WCKD is a protocol for getting anything made — software, hardware, a product, an operation — by a council and a crew of specialized agents. First it learns what exists and why the dead attempts died. Then it judges every capability, designs only what is missing, prices it, proves it, breaks it, and ships it with a rehearsed way back.
A build that could have been a download, a purchase or a configuration change is a failure, even if it works.
What do we want, to what standard? A brief with capabilities and a Bar: a threshold and a measurement method per dimension, signed by a human.
brief.jsonWhat exists, what works, what died and why? At least five source classes per capability, the top candidates actually tried, and the Graveyard mined for the real spec.
landscape.jsonPer capability: keep, bend, build, wait or kill. BUILD must say why not KEEP and why not BEND. Only the gap is designed, and everything is priced with a confidence class.
verdict · blueprint · cost · prospect · planBuild, prove from clean, break as an adversary, ship with a rehearsed rollback, then keep it with an owner, SLOs, a budget and lessons that make the next project cheaper.
workorders · proofs · breaches · release · runbook| Gate | Question | Human veto |
|---|---|---|
| G-W Briefed | Is the want measurable and the Bar written? | yes |
| G-C Canvassed | Did we look wide, try things, and mine the Graveyard? | — |
| G-K Judged | Every capability judged, only the gap designed, everything priced? | yes |
| G-D1 Built | Does every work order have a build log? | — |
| G-D2 Proven | Proved independently, from clean, against the Bar? | — |
| G-D3 Hardened | Attacked, and every finding dispositioned? | — |
| G-D4 Shipped | Rollback rehearsed and monitoring live before cutover? | yes |
| G-D5 Kept | Owner, SLOs, opex, lessons, ledger row? | — |
One agent per role, one artifact per hand-off. Whoever builds does not prove; whoever proves does not break; whoever breaks does not fix. Independence is structural, not a promise.
| Transport | For | Start |
|---|---|---|
| CLI | a shell, a person, a script | bin/wckd.py — stdlib only |
| Handoff | a model that only reads text | wckd handoff aqm --inline — SOUL, role, gate criteria, inputs, output path, schema |
| MCP | Claude, Cursor, Windsurf, any MCP client | bin/wckd_mcp.py — 12 tools over stdio |
| HTTP | a custom loop, n8n, a service | wckd serve --port 8787 — REST, optional bearer token |
| Orchestrated | “run it against my agent” | wckd run aqm --agent "claude -p" — stops at every human gate |
Both realms carry their own obligations: digital
(itemised run cost, priced exits, watched kept dependencies, named environments) and physical (certification,
hazard analysis, pilot batch, rehearsed recall, recalibration). machine/ is the protocol compiled to JSON —
phases, gates, roles, tolerances, campaigns, physical obligations and one self-contained schema per artifact —
so an implementation in any language gates at the same place. wckd conformance proves the schemas
agree: 16 cases, valid artifacts must pass, mutants must fail.
# the protocol is files; the CLI enforces the gates pip install pyyaml jsonschema git clone https://github.com/WCKDai/wckd && cd wckd make test # 47 tests, including a full W→D5 run wckd new aqm "A home air-quality monitor that alerts my phone" wckd next aqm # who runs next, what to hand them, what to write wckd gate aqm G-W --by "your name" --human # human-veto gates: G-W, G-K, G-D4 wckd score aqm # GOALS.md G1–G12, scored from the artifacts # kill is a win, and it has to be recordable wckd gate aqm G-K --by "you" --decision KILL --note "already solved; nothing to build"