EKOS — Enterprise Knowledge Operating System

The compiler enterprise knowledge needs before AI can be trusted with it.

Every AI agent an enterprise deploys today reasons over raw files, stale wikis, or a vector index that can't tell fact from guess. EKOS compiles the enterprise itself — code, SQL, ETL, Git history, docs — into an append-only, evidence-backed knowledge ledger. Every answer an agent gives cites the exact source it came from, or it doesn't answer.

Enterprise systems Knowledge Compiler Evidence-backed ledger MCP Any AI agent
§ 01 / problem
The knowledge is already there. Nobody compiles it.

Enterprises don't lack information. They lack a compiler for it.

Business logic sits in a Pentaho job nobody's opened in years, a stored procedure with no comments, a Confluence page that stopped being updated the day its author left. Every system holds a partial, decaying truth — and every AI assistant pointed at that mess inherits the same fragmentation, then papers over the gaps with a confident-sounding guess.

  • ScatteredSQL in one repo, ETL XML in another, tribal knowledge in nobody's head anymore
  • UntraceableA dashboard number nobody can trace back to the transformation that produced it
  • PerishableThe one engineer who understood the legacy pipeline leaves; the code stays, unread
§ 02 / why now
The AI-native inflection point

Agents don't fail because they're not smart. They fail because they're not grounded.

Every enterprise is wiring AI agents into their workflows right now — and every one of those agents is only as trustworthy as the context it's given. Vector-search RAG retrieves similar-looking text; it can't tell you a fact is true, only that it's nearby. MCP (Model Context Protocol) is emerging as the standard wire format for how agents reach tools and data — but a protocol is only as good as what's on the other end of it.

  • RAG's limitSimilarity is not truth — a fuzzy match is not evidence
  • MCP's gapA universal wire protocol with nothing structurally trustworthy behind most of its endpoints yet
  • EKOS's betBe the evidence-backed thing an agent talks to over MCP — not another retrieval index
§ 03 / product
How it works

Six verbs. Every one deterministic. Nothing silently dropped.

01
init
02
build
03
recover
04
resolve
05
compile
06
commit
§ 04 / proof, not a slide claim
Tested cold against real code, not fixtures

Cloned two real GitHub ETL projects. Ran the full pipeline. Cross-checked every claim.

transforms/load-fact-coded-values.ktr
Object :0  node_type: "Source"
  object_name: "omrs_obs"
Object :1  node_type: "Sink"
  object_name: "fact_patient_coded_value"
  columns: [patient_id, gender_id, ... ]

The real Kettle steps: Import Coded ObsWrite Coded Obs. Exact match — parsed, not generated.

375
objects, one real repo
100%
SQL nodes mapped
0
facts fabricated on parse failure
§ 05 / ai interaction log — part 1
Claude Code, using EKOS's own MCP server, found a real bug in EKOS

Not a demo. A live audit, by an Anthropic model, of its own infrastructure.

This deck itself was built inside a Claude Code session backed by Anthropic's Claude models. While proving EKOS's Pentaho/SQL recovery worked, that same session queried EKOS's ekos_dependents MCP tool on a table it had already shown a real ETL job writes to — and got back dependents_count: 0.

live JSON-RPC over stdio — before the fix
→ tools/call ekos_dependents { "id": "...fact_patient_coded_value" }

← result { "target": { "kind": "Table" },
           "dependents_count": 0, "dependencies_count": 0 }

No hand-waving, no "that's probably fine" — the agent treated its own product's wrong answer as a bug to root-cause, the same way it would for a user's code.

§ 06 / ai interaction log — part 2
Root-caused, fixed, tested, then re-verified live

The agent didn't just patch the symptom — it found two independent bugs.

  • Bug 1Cross-system identity resolution skipped the single most confident match it could ever see — an exact name shared between a SQL table and the Pentaho step that writes it
  • Bug 2Impact/dependents queries had no concept of "unconfirmed" — thousands of unreviewed hypotheses were being traversed as if they were facts
  • FixTwo small, targeted patches + 6 new tests, written before the fix, per the repo's own test-first workflow
same query, after the fix + human-style review
→ tools/call ekos_dependents { "id": "...fact_patient_coded_value" }

← result { "dependents": [{
  "name": "load-fact-coded-values.ktr:1",
  "relationship": "SameAs",
  "properties": { "confidence": 0.95,
    "status": "confirmed" } }],
  "dependents_count": 1 }
§ 07 / engineering discipline
Why a technical buyer should trust this codebase

RFC-first. Test-first. Append-only. Zero shortcuts on the ledger's invariants.

738
tests passing, 0 failing
43
accepted RFCs
43
devlogs — full build history
0
unsafe without a written justification

Every capability is designed in writing before it's built. Every devlog captures what shipped, why, and what broke — including this session's own bug fix. Nothing here is "trust me" — clone it, run cargo test --workspace, read the RFC.

§ 08 / who this is for
Positioning

For teams who need an AI agent to be right about their business, not just fluent about it.

Data & platform teams

Inheriting legacy ETL (Pentaho, SQL, stored procs) with no owner and no docs — get an evidence-backed explanation instead of reverse-engineering XML by hand.

Teams building AI agents

Need grounded enterprise context, not another vector index that can't distinguish a fact from a plausible guess.

Engineering orgs, generally

Any team where "what does this actually do, and how do you know" keeps landing on a person instead of a system.

Not for

Teams wanting a BI dashboard or a data catalog with a nice UI — EKOS is the compiler underneath that kind of tool, not a replacement for one.

§ 09 / where this stands today
Honest about the stage

Early. Open source. Built in public, bugs and all.

EKOS is early-stage and open source — this deck's own §05–06 is the point: the product is far enough along to be genuinely useful, and immature enough that testing it against real, uncontrolled data still finds real bugs, which get fixed in the open. Real/tested connectors today: file, Git, GitHub, Confluence, local docs (PDF/DOCX/etc.), Pentaho, crypto, Python/PySpark (real AST + DataFrame chains), Rust source (real AST + function-call graph). Several more (Salesforce, SAP, Oracle, Snowflake, Fabric) are scaffolded but not yet proven against live accounts. A built-in secrets/PII redaction baseline (RFC 0043) runs on every connector automatically, on by default, not something a workspace can turn off.

Design partners wanted· Contributors wanted· Feedback wanted
Try it

Point an agent at your enterprise's real knowledge — not its raw files.

terminal
$ git clone https://github.com/alexeyban/EKOS.git && cd EKOS/ekos
$ cargo build --workspace

# compile a workspace once
$ ekos init && ekos build && ekos recover && ekos resolve && ekos compile && ekos commit

# hand it to Claude, or any MCP-speaking agent
$ ekos mcp serve --workspace .
EKOS · Rust 2024 · MIT-licensed · github.com/alexeyban/EKOS · @ekosproject for updates