Balanov/baiagent-director

BAIagent Director is a business-advice application: given a company website, it studies public pages and outside mentions, separates facts from hypotheses, links sources, and helps build a measurable pilot and implementation outline. It suits founders or managers who want a concrete next step with cited evidence rather than an automatic profit promise. It is not an autonomous employee — it does not operate a CRM or execute business operations, and its recommendations remain hypotheses until checked against real data; a citation only confirms text appears in a source, not that the claim is true or commercially useful. Fresh market collection needs internet access.

How the advice is produced

The advisor looks at products, services, prices, competitors, reviews and market signals, and frames findings as causal chains: observation, mechanism, action, and how to test it repository. Users can also enter their own hypothesis through a button or dialogue command, then save it, select it and build a plan. A separate implementation view adds stages, rough timelines, owners, a budget formula and a stop criterion. After the task is defined, it can suggest solutions from a verified BAIagent catalogue — but a catalogue recommendation is not the same as an already running autonomous agent.

Three entry points, different capabilities

The project shares one codebase across BAIagent, web-audit and an extended Balanov.PRO enterprise audit, but the README warns these use different entry points and server setups and should not be treated as interchangeable setup and development notes. The Windows desktop app adds work with selected local documents (TXT, Markdown, CSV, JSON, PDF, DOCX) and a local dialogue with Qwen, either through an existing local Ollama installation or by downloading a Qwen3-4B model — no mandatory cloud API key. The web audit is deployed to PHP 8.1+ with cURL, DOM and SimpleXML and calls a research endpoint; it does not claim to run Qwen locally in the browser. Public web-audit entry is available web audit.

Requirements and practical constraints

• Windows install: the two installers are separate and unsigned (no Authenticode), so Windows may warn of an unknown publisher. Director is at 1.2.0; Enterprise is 1.1.1 and is not included in that release installation section.
• Disk and memory: on first launch without an existing Ollama, the app may download about 2.4 GB of model files; the README recommends at least 8 GB RAM and free disk space.
• Building from source: Node.js 24 and npm are required.
• Offline use: local dialogue and saved materials can work without a network; fresh market collection cannot.

Verification scope and licensing

The README reports 363 automated tests in 28 files plus the five listed test/build commands, and separately checked a local Qwen3 14B orchestrator and seven desktop scenarios development checks. The result check for the built Windows application is recorded separately. macОS and Linux packaging workflows exist, but the README states this does not mean the macOS package is notarized or that all platforms were manually verified.

The code is AGPL-3.0-or-later; Qwen and third-party dependencies keep their own licenses. Published installers include SHA-256 hashes in the README.