Brok's ForgeAI Engineering
WhyHow it reasonsArchitectureProductCompareDocsGitHub
Get started
WhyHow it reasonsArchitectureProductCompareDocsGitHubGet started
Open source · Apache 2.0

The AI Engineering Operating System.

Your AI system is a sequence of engineering decisions — a prompt promoted, a dataset regenerated, a model swapped. Broks Forge records those decisions with the evidence and the reasoning behind them, then answers the questions that follow: why did this fail, what changed, has this happened before, and can we defend it?

Get startedSee how it works
Build
agents, prompts, datasets, providers
Evaluate
reproducible measurement
Understand
intelligence, knowledge, memory
Evolve
AI Git, promotion, rollback

An example answer from Brok. Question: has this happened before? Answer, near-certain: yes. Checkout Quality number one failed nineteen days ago against the same agent and dataset, with an identical recorded cause, which makes this a recurrence rather than a coincidence. This is derived, read from earlier evaluations sharing an artifact with this one. Both failures recorded “Connection refused”. Engineering memory records “Moved the endpoint behind the internal gateway”. Two downstream artifacts are left without evidence.

Ask Brok · Has this happened before?

Yes — this has happened before.

near-certain

Checkout Quality #1 failed 19 days ago against the same agent and dataset. The recorded cause is identical, which makes this a recurrence, not a coincidence.

Derived · read from earlier evaluations sharing an artifact with this one

ReasoningBoth failures recorded: “Connection refused”.
Memory“Moved the endpoint behind the internal gateway.”
Impact2 downstream artifacts left without evidence.
View the failure graphCompare revisionsInvestigate the precedent

Every statement declares how it is known. Nothing is generated.

The problem

Teams lose the reasoning behind their AI systems.

The prompt is in version control, but why v8 replaced v7 is in a Slack thread. The evaluation results are on a dashboard, but which decision they justified is nowhere. The engineer who knew has left. Six weeks later something fails, and nobody can reconstruct which change mattered.

An ordinary engineering question

Tracing & eval tools

Broks Forge

What did this request do?

Which calls were slow or expensive?

Did this prompt change break anything?

Why was this prompt promoted?

What evidence supports the current configuration?

Has this failure happened before, and what did we do?

Which of our decisions have no evidence behind them?

What would break if I changed this dataset?

Why is the system the way it is?

The bottom half of that table is not a feature gap a better dashboard closes. Those questions are about artifacts, versions, decisions and evidence — objects that tracing does not model, because it was never trying to. Answering them needs a different data model.

The solution

Record the engineering act. Then reason over it.

You register artifacts and evaluate them — work you were doing anyway. From that, Broks Forge derives observations, claims, decisions, evidence and durable knowledge. Nobody writes any of it down.

Derived, not authored

Promoting a version is a decision. Running an evaluation is evidence. The record is a by-product of engineering, so it never rots.

Nothing is fabricated

The reasoning layer is deterministic over real rows — no language model. When the record cannot answer, it says so.

Every claim is labelled

Derived, inferred, suggested or unknown — with a three-step confidence ladder instead of an invented percentage.

Absence is not health

An artifact nobody measured is reported as unknown, never as passing. A distinct verdict state, on purpose.

What is an AI Engineering Operating System?

Deterministic reasoning

There is no language model in the reasoning layer.

Not as a limitation — as the point. An answer you have to go and verify has not saved you the work of finding it. Broks Forge resolves your question to an engineering intent and composes the answer from real rows, so the same question always returns the same answer, and every sentence can be traced to the record that produced it.

A generated answer

“The evaluation likely failed due to a configuration issue with the agent's credentials. You may want to check your provider settings and re-run the test.”

  • Plausible whether or not it is true.
  • No way to tell which part came from your system.
  • Ask twice, get two different answers.
  • You still have to go and check.

A derived answer

“6 of 8 sampled runs returned HTTP 401 against Local OpenAI-compatible. Credentials on that provider were rotated 3 hours before this ran.”

  • Derived — the run rows say so.
  • Traceable — every clause opens the row behind it.
  • Reproducible — the same question returns the same answer.
  • Bounded — proximity is offered as proximity, not as proof.

Knowledge

Derived from work already done — never written, never generated.

Documentation rots because keeping it true is a second job nobody is promoted for. Broks Forge never asks. Promoting a version is a decision. Running an evaluation is evidence. Knowledge appears only where a decision and its evidence genuinely both exist — which is also why the platform can tell you which of your decisions have nothing behind them at all.

Derived

Read directly from the record.

“4 evaluations cover this revision.”

Inferred

A conclusion drawn from what the record holds, and labelled as one.

“These failures read as infrastructure, not quality.”

Suggested

A recommendation, carrying the evidence behind it.

“Do not promote v4 — nothing has measured it.”

Unknown

The record cannot answer. Said plainly, never filled in.

“No evaluation has run against this dataset.”

Absence is never reported as health. An artifact nobody has measured comes back as unknown — a distinct verdict, on purpose, because “no failures” and “no evidence” are not the same claim.

How the reasoning works

Architecture

Five layers, each earning the one above it.

A layer may read everything beneath it and nothing above it, and no layer may duplicate a layer below it. That rule is why adding a reasoning surface never means adding a table.

Engineering Applications

Brok · Root Cause Explorer · Briefs

Reason over everything below. Own no data, so they cannot drift from the truth.

Layer 5

Forge Graph

the living map

Artifacts, their real relationships, and reasoning layered on top. The Execution Graph and Failure Graph are the same map narrowed to a single run.

Layer 4

AI Git

the evolution engine

Revisions, promotions, rollbacks and the rationale behind every change. Evolution reads that history back as lineage, dependents and blast radius.

Layer 3

Registry

the engineering catalog

Every artifact and every derived knowledge object, discoverable in one place.

Layer 2

Forge Kernel

the invisible foundation

Identity, tenancy, persistence, execution. You are not supposed to notice it.

Layer 1

Brok

An engineering partner, not a chatbot.

Ask why something failed, whether to promote it, or what the team should do next — and get an answer read from your own engineering record. There is no language model: Brok resolves your question to one of 25 engineering intents and composes the answer from real rows. Ask it something the record cannot support and it refuses, then offers what it can answer.

How Brok works

Ask Brok · Should I promote it?

Nothing has measured v4, so promoting it would be an act of faith.

A promotion with no evidence behind it cannot be defended later and cannot be safely reversed either.

DerivedNo evaluation has run against Support Prompt since v4 was created.
InferredThe failures read as infrastructure rather than quality.

Investigated 4 evaluations · 2 AI Git revisions · engineering memory

Root Cause Explorer

When you ask why, you get an investigation.

Open a failure and everything the record holds about it is already assembled: a dated chronology of the engineering that led there, the cause at four depths, the evidence and AI Git chains, the engineering memory, and every earlier failure on the same ground. No hunting across five pages.

How the Explorer works

Investigation · Checkout Quality #2

Immediate

The agent's credentials are being rejected

6 of 8 sampled runs returned HTTP 401.

Contributing

Every failing run reached the same provider

Inferred · likely

Historical

This has happened before — 19 days ago

Identical recorded cause.

Related change

Credentials rotated 3 hours before this ran

Proximity, not proof.

AI Git

Version control for engineering reasoning.

Not source control. AI Git answers engineering questions: what was promoted, why, what it replaced, what evidence covered it — and whether production is running the newest revision or an older one somebody rolled back to at 2am. The reason you record becomes Engineering Memory, recalled verbatim forever.

How AI Git works

Support Prompt · deployment timeline

Rolled back. Production is on v3 even though v4 exists.

  1. v4Rolled past

    “Trimmed the system preamble.”

  2. v3Current production

    “Softer tone after complaints.”

  3. v2Superseded

    “Added the refusal instruction.”

Engineering Intelligence

The reasoning layer nobody has to write.

Observations, claims, decisions, evidence and knowledge — all derived from real engineering work. Promoting a version is a decision; running an evaluation is evidence. Knowledge exists only where both are genuinely present, which is why it can be trusted, and why the platform can also tell you which of your decisions have nothing behind them.

How Engineering Intelligence works

Derived from real work

Observation
Checkout Quality measured Refund Agent: 2 of 2 items failed.
Decision
Support Prompt v3 promoted — “Softer tone after complaints.”
Evidence
4 evaluations cover the promoted revision.
Knowledge
Support Prompt's canonical revision is v3, backed by 4 evaluations.

2 decisions have no evidence behind them. A promotion nobody measured is a position carried on faith.

Forge Graph

The architecture diagram that cannot go stale.

Every architecture document is out of date the week after it is written, because keeping it true is manual. The Forge Graph is not maintained — it is the relationships the system actually recorded while you worked. Ask what depends on a dataset before you regenerate it and the blast radius is a fact, not a recollection. Narrow the same map to one run and it becomes the Execution Graph; narrow it to one failure and it becomes the Failure Graph.

How the Forge Graph works

Blast radius · Support Dataset v2

Changing this dataset touches 5 artifacts.

Derived from recorded relationships, not a declared manifest.

Support Dataset v2the change
Checkout Qualityevaluates it
Refund Agentmeasured by it
Support Prompt v3evidence would be invalidated
2 knowledge objectswould lose their basis

The product

Five surfaces, one engineering record.

No surface owns its own data. Each is a different lens on the same record, which is why they can never disagree with one another.

What should I know this morning?

The record, briefed.

Brok opens on standing briefs read straight from the engineering record — what ran, what was promoted, what is failing, what knowledge was derived overnight. Ask it a question and the answer is composed from real rows, each statement declaring how it is known.

How it works
Brok's briefing surface, showing daily, deployment, incident, prompt, evaluation, dataset, knowledge and architecture briefs derived from the engineering record.

Captured from a running instance with seeded data. Nothing here is a mockup.

The engineering journey

A loop, not a set of pages.

Every surface in the product serves one step of this loop — and the loop closes. Learning becomes the precedent the next failure searches.

01

Problem

Quality dropped, cost rose, something failed overnight.

02

Execution

Run a reproducible evaluation with a pinned configuration.

03

Evidence

The result enters the record as an observation, then as evidence.

04

Knowledge

Where a decision and evidence both exist, a durable fact emerges.

05

Decision

Promote or don't. Brok refuses to bless an unmeasured revision.

06

Revision

A new version — and one honest sentence saying why.

07

Promotion

AI Git records what was promoted and what it superseded.

08

Deployment

Production runs it. A rollback is displayed as a rollback.

09

Learning

It becomes precedent. The next failure can find it.

The shortest useful loop: register, write a reason on every version, evaluate before you promote, investigate instead of re-running.

The full workflow

Comparisons

A different scope, not a better dashboard.

Tracing, evaluation and gateway tools are good at what they do, and several are more mature than this project. Most teams should run one alongside Broks Forge. The useful question is not which is better — it is what each one is for.

CapabilityTracing / eval toolsBroks Forge
Production request tracing
Evaluate against datasets
Prompt versioning
Decisions as first-class objects
Evidence linked to decisions
Engineering memory — why it changed
Artifact dependency graph
Precedent search over failures
Assembled root-cause investigation
Grounded Q&A with no LLM
Maturity and ecosystem

Generalised across the tools compared below; capabilities differ between them and change over time. Each comparison page is specific, and cites what the other tool does well.

vs LangFuseTracing & evaluation for LLM appsvs LangSmithLangChain-native observabilityvs PromptfooDeveloper-first prompt testingvs HeliconeLLM gateway & observability proxyvs Weights & BiasesExperiment tracking

Who it is for

Built for the questions asked after the demo works.

Getting an AI feature working is the first week. Explaining, defending and evolving it is every week after that.

AI / ML engineers

“Why did this fail, and has it failed before?”

Reproducible evaluations, a versioned history of every prompt and agent, and an assembled investigation instead of a log search.

Staff & lead engineers

“Can we defend this configuration?”

One engineering record to reason over — including precedent, contradictions, and decisions carrying no evidence.

Engineering managers & CTOs

“What can we actually prove?”

Evidence behind decisions. Which promotions are defensible, which are being carried on faith, and what the system cannot yet prove.

Teams inheriting a system

“Why is it like this?”

Engineering Memory. The reasoning behind the current state survives the person who made it.

FAQ

Frequently asked questions

Is this an observability tool?
No. Observability answers what happened. Broks Forge answers why the system is the way it is, what evidence supports it, and what to do next. They model different objects, and running both is sensible.
Is Brok a chatbot? Does it use an LLM?
Brok is not a chatbot and contains no language model. It resolves your question to one of 25 engineering intents by deterministic phrase scoring, then composes the answer from real database rows. Ask something the record cannot answer and it refuses.
Do I have to use a specific framework?
No. Agents are registered by HTTP endpoint. LangChain, LlamaIndex, a custom FastAPI service, a serverless function — if it is callable over REST, it works. There is no SDK to adopt.
Is it in my production request path?
No. Broks Forge is not a proxy or a gateway. It calls your agent endpoint during an evaluation and never otherwise, so it adds no latency to your traffic.
Is it open source? What does it cost?
Open source under Apache 2.0 and self-hostable with Docker Compose. No licence fee, no per-seat cost. You pay for the infrastructure you run it on and any model calls your evaluations make.
How mature is it?
Early. It is a complete, tested system — 499 backend tests running against real PostgreSQL — but it is a young project without the production track record or support organization of the commercial tools it is compared with.
What is the minimum useful setup?
An agent, a dataset, a prompt and one evaluation. That already produces observations, evidence, a graph, an AI Git timeline and answerable questions. The habit that matters most is writing one honest sentence on every version.
Why does it say “unknown” instead of “healthy”?
Because nothing has measured that artifact. Absence of failure is not evidence of health, and the platform refuses to imply otherwise. Run an evaluation and it becomes a real verdict.
Read the full FAQ

Give your AI system an engineering record.

Self-hosted, open source, and running in about fifteen minutes — most of it waiting on the first Docker build.

Get StartedRun it locallyRead the source

New to the category?What is an AI Engineering Operating System?

Brok's ForgeAI Engineering

An AI Engineering Operating System. Open source under Apache 2.0.

Product

  • What is Broks Forge?
  • The Five Layers
  • Brok
  • Root Cause Explorer
  • AI Git

Learn

  • Getting Started
  • Core Concepts
  • Engineering Workflow
  • Examples
  • Glossary

Compare

  • Overview
  • vs LangFuse
  • vs LangSmith
  • vs Promptfoo
  • vs Weights & Biases

Develop

  • Architecture
  • REST API
  • Data Model
  • Extension Points
  • GitHub

© 2026 Brok's Forge. Apache 2.0.

Built by Gokulraj