Braintrust

Braintrust combines three things AI teams usually stitch together from separate tools: tracing/observability for production agent runs, an evaluation framework (offline experiments plus online scoring), and a prompt/eval playground with human review. It ships SDKs for Python, TypeScript, Go, Java, Ruby, Rust, and .NET (all on GitHub under the braintrustdata org), a CLI (bt), and a Loop feature that runs an LLM investigator against your traces to surface recurring failure patterns on a schedule. Docs: braintrust.dev/docs.

Braintrust built and open-published the architecture of Brainstore, a custom columnar store designed to handle traces that can run hundreds of megabytes per interaction — a scale traditional observability backends struggle with, per Braintrust's own engineering writeups (^[2], ^[3]). This is a vendor claim about their own system, not independently benchmarked, but the technical detail in the posts (shingled bloom filters for phrase search, an async execution engine for streaming log search results) goes well beyond marketing copy, which is a reasonable signal for a team evaluating whether the product will hold up at real trace volume.

Braintrust splits into a control plane (UI, metadata, hosted by Braintrust) and a data plane (your actual traces, logs, and datasets). In hybrid mode you self-host the data plane — via a published Docker Compose config or Terraform modules for AWS, Azure, and Google Cloud (all public repos under braintrustdata) — while Braintrust hosts the control plane, so sensitive trace data never leaves your environment even as you get UI updates automatically ^[4]. Notion and Ramp are cited by Braintrust as hybrid-deployment customers. Fully self-hosted (both planes) is also available via the same Terraform/Helm tooling but is aimed at Enterprise customers per the docs.

Three tiers, verified on the pricing page 2026-09-29: Starter is free — $10 of model credits, 1 GB of processed trace data ($4/GB after), 10,000 scores ($2.50/1k after), 14-day retention, unlimited users and projects. Pro is a flat $249/month — $100 credits, 5 GB processed data ($3/GB after), 50,000 scores ($1.50/1k after), 30-day retention (extendable at $0.50/GB/month), plus custom charts, environments, priority support, and RBAC. Enterprise is custom-quoted and adds SAML SSO, a BAA for HIPAA, an uptime SLA, and custom data retention/export — none of which are on Pro. Pro is a flat team price, not per-seat, but it's also not tiered by usage below the included credits/data/scores, so a small team doing meaningful trace volume can hit the $249 floor quickly. Source: ^[1].

Braintrust has raised three public rounds: a $5M seed (Dec 2023), a $36M Series A (Oct 2024), and an $80M Series B led by ICONIQ with returning investors Andreessen Horowitz, Greylock, Elad Gil, and basecase capital (Feb 2026) — all announced on the company's own blog (^[5], ^[6], ^[7]). Founder/CEO is Ankur Goyal, previously at Impira and Figma. The company publishes customer video testimonials from Notion, Cloudflare, Replit, Dropbox, Box, Vercel, Bill.com, Navan, and Browserbase ^[13] — these are vendor-selected and vendor-hosted, not independently sourced, so treat them as marketing evidence of adoption rather than a neutral customer list.

The core Braintrust platform (tracing backend, eval engine, Brainstore, UI) is closed source and delivered as SaaS or hybrid/self-hosted deployment — there is no public repo for the product itself. What is open source: the client SDKs (all under MIT, per GitHub), and notably autoevals (github.com/braintrustdata/autoevals), a standalone scoring/evaluation library with 1,041 GitHub stars and commits as recent as 23 September 2026 — actively maintained, not abandoned. Treat the SDKs and autoevals as complementary tooling, not the product; deciding to use Braintrust means buying into the closed-source hosted or hybrid platform behind them.

Braintrust achieved SOC 2 Type II compliance, announced 15 July 2024 ^[8]. It runs a Vanta-hosted trust center at ^[9] listing current compliance status, and publishes a public status page at ^[10]. Per the pricing page's feature comparison, MFA and SOC 2 are available on all plans; SAML SSO, a BAA, and an uptime SLA are Enterprise-only; a DPA is available with a click-through agreement on Pro and a negotiated version on Enterprise.