steps: convert traces into a versioned dataset, run an experiment comparing a candidate prompt against the baseline, review scores. Output: an experiment comparison showing scored results per prompt or model.

Before you choose

• Model credits cover built-in models and features like Topics; usage beyond included credits is billed at token rates.
• Starter includes 14-day retention. Pro includes 30 days, with additional retention listed at $0.50 per GB per month; Enterprise retention is custom.
• Pro includes basic role-based access controls. SAML SSO, custom roles and hybrid deployment require Enterprise terms.

Listed under

• Coding
• Code review and testing
• DevOps and observability
• AI agents
• AI product teams
• Developers
• Evals
• LLM monitoring
• Observability

Get started

No credit card required for the Starter plan.

1. Create an account at braintrust.dev/signup
2. Install a native SDK and add tracing to your application
3. Log your first trace and run your first eval from the web app

Starter is free with $10 model credits, 1 GB processed data and 10k scores included, then per-unit overage. Pro is $249 per month with $100 model credits, 5 GB processed data and 50k scores included. Model token rates are on the linked plans-and-limits page and were not supplied.

Source: braintrust.dev/pricing, checked 29 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 1 Nov 2026. Each source supports part of the overview; availability and prices can change.

• Braintrust - The active observability platform for agents ↗www.braintrust.dev
• Get started with Braintrust - Braintrust ↗www.braintrust.dev
• Pricing - Braintrust ↗www.braintrust.dev

Company website

braintrust.dev: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 1 Nov 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.