Frontier-model AI, agentic applications, and generative video — built into systems that do real work.
Empower CLI
Agentic AI · Release 2026.8.11 · Built June 2026
A command-line client for Encore Analytics Empower. It makes the same reads the Empower screen makes, under your own credentials, from outside the browser. Nothing is installed into Empower and nothing on the server changes.
Reports, charts, views, dashboards and schedules come out as Excel, CSV, JSON, PDF, HTML or image. The server declares which formats an artifact supports; the CLI does not assume
Every output carries the contract, the status period and how it was obtained, with a byte count and a SHA-256
bundle verify re-proves a package offline, with no Empower access at all
Read-mostly. The only write path is Empower’s own Input functions, and only the workflow transitions Empower says your account may perform
Custom report and chart ids belong to the user and the server that made them, so commands discover what exists before they act
Profiles name where credentials come from — an environment variable, a protected file — never the values themselves
Python 3.10–3.14, one runtime dependency, 986 tests
Source: the Empower CLI documentation (docs.kuon.ai/cli.html) and the repository README
Empower CLI Trainer
Agentic AI · Release 2026.8.18 · Built July 2026
An agentic development tool for testing CLI functionality while building applications. Pick a contract, an artifact and a scope; the Trainer composes the command, runs it against a real server and returns the actual payload — so a command and what it produces are confirmed before an application depends on them.
Every artifact family — reports, charts, views, dashboards, schedules — can be exercised against a real contract, with the output rendered rather than described
Selector lists are discovered live from your own server, so what is offered is what your account can actually reach
The write side covers all ten Empower Input functions: variance narratives, scope of work, user EAC, EAC narrative, VAR required and categories, action items, and the action-item and VAR kanbans
Before offering any action it asks Empower which workflow transitions your account may perform on that element in its present state. Only those appear
No offline draft, no local copy, no queue of pending saves. A commit is one submission, never retried, followed by a complete re-read of the record
It runs on your own machine and reaches Empower only through an installed Empower CLI, pinned to a specific release — a mismatch stops it starting
The browser never receives an Empower password; credentials stay in the CLI’s protected storage
Windows, macOS and Linux, with the Windows path verified end to end on Windows Server 2022
Source: the CLI Trainer documentation (docs.kuon.ai/trainer.html) and the repository README
Empower OData
Agentic AI · Release 2026.8.22 · Built August 2026
Connects to the Empower CLI and republishes what it reads as a read-only OData v4 service, so Power BI and Excel pull from Empower in real time — or the same path feeds a data warehouse and the dashboards built on it. The CLI is the only component that talks to Empower.
The failure it is built against is quiet: a Power BI card showing a CPI that disagrees with the Empower report open in the next window, and nothing on either screen saying which is right
Three distances from the server rather than competing options — read on demand at refresh, extracted to files on a schedule, or exported once by hand
Every view is read twice: once through the path the feed publishes, once as Empower’s own Excel export of the same view and period, then compared cell by cell. Columns Empower renders as colour or arrows are reported as not verified, never counted as passes
Every refresh sums the lowest-level elements against the contract total Empower reports for that period. If they disagree the refresh stops and publishes nothing
CPI is total BCWP over total ACWP and SPI is total BCWP over total BCWS, rebuilt from summed components each time a filter changes — never an average of account-level indices
A view’s columns and types are approved by a person once and then frozen. A frozen column that disappears or changes type stops the refresh; a new column is recorded as a note
What exists is asked of Empower live, under the reader’s own authorization, every time. Nothing carries a stored list of view identifiers
Read-only OData v4, so Excel and anything else that speaks OData read the same service. Zero runtime dependencies
Source: the Empower to Power BI documentation (docs.kuon.ai/powerbi.html)
RapidClaw
Agentic AI · Built June 2026
What the CLI surface makes possible, taken to an interface Empower could never have offered. You name the contract, the status period and the WBS element, and RapidClaw answers with cumulative and current-period CPI and SPI at that level, then names the work packages driving the variance — with the matching Empower chart on the watch pane, so the spoken number and the displayed number are the same number, at the same level, for the same period.
A spoken request has to resolve to a real artifact, in a real scope, for a real status period before anything renders. Where it cannot, it says so rather than describing a chart it never fetched
Cost and schedule status first, then the lowest-level work packages driving the variance, each named with its own CPI and CV
CPI, SPI, EAC, variance and corrective-action questions in plain English
The chart is visualized in the browser from the raw Empower chart payload; the native chart-data XLSX remains the source artifact
Data-window scope and watch-pane scope are held separately, so the spoken answer and the displayed chart cannot drift to different elements
Re-engineering access is only half of it. An agent also needs a world it cannot step outside. The ontology is a declared, typed set of objects and a fixed verb list, with a parser that refuses anything not named in it.
A declared, typed set of objects every agent can see and act on: sessions, jobs, runs, tools, channels, configs, skills
The same small set of verbs on every object. Commands, state, audit, rollback and approval are all named in the declaration
Parser-enforced, replayable and scoped per agent, refusing anything the agent cannot name. The capability is absent from the surface rather than rejected at runtime
Where most agent platforms treat ontology as a data layer behind a query language, this one treats it as the operational layer the runtime executes
Agentic AI · Begun May 2026 · Rebuilt August 2026 · In build
Dungeon Claw uses frontier-model agentic AI to place multiple independent agents inside the legacy framework of Zork. They collaborate across narration, characters, memory, and voice to create a deeper, more characterful experience — while preserving the rules, mystery, and nostalgia of the original.
Zork was written on a PDP-10 mainframe in the late 1970s; this is that game reconstructed as a collaborative agent experiment
Agent memory persists across a long session, and imagination is bounded by what the engine will declare true
Several agents run concurrently and hand off to each other; conversation routing carries no hardcoded lists
Nothing an agent believes becomes a fact unless the engine states it
Agentic AI · Published on ClawHub · v1.0.1 · MIT-0
Launch verification and watchdog discipline for delegated, backgrounded, browser-driven, and long-running tasks. A backgrounded agent that dies is silent by default; this makes it loud.
Prevents false "running" claims, which is how an orchestration reports success over a dead job
Clarifies failed_to_start against failed_after_work_started, because those two need opposite responses
Treats heartbeat as a watchdog rather than the main progress loop
Published on ClawHub, version 1.0.1, MIT-0, security audit passing
Generative AI video · Productions from March 2026 · Board data 15 April 2026
Agents handing typed artifacts down a line, in a domain with no legacy application at all. Five agents carry a production from a one-line brief to a finished cut: screen writer, casting director, location scout, cameraman, film editor. A board tracks every asset each stage produced.
Each stage hands the next a typed artifact, so a failed production names the stage that failed rather than the pipeline
28 productions and 391 tracked assets as of 15 April 2026: 25 screenplays, 53 character references, 217 keyframes, 86 filmed scenes, 10 final ensembles
Character references are versioned per project and reused across its scenes, keeping a character's appearance consistent between shots
The board is reconciled against the physical project storage, so it cannot list an asset that is not there
1920×1080, H.264. The still shown here is one of those outputs
Source: the production board
Runtime
Backing OpenClaw and open source
KUON.AI is a committed supporter of OpenClaw and open-source software. Open tools, transparent command surfaces, and shared infrastructure are essential levers for advancing agentic AI from frontier models into useful real-world applications.
