Personal companion · source available
Make the workspace yours.
Choose the welcome mix, appearance mode and palette while keeping personalization out of model memory.
More than a model. A private system for real work.
Rangabot coordinates local models, approved knowledge, selective memory and bounded tools—then shows enough of the process for you to remain in control.
Real product · synthetic data
Four short tours. No staged lifestyle photography.
These loops were captured from a clean local source candidate using an isolated temporary data root. The model endpoint was disabled; no personal profile, installed model, chat, document or local path was used.
Scholar · source available
See knowledge and its boundary.
Move from current developments to vault health and the local changelog without exposing private documents.
Local tools · mixed maturity
Open only the tool the work needs.
Memory and approved folders are source-available; bounded SQL analysis remains experimental and visibly permissioned.
Open platform · source available
Inspect progress, not promises.
Open the public capability map, then inspect the criteria, dependencies and evidence behind one score.
Honest availability
One vocabulary for what works—and what does not yet.
Present in the public source tree; inspect the exact commit before installing.
Usable for bounded work, with an important quality limitation still open.
Implemented for a controlled preview, but not a supported public release.
A declared direction, not current functionality.
One calm workspace
Conversation first. Tools appear only when the work needs them.
Start with a question. Smart mode can retrieve relevant local knowledge; Teacher mode tightens the evidence boundary; the workbench opens memory, analysis, mastery and approved folders without turning the composer into a control panel.
Your requestRangabot control planeApproved local contextYour local model
One calm workspaceAsk first. Add tools when needed.
Explain this result and show the smallest useful check.
Current chatLocal memoryKnowledgeData
Model, source boundary and any calculation receipt remain visible.
Capability map
Nine paths through the current product.
Each path names the useful job, the control that keeps it trustworthy, and the boundary that should not be hidden by polished marketing.
A local conversation with a control plane above the model.
Rangabot frames the current request, recent chat, approved context, tool boundaries and honest limits before the selected local model answers.
- Streaming Markdown and code
- Reply, Stop and durable turn recovery
- Local only, Smart and Teacher modes
A welcome that feels personal without becoming hidden memory.
Choose a local name, light or dark appearance, one of eight palettes, and a fresh-chat mix of quotes, jokes, thoughts or cited lines from your own books.
- Optional name and rotating greeting
- Mix, Quotes, Jokes, Thoughts or My books
- Preferences stay separate from model memory
Conversations stay findable, portable and under your control.
Search, pin, reopen and delete local chats; start project-scoped work; and move readable transcripts between installations with bounded Markdown import and export.
- Local conversation history
- Projects, search and pinning
- Markdown import and export
Remember only what the user explicitly approves.
Local memories are visible records—not an invisible profile. Create, inspect, edit, delete, export and review conflicts before importing anything.
- Preference, fact or instruction
- Relevance-filtered use
- Conflict-aware JSON import
A private library that shows where an answer came from.
The Knowledge Vault ingests supported local documents, combines keyword and embedding retrieval, and keeps source passages attached to teaching and synthesis.
- PDF, DOCX, HTML, Markdown and text
- Hybrid or visible keyword-only retrieval
- Teacher Mode citations and evidence limits
Questions can become inspectable local calculations.
Approve a CSV, Parquet or DuckDB file, attach it to one conversation, and let Rangabot link ordinary language to its schema before proposing a bounded read-only calculation.
- Relationship-aware schema linking
- Typed descriptions, aliases and confirmed joins
- Query, row count, timing and fingerprints
Use only the piece of local work you intentionally attach.
Allow a repository folder, search eligible text on demand, attach a bounded line-numbered preview, or create a validated Word document through normal chat.
- Reversible repository approval
- Secret-filtered scoped previews
- DOCX validation, preview and download
See, choose and deliberately download local models.
The desktop candidate distinguishes reviewed chat models, embedding models and unqualified installations, with hardware and license guidance before a large download.
- Installed-model inventory
- Explicit general-model selection
- Click-to-install reviewed choices
Trust is visible in the interface and the roadmap.
Privacy boundaries are explicit, and Path to Mastery separates merged work from capabilities that have actually earned their gates.
- Loopback and owner-private storage boundaries
- Criterion-level public mastery map
- Named evidence and honest release states
What comes next
Progress without pretending it has shipped.
The open-source application and the packaged desktop experience move at different speeds. These labels keep that boundary clear.
Rangabot desktop editions
Full and smaller Light packaging paths are implemented for macOS arm64 and Windows x64. Light reuses an already-running local Ollama service; signing and clean-machine evidence still block public binaries.
Read the candidate update →More artifact skills
Editing existing Word files, PDF reports, presentations and spreadsheets follow the available Word-creation foundation. Email drafting is local text work today; sending email is not implemented.
Inspect the acceptance criteria →Windows, Linux and mobile apps
A governed Windows x64 MSIX path is implemented but remains unsigned and untested on clean Windows machines. Packaged Linux and mobile applications remain roadmap work.
See how release claims are earned →Local response feedback
Helpful and Needs improvement can be stored reversibly on one exact eligible candidate. Ordinary source builds keep the control disabled when their candidate manifest is unknown or mixed.
Read the candidate boundary →