An operating system for running a personal team of AI agents. KRANG owns the context, the memory, the permissions and the state. Claude, Codex and local models do the work, and you can swap them without rebuilding anything.
$ krang
.-~~~~~~~-..' ~ ~ './ ~ (o o) ~ \| ~~ \___/ ~~ || ~ K R A N G ~ |\ ~~~~ ~~~~ /'. ~~~~~ .''-._____.-'instance personal
providers claude ok codex ok ollama okboards engineering 4 open · business 2 open
running rocksteady → card e7 (feat/legal-pages)
needs you 1 decision waiting
$ krang briefing
0runtime dependencies
15,409lines of TypeScript in the core
19decision records, including the wrong ones
1machine. State on disk. No cloud service required.
Five roles, no fixed model
Each role has a fixed remit. The model behind it is whatever you configure.
Change provider and the roles stay. That is the whole point.
KRANGThe brain
Holds context, decides what needs doing, delegates, tracks what is open. Never does the specialist work itself.
SPLINTERBusiness
Why are we doing this, what does it cost, what is the risk.
ROCKSTEADYEngineering
Code, git, PRs, tests, deployment. Works through whichever coding model is configured.
BEBOPResearch
Gathers evidence, compares options, cites sources. Cheap tier by default.
SHREDDERRed team
Assumes the others are wrong and tries to prove it. Strongest model available. Invoked deliberately, never on routine work.
SHREDDER is the role most systems do not have. Every other role is cooperative. For code that is survivable, because tests are an independent adversary. For a VAT position there is no CI, and a confident wrong answer travels unchallenged until it becomes a filing.
So the adversary is structural, not a disposition the other roles are asked to adopt. An adversarial pass on everything trains you to skim it, so SHREDDER is reserved for the work that matters.
How it works
One core. Any number of instances.
The core never names an instance
CLI, orchestration, routing, retrieval, policies and skills take an instance name as input and assume nothing about which one. Adding a second instance is a sibling directory, not a change to the engine.
The boundary is enforced, not documented
A pre-commit neutrality check scans tracked content for client names, product names, home paths and credential shapes. The patterns it hides are supplied at runtime and never committed.
State lives in append-only files
Ledger, tracker and boards are plain files on disk. Nothing financial is committed. Nothing about a client is committed anywhere.
# daily use
krang # the status screen
krang ledger add # log an achievement, decision or result
krang track add # capture a requirement before it's lost
krang track ls # what's open, by status and priority
krang briefing # priorities, open items, recent ledger
krang board ls # cards, per domain
krang run once # claim a card and run it# layout
krang/
├── core/ engine: never references an instance
└── instances/
├── personal/ your life, company, side projects
└── work/ same shape, stood up when ready
Honest status
What is built, and what is not.
WORKING TODAY
The CLI and its stores: ledger, tracker, boards
The supervisor that claims cards and runs them against local models
Retrieval with a scored evaluation harness
Workspace creation with git worktrees and per-workspace ports
Usage analytics and launchd scheduling
NOT BUILT YET
The provider layer. The five roles are prompt templates and a declared vocabulary until the model abstraction lands.
Routing beyond two local tiers behind one host.
Decision record D-0017 records that gap rather than papering over it.
Decisions
Every decision, with the context that produced it.
Nineteen records so far. They carry the reasoning, including the ones that turned out wrong. If you want to know why KRANG stores state in files rather than SQLite, or why it runs on your own hardware, the answer is written down.
D-0003KRANG is the orchestration layer, not a client of one
D-0005Runs on own hardware, costs nothing, keeps data local
D-0013Boards are directories and append-only logs, not SQLite
D-0017Five named roles, and one of them exists to attack the others
D-0018Providers are a chain resolved late, and availability is inferred rather than asked
Learn to run your own
A founding cohort for professionals who want an AI brain they own.
The code is free. The cohort is how you stand it up for your own work in weeks rather than months: live sessions, the setup done together, and a community brain trained on what the group learns. Founding places are limited and priced once.
KRANGBuilt by cteic Ltd, Dublin. Apache-2.0. The name and marks are not licensed.