v0.21.0A security release: what an agent or a client sends Iris is bounded before Iris reads it, and a tool argument can narrow what the operator allows but never widen it→

Comparison · Evaluation

Iris vs Braintrust

MCP-Native Eval vs Experiment-Driven Evaluation.

TL;DR

Iris is an MCP server your agent discovers and uses on connect — one config block, no SDK, one SQLite file, every rule's precision and recall published at iris-eval.com/proof. Braintrust — from its own pages, read 2026-10-01: Braintrust is an eval and observability platform with dataset management, experiment tracking, a prompt playground, and span-level tracing. If you need production-grade experimentation workflows, human review, or CI-integrated regression testing, Braintrust is the deeper eval platform. Online scoring rules score production logs at span, trace or group scope. (source)

Iris grades what an agent did with its tools — the trace, the answer, the cost — not whether an MCP server honours its own contract; a server test harness answers that question, and Iris runs beside it. For the method, see the agent eval guide.

Feature comparison

Side by side.

13 features, the same 13 on every comparison. Every Braintrust cell links the page it was read from and the date; where the sentence it was read from was found verbatim on a plain download of that page (checked 2026-09-22), the link carries it as its title, and where it was not, the cell says so. The highlighted cells are Iris's own call on which side is stronger for a team running MCP agents — 4 to Iris, 3 to Braintrust — not a measurement.

FeatureIrisBraintrust
Integration methodOne block in the MCP config, no code — the agent discovers Iris and its tools on connectSDK imports or OpenTelemetry (Python, TypeScript, Java, Go, Ruby, C#)Braintrust's page · read 2026-10-01
Self-hostingOne process, one SQLite file; Docker image with a health checkEnterprise plan only; hybrid: you run the data plane, Braintrust hosts the UIBraintrust's page · read 2026-10-01
Where it runsNothing in the agent's process — Iris is a separate server the agent callsBackground-thread logging by default; synchronous flush optionalBraintrust's page · read 2026-10-01
Evaluation26 built-in deterministic rules and 9 custom-rule types, in-process; 7 judge templates on a key you supply; every rule's precision and recall publishedLLM, code, and human scoring + datasets + experimentsBraintrust's page · read 2026-10-01
Cost trackingPer-trace USD cost and tokens; a cost spike judged against the agent's own historyLatency, token counts and cost captured per span and trace; dashboards filter and group by metadataBraintrust's page · read 2026-10-01
MCP supportProtocol-native — Iris is an MCP server with 12 tools; OTLP traces inMCP server that creates scorers and evaluators, runs evals, and edits datasets (Aug 2026)Braintrust's page · read 2026-10-01
LicenseMIT, the whole packageProprietary (proxy is MIT)Braintrust's page · read 2026-10-01
OwnershipIndependent and founder-ledVenture-backed company; $80M Series B led by ICONIQ (Feb 17, 2026)Braintrust's page · read 2026-10-01
DashboardA local dashboard on its own port: traces, moments, regressions, five viewsDashboards track requests, latency, tokens, cost, scores, topics; filters and groupings apply to all chartsBraintrust's page · read 2026-10-01
Framework supportAny MCP client (2 verified, 8 claimed — see /clients); OTLP/HTTP from anything elseIntegrations for OpenAI, Anthropic, Vercel AI SDK, LangChain, LlamaIndex, OpenTelemetry, LangGraph, CrewAI + moreBraintrust's page · read 2026-10-01
Prompt managementNot includedVersioned prompts (new version per save); invoke() by slug from code, pin versions in productionBraintrust's page · read 2026-10-01
Enterprise and complianceSelf-hosted. Nothing leaves your machine unless you turn one of these on: an OpenTelemetry endpoint (IRIS_OTEL_ENDPOINT), which exports traces to the collector you name; the LLM judge with your own key, which sends the text it judges to that provider, and whose citation check fetches the pages an output cites; or a webhook, which posts ids, the verdict and rule names, never the text, to the address you set. No compliance certification is claimed before it is heldSOC 2 Type II, SAML SSO, BAA (HIPAA), RBAC, on-prem or hosted deployment, premium supportBraintrust's page · read 2026-10-01
Cost to runFree — MIT, one process on your machine; the only spend is a judge call on a key you supply, when you opt inStarter $0/month (1 GB processed data, 10k scores, 14-day retention); Pro $249/month (5 GB, 50k scores, 30-day retention); Enterprise customBraintrust's page · read 2026-10-01

Decision guide

Which one fits your stack?

When to choose Iris

  • You are building with MCP-compatible agents and want the integration to be one config block
  • You want the evaluation to be deterministic and local — no model call, nothing leaving the machine
  • You want to read what each rule is worth before you trust it: every rule's precision and recall is published
  • You want self-hosting to be one process and one file
  • You want a fully permissive MIT license on the whole package

When to choose Braintrust

  • You need deep eval capabilities — datasets, experiments, human review, LLM scoring source
  • You need a prompt playground for iterating on prompts with real data source
  • You need enterprise compliance today (SOC 2 Type II, SAML SSO, on-prem or hosted deployment) source
  • You need multi-language SDK support (Python, TypeScript, Java, Go, Ruby, C#) source
  • You need cost and token analytics you can filter and group across dashboards source
  • You need CI/CD-integrated regression testing against production datasets source

FAQ

The questions buyers ask.

What is the difference between Iris and Braintrust?
Iris is an MCP-native agent eval server that needs no SDK. Add the config block, restart your client, and every session lists Iris's tools on connect. Iris never intercepts: it runs when your agent calls one of its tools, when a host hook or `iris-eval ingest` hands it a trace, or when you POST one to its HTTP API. Braintrust is an SDK-based eval and observability platform with datasets, experiments, a prompt playground, and CI/CD-integrated regression testing. Braintrust offers deeper evaluation workflows with LLM-as-a-judge, code scorers, and human review. source
Is Iris better than Braintrust for MCP agent evaluation?
Braintrust requires SDK imports (or OpenTelemetry) but offers experiment workflows, dataset management, and SDKs for Python, TypeScript, Java, Go, Ruby, and C#. source
How much does Braintrust cost to run?
Iris: Free — MIT, one process on your machine; the only spend is a judge call on a key you supply, when you opt in. Braintrust, from its own page read 2026-10-01: Starter $0/month (1 GB processed data, 10k scores, 14-day retention); Pro $249/month (5 GB, 50k scores, 30-day retention); Enterprise custom. source
Can I self-host Braintrust?
Iris: One process, one SQLite file; Docker image with a health check. Braintrust, from its own page read 2026-10-01: Enterprise plan only; hybrid: you run the data plane, Braintrust hosts the UI. source
Does Braintrust work with MCP agents?
Iris: Protocol-native — Iris is an MCP server with 12 tools; OTLP traces in. Braintrust, from its own page read 2026-10-01: MCP server that creates scorers and evaluators, runs evals, and edits datasets (Aug 2026). source

Sources

Every Braintrust statement on this page was read from one of these pages on the date shown. The file behind this page is website/src/lib/compare/braintrust.json.

Last verified: 2026-10-01. This comparison is based on publicly available documentation and may not reflect recent changes to Braintrust. We aim to keep this page accurate and fair.

See something outdated or incorrect? Report an inaccuracy — we review and update within 48 hours.

Ready to see what your agents are doing?

Add Iris to your MCP config. First trace in 60 seconds. No SDK, no signup, no infrastructure.

Install IrisTry the playgroundEvery comparison