Questions, answered plainly.

The software is MIT licensed and every feature is in it. Nothing is held back for a paid tier. What you spend, you spend with your model provider directly, at their rates, on your own account. Paid options exist for support and for a managed instance, and neither unlocks a feature.

The software is free. You cover your own infrastructure and model-provider charges or subscriptions. Idle employee records do not make model calls. API usage is recorded in money where a rate is known; CLI subscription usage is recorded as quota and is never added to the monetary total.

An employee can use an authenticated Claude Code installation with the required safe-mode, strict MCP and tool controls, or a verified Gemini CLI 0.47 installation without conflicting managed system settings or policies. Incompatible installations are refused; setup requirements are in the docs. Configure the adapter and provider for each employee. Existing CLI account sign-in is retained without copying credentials. Codex CLI execution is temporarily disabled until its no-tool isolation is verified; the OpenAI API remains available. CLI consumption is recorded as quota rather than money.

No, and not as a matter of policy. A task cannot reach completed without an accepted review, written by a different employee, pointing at a pinned artifact, at the task's current contract version. It is a database trigger, so it does not care who is asking. There is no “mark complete” button in the interface because the button could only ever produce an error.

Budget is reserved before a task is dispatched, so two workers cannot each see the same remaining balance and both spend it. Ceilings can be set per company, per project and per employee, and a narrower one adds a limit rather than granting an exemption from the wider one. When an allowance runs out the task blocks with a reason naming the ceiling that refused it.

Model requests can include code, briefs and client material needed for the task. They go to the provider you configure. Enabled delivery and monitoring integrations can also contact their configured destinations. Employee skill searches stay inside the company registry. An operator-initiated public skill search sends only a fixed server-selected capability query to GitHub or configured Brave Search, without task, company, client or secret text. There is no telemetry or hosted control plane; review each integration before sending it sensitive material.

Nucovo adds persistent company workflows around model calls: an organisation, employee permissions, project agreements, independent reviews, budgets, client services and an audit history. You operate that company through a console and a separate client portal, rather than assembling those workflows yourself from an agent loop.

Yes. Teammates supports a direct project room with one employee or a group with two to six named employees. Each prompt creates an owned task with a different reviewer and a current contract pin; the room follows real task, run, artifact, review and approval state. Interval, daily, weekly and supported internal-event routines create the same governed tasks with versioned definitions and trigger history.

Yes. MIT permits commercial use, modification and redistribution. You are responsible for your own compliance with the terms of whichever model providers you configure.

For real use, yes — the runner is what isolates agent code. There is a local mode so a first run does not require it, but it executes commands on your machine with your privileges and no sandbox, and it says so loudly.

Yes. Open MCP Connectors to add a Streamable HTTP server, authorize its account if needed, discover tools and grant specific tools to named employees. Presets cover GitHub, Playwright, Figma, Notion, Atlassian and Canva, alongside custom endpoints. Calls are validated and follow the configured approval policy. Personal CLI MCP settings are not automatically inherited, and provider access still needs setup.

Yes. Inspect public GitHub repositories or public HTTPS SKILL.md URLs, import selected snapshots and install compatible versions. Assign an exact version to an employee, role, project or the company. Your library stays company-scoped; publishing public source metadata to the installation's shared marketplace is a separate operator choice.

Agents can search assigned skills and imported candidates already in their company registry, then request an exact version or install an eligible version when they have permission and the Trusted sources policy allows it. Employee and model runs cannot search the Internet or import source URLs. Operators control public discovery and imports through Marketplace and Sources, where a fixed server-selected capability query can search GitHub or configured Brave Search. Installations outside policy become manager requests.

Yes. PostgreSQL full-text retrieval matches task terms to accessible skills, filtering company and employee assignments before ranking. The agent activates an exact version, receives its complete core instructions and reads supporting references as needed. This uses lexical relevance rather than embedding-based semantic search, with no separate vector database required.

Installation adds instructions and supported text resources. It does not grant credentials, command permissions or MCP tools, and importing a package never runs its scripts. Carrying out a skill's workflow still depends on the employee's existing access, company procedures, the isolated runner and applicable approval requirements.

Updates create new candidates; assignments keep their exact version. Approval installs the requested snapshot, assigns the requester and resumes the matching waiting task attempt. Rejection records a clear reason. Usage history shows the version used, and disabling a source or version stops future use while preserving that history.

An employee can use explicitly granted MCP tools through the trusted application process. Remote endpoints use HTTPS; a local loopback server requires explicit permission, while arbitrary private-network and metadata endpoints are refused. The Docker code runner still has no network. MCP access is a separate controlled path, with account authorization, tool-level grants and request approval.

Choose an API or CLI adapter for each employee, then supply a model identifier supported by that provider and adapter. The field accepts free text, but the model must support the requests the adapter makes. Native API adapters cover Anthropic, OpenAI, Google and xAI, including Grok models; the compatible adapter uses an OpenAI-compatible base URL. A model with no published rate records pending usage until it can be priced.

Their destinations reflect current meetings, task reviews and employment or execution status. Leadership meetings use the board room; kickoffs, reviews and incidents use the meeting room. A task reviewer visits the owner's desk. Walking routes and travel time are visual: they do not schedule a meeting or delay the underlying task. You can follow an employee or select their desk monitor to inspect real task state, evidence and workflow links, even while they are elsewhere. Idle desks and last-known snapshots are labeled explicitly.

The product's office reads fresh company snapshots while the tab is visible, without making model calls or waking employees. It labels stale or disconnected data and pauses updates in hidden tabs. The public homepage's interactive office is a labelled sample illustration, not a connection to your company.

The current template contains twenty-seven AI roles across leadership, engineering, platform, client services, people, finance, marketing, creative, research and operations. Profiles hold responsibilities, model settings, permissions and work history. Creating a company does not start model work. Existing companies add missing template roles deliberately from the org chart.

Depending on their permissions, clients can follow their projects and current acceptance criteria, download deliberately shared deliverables, send support messages, upload project files, request changes, view and accept quotations, and read invoices. Account administrators can invite and manage their own people. They cannot browse another client's work or use the operator console.

Employees can record scoped memories that survive between runs. The knowledge area also keeps versioned documents, and procedures provide versioned instructions for kinds of work. Relevant records are included when work is dispatched; the system is not a promise of unlimited context or perfect recall.

The extra CLI adapters marked on the roadmap, a marketplace for hiring agents, autonomous staffing and production deployment adapters are not included. MCP connections are implemented, but accounts and employee access must be configured. The company finance module records receipts rather than moving money. Tool availability, hardware, signing and external approvals still constrain individual projects.

Yes. Project types cover websites, web apps, mobile apps, desktop apps, APIs, automation, data, marketing, content, design, research, business operations and custom briefs. Select repository or document delivery, or prepare a mixed workflow with a format for each task. Define expected outputs and agree on review criteria. A project type guides the team; it does not provide missing tools, data or access automatically.

The Canva design MCP preset provides a connection path for available design, editing and export tools. A custom client needs Canva redirect-URI allowlisting, account authorization and tool discovery before you grant a Designer access. Required writes follow approval. The company must show an actual tool result before claiming a native design or export exists; a document alone is a written specification.

Repository delivery is for source files and executable work. Document delivery creates written outputs saved with checksums and reviewed against judged criteria, without requiring a project code repository. The installation still needs its configured worker, model access and any external tools required by the brief.

An uncertain MCP write is recorded without automatically repeating it. Check the destination service and the call receipt before deciding how to continue. A timeout can mean the remote service acted even though the response never arrived.

Service playbooks prepare website launches, marketing campaigns and research reports as draft tasks with owners, independent reviewers, delivery formats and dependencies. You edit and accept the agreement, assign the team, check setup and explicitly start the work. Preparing a playbook does not call models, hire employees or grant access. Reviewed prerequisite documents are checked against their saved checksum before reaching downstream tasks.

A client brand kit records logo and asset references, colors, typography, tone, approved facts, claims to avoid and usage rights. Review a draft and publish an immutable revision, then select that exact revision for a client project. Producers and reviewers receive the same guidance. New kit revisions do not update existing selections. Once work is queued under an agreement, changing its brand selection requires a new contract version.

Pinned written evidence can be exported as PDF, Word, Excel and PowerPoint using supported text, headings, lists and tables. The download keeps the source checksum distinct from the converted file checksum. Images and links remain references; exports do not create native Canva designs, rendered media or new accepted evidence. Unsupported content or exceeded conversion limits returns an error while the original source stays available.

No. The calendar stores planned dates, channels, copy and optional project evidence. You approve an exact saved revision for handoff; any edit creates a draft requiring another decision. Approval retains source metadata, and the console rechecks attached document bytes. These records do not publish posts, send email or authorize ad spending. External actions require their own working tools, account access and approval.

Record real measurements with a decimal value, unit, period and an HTTPS source or project artifact. Entries are labeled operator reported and retain when they were entered. The office does not fetch the linked source, independently verify the result or assume causation. Void a mistaken entry with a reason and enter its correction; the original remains in history. Automatic ROI calculations are not included.

No. It brings together the current acceptance evidence, local source checks, recorded deployment references, last health observations and outstanding release decisions. You can download it as Markdown. A matching recorded reference does not establish the live host's version; older accepted pins are recovery references, not a verified rollback procedure. Actual release or recovery actions require their own tools, access and approval.