This repository contains official public beta downloads, release notes, and product documentation. The OverShoulder source code is maintained privately.
OverShoulder gives CEOs, founders, and project leaders an AI advisor that understands their company, sees work as it happens, and helps them make better decisions across strategy, product, finance, people, operations, and execution.
Unlike a generic AI assistant, OverShoulder does not start every conversation from zero. Its intelligence core, Cortex, builds an evidence-grounded understanding from company documents, organizational memory, prior decisions, and live work context.
Download links and release updates will be shared here on launch day.
Important company knowledge is usually fragmented across contracts, reports, presentations, conversations, and the people who remember why a decision was made. Conventional AI tools can answer a prompt, but they rarely understand the organization behind it.
OverShoulder brings that context together so leaders can ask better questions, verify what is true, identify what is missing, and act with greater confidence.
With user-permissioned screen sharing, OverShoulder can understand the work currently in front of you. It can follow a live workflow, recognize when context changes, and provide guidance without requiring you to explain every visible detail.
Cortex connects company files, organizational memory, historical decisions, and current context into a persistent evidence layer. It retrieves across document formats, follows related evidence, preserves document history, and rereads original pages when extracted text is not enough.
OverShoulder turns that context into grounded advice through three complementary experiences:
- Company Brain Chat for questions about the business, its history, and its evidence.
- Council for high-stakes decisions that benefit from multiple AI perspectives.
- Active Guide for step-by-step help while work is happening on screen.
- Ask company-wide questions and trace answers back to local evidence.
- Compare agreements, reports, financial records, and historical versions.
- Surface risks, contradictions, missing information, and unresolved decisions.
- Bring multiple AI models together for structured decision review.
- Receive contextual guidance while navigating unfamiliar tools and workflows.
- Preserve organizational knowledge so important context does not disappear with time or turnover.
flowchart LR
DOCS[Company Documents] --> CORTEX[Cortex Company Brain]
MEMORY[Organizational Memory] --> CORTEX
DECISIONS[Decisions and History] --> CORTEX
SCREEN[Permissioned Screen Context] --> PERCEPTION[Visual Perception]
QUESTION[Leader's Question] --> ORCHESTRATOR[Context Orchestrator]
CORTEX --> ORCHESTRATOR
PERCEPTION --> ORCHESTRATOR
ORCHESTRATOR --> REASONING[AI Reasoning]
REASONING --> CHAT[Grounded Advice]
REASONING --> COUNCIL[Council]
REASONING --> GUIDE[Active Guide]
Every interaction starts with the user's intent. OverShoulder then combines the most relevant company evidence, conversation scope, and optional live screen context before asking an AI model to reason. The final response is checked against the evidence actually retrieved.
OverShoulder supports current model families from Anthropic, OpenAI, and Google, including Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5; the GPT-5.6 family; and the Gemini 3 family. Model selection is isolated per chat tab, and provider-specific fallback chains keep a session usable when a selected endpoint is temporarily unavailable.
Council uses several models as independent reviewers instead of asking one model to imitate a debate. Its default panel combines Claude Opus 5 as moderator with GPT-5.6 Sol and Gemini 3.1 Pro, while allowing the user to change the participants.
Cortex is more than document search. It is the system that keeps OverShoulder's advice connected to the company it serves.
- Local-first evidence vault: company files remain under the user's control and are indexed for retrieval.
- Multi-format understanding: PDFs, office documents, tables, OCR content, and visual pages can contribute evidence.
- Context continuity: follow-up questions retain the relevant document scope without silently inheriting unrelated context.
- Deep Read: original pages can be revisited when structure, layout, or visual evidence matters.
- Version awareness: newer records can be preferred without erasing historical evidence.
- Grounding controls: unsupported citations and unverified claims are blocked or returned as honestly incomplete answers.
Company decisions require a higher standard than fluent text. OverShoulder is designed around four principles:
- Evidence before confidence. Advice should be traceable to the material that supports it.
- Permissioned perception. Screen context is used only when the user chooses to share it.
- Memory is not proof. Organizational memory can guide retrieval, but source documents remain the grounding authority.
- Honest uncertainty. When evidence is incomplete or conflicting, OverShoulder should say so instead of filling the gap with invention.
OverShoulder will be distributed as a packaged desktop application. Beta users will not need to install developer tooling or run the source code.
The first build will be published after code signing, notarization where required, and release verification are complete. Supported systems, known limitations, release notes, update instructions, and a feedback channel will accompany the download.
For Cortex internals, query flows, screen-sharing orchestration, retrieval, Deep Read, grounding, and evaluation architecture, see the Company Brain Technical Deep Dive.