Deckent
Give the AI a goal. Deckent plans it, routes it to the right hands, proves the result and accounts for it.
Deckent is an Agent OS that plans, executes, supervises and proves the work AI agents do, from software development and product management to everyday office tasks and enterprise processes. The same core serves a single person working alone and the largest organisations alike. It is open source for individual use. For teams and organisations, the process, management and Enterprise layers come under a single product licence, with no per user fee.
Not being able to hand over a job end to end
Getting an AI model to write code, summarise a report or draft a message is easy today. What is hard is entrusting it with a job from beginning to end.
An assistant living in a chat window has a single context. One task at a time, one point of view, one provider's blind spots. Neither a development job touching five files nor an operations job that must pull data from three systems and prepare a decision fits into that window. When you force it, the work does not merely slow down; it becomes unreliable in places you cannot see.
The second problem is more insidious. The tool says "done" and you are left having to believe it. Yet a green test does not show the work is correct. The model's own report is not independent evidence. A table that looks plausible on screen does not mean the reconciliation is right. The more authority you hand to AI, the more an unproven "done" costs you.
Third, dependency. Model choice, execution environment, memory and audit records are increasingly packaged inside one provider's cloud. That arrangement is comfortable right up until the day that provider changes its pricing, has an outage, or your data must not leave the company.
These three walls are not only in front of software teams. The finance team trying to shorten the monthly close, the sales team preparing quotes, the HR team running a hiring process all hit the same thing: single context, unproven results, dependency on one provider. Today it is all handled by hand; copy and paste between windows, a glance at the result, carrying on with "hopefully that is right". Workable for one person. It does not scale for a team or an organisation.
An operating discipline, not a conversation
Deckent turns working with AI from a conversation into an operating discipline. That is exactly what "Agent OS" in the name means. Just as the operating system on your computer manages programs, Deckent manages AI agents. Who will work, what they may touch, how much they may spend, and what evidence will be required when they finish: all of it is settled in advance.
Using it is plain. You install it into your workspace, connect the AI providers and business systems you use, and write a goal.
A software team writes: "Add a refund flow to the payments module, write the unit tests, update the user documentation." An operations team writes: "Pull last week's open orders from the ERP, compare them against stock, draft purchase requests for the missing items and submit them for my approval."
Both goals go through the same loop, and the loop runs in the same order every time.
Understand
Deckent scans your workspace. For a codebase it reads the language, framework and test commands; for a business process it reads the connected systems, the data sources and the decisions already taken. It interprets the goal in that context.
Plan
It breaks the goal into small, interdependent tasks. Which task waits on which, which can run at the same time, which file or system will be touched: it produces a dependency map.
Route
For each task it selects the right agent and model. Work that needs deep design and reasoning goes to a strong model, repetitive routine work to a cheaper one, and data that must not leave the company to a local model.
Execute
Agents (Deckent calls them workers) run in parallel. Each worker can only write to the files or systems assigned to it and cannot touch its neighbour's area. Steps with an effect on the outside world wait for approval.
Verify
Each task's result is tested against criteria appropriate to the kind of work. For code, the project's own test and build commands run. For a report, the output is cross checked against the source data. For a document, the content is read against the criteria. The worker saying "I finished" is not enough; Deckent inspects the produced output, the run records and the result file itself.
Evidence and memory
What was done, by whom, under which authority, in how long and at what cost: all of it is recorded. Decisions taken and patterns learned are stored alongside the workspace. The next job is planned more accurately.
This loop is the heart of Deckent and is not left to the model's improvisation. The order is fixed, repeatable and auditable.
Code, product and enterprise processes
Deckent is not a coding tool; coding is only one of the kinds of work it covers. The same engine carries four areas at once.
Software and code
New feature development, bug fixing, code review, database migrations, writing tests, dependency updates, maintenance on legacy codebases. Each worker operates in a narrow area, the result is verified with the project's own test commands, and the change reaches you with its evidence.
Product development
The first release of a product from scratch, its documentation and its tests. Release notes, multilingual user documentation, design review, recording product decisions. Deckent's memory remembers the product's past decisions; if a new request contradicts a decision made three months ago, it tells you.
Everyday work
Summarising long email threads, drafting replies, updating checklists, turning meeting outcomes into tasks. Personal memory stays with you; every outward action (sending a message, changing a record) asks for your approval.
Enterprise processes
This is where Deckent meets Verhex's own area of expertise. Flows connected to ERP, CRM and HR systems: open order checks, stock and purchasing comparison, invoice and payment reconciliation, tracking hiring steps, periodic reporting. Every flow knows which system it talks to with which identity, on behalf of which department it runs, and at which step it must ask for approval. It runs on a schedule, resumes from where it stopped after an interruption, and writes every step to the audit trail. ERP integration is one of Deckent's principal directions; the beta programme advances together with these flows. System connections are built with MCP tools, the HTTP API and process specific connectors. Verhex's principle holds here too: process first, system second.
One engine, three faces
You use Deckent in three forms; all three rest on the same core.
As an assistant
You plan by talking, ask questions and give approvals. This is the side that turns your intent into structured work.
As a workforce
Agents running in the background and on a schedule: code, tests, documentation, data work, repetitive operations, business system flows. This is the side that executes accepted work within its boundaries.
As a platform
You add your own agents, skills, providers and system connections. This is the durable orchestration, memory, approval, routing, recovery and audit infrastructure.
The primary surface is the terminal. It gives full control, avoids unnecessary visual load and reveals information progressively. The desktop application is the visual face of the same capability. Telegram, Discord and WhatsApp connectors, the HTTP API and the MCP interface (the standard by which AI tools connect to each other) use the same core through different doors. The observability dashboard only observes; it never decides.
Who did what, under which authority
What separates Deckent from its peers is not speed but accountability. "Governance" here is not corporate decoration but part of how the work is done. For one person it is a safe working habit; for an organisation it is compliance and audit infrastructure. They are the same mechanism.
Chain of authority
Every effect in the system is bound to a chain: from goal to task, from task to attempt, from attempt to a single operation. Look upward and you see which goal a change served; look downward and you see the action that carried it out. When one agent delegates to another, its authority narrows and never widens.
Write boundary
Each worker receives the minimum set of files and systems the job requires. An attempt to write outside that boundary is flagged and the job's result fails. You do not get a surprise file change or an unexpected system record.
Approval gates
Risky operations (deletion, writing to a business system, sending a message to a customer, exceeding the budget) wait for your approval. Approval is given only in the interactive command line where your identity is verified. Remotely connected tools can see the approval inbox but cannot grant permission. There is no path by which an agent approves itself.
Budget and cost
Every run's budget is known from the start. Provider usage is measured task by task, reported, and attributed to the goal, the project and the department. You do not get a surprise at the end of the month.
Audit trail
Closure records are written to an append only ledger that cannot be deleted, and sealed with a digital signature. The signing key is held separately from the system, in its owner's hands. There is always an answer to "who made this change, when, and on what basis".
The honesty rule
If Deckent cannot do something, it says so plainly. An unsupported platform, an unreachable provider or an insufficient resource does not quietly divert to another path; it says "on hold" and writes the reason. A failed run is never relabelled as successful.
Provider independent, your data stays with you
No AI provider is part of Deckent's identity. You use OpenAI, Anthropic, Google or an open model on your own server under the same core. Which work goes to which model is determined by your configuration, live access status and cost policy. If a provider has an outage, that is not a stop but a routing decision.
Your data stays with you as well. Workspace state, memory, evidence and work outputs live on your own disk or your own infrastructure. The memory layer is a local database with full text search. Running Deckent does not require a Deckent cloud; you can work without moving regulated data outside the company.
The verifier differs from the producer
Having the same model audit work it produced is a closed circuit; a model cannot see its own blind spot. Deckent requires verification to always be performed by a provider different from the one that produced the work. Verification by the same provider is refused. If there is no second provider the result is "on hold"; no false approval is manufactured. The same rule sits at the core of Verhex's Xerify tool.
See XerifyWho it is for, and under which licence
Individual use
For a single person working on their own projects and their own work, Deckent is open source. The core is complete; it is not a limited trial. You give the goal, parallel workers run, and the result reaches you with its evidence.
Teams and organisations
For team and organisational use, Deckent's upper layers come into play: business process flows and enterprise system connectors, multi tenant administration, role based authorisation, integration with corporate identity and security systems, policy packs, central budgeting and audit. These layers are licensed as Verhex Enterprise. The Enterprise layer does not invent security absent from the core; it carries controls already structural in the core to organisational scale. Setup, process analysis and integration are carried out together with Verhex consulting.
Licence model
The licence is per product, not per user. One licence covers an unlimited number of users in the organisation. When you spread AI from a team of ten to an organisation of a thousand, the invoice does not grow; adoption is not penalised, it is encouraged.
Where it stands today
Deckent is at version 0.100.0 and in beta. It is written in TypeScript and requires Node.js 24 or later. macOS, Linux, Windows and WSL2 are the target platforms; continuous integration runs on Ubuntu and Windows.
In 0.100.0 the product offers more than 81 command line commands, 51 MCP tools and 8 MCP resources, 22 built in agents, 35 built in skills and a 20 page observability dashboard. Documentation is available in English and Turkish.
The release process is deliberately human approved. Continuous integration builds, verifies and produces supply chain attestation; only a human publishes a release.
Most importantly, Deckent builds itself with itself. At Verhex, Deckent's development is run through Deckent's own runs. More than seven hundred numbered sprints have gone through this core. Failed runs remain recorded as failed, together with their root cause. Every friction you would meet, we meet first in our own work.
Join the beta
Deckent is in active development and moves forward together with its beta participants. If you would like to try it on a codebase, a product process or an enterprise workflow, fill in the form and we will discuss your scope and use case together. Early participants have a direct say in how the product takes shape, from routing rules to approval flows and process connectors.
Apply for the beta