Conversation candidate
Latest complete v1.0.11 candidate; repeat and blind-review release gates remain open.
Private AI · faithfully localYour knowledge.
Your knowledge.
Your computer.
Your Ranga.
Rangabot is a local-first personal assistant for conversation, learning, coding and analysis. Your chats, memories, documents and data stay with you.
Open sourceRuns locallyNo accountCloud disabled by default
LOCALNothing sent
Useful because it knows
when to think—and when to ask.
Rangabot adds a local control plane above the downloaded model: selecting context, respecting permissions, running verified tools and showing the evidence behind important answers.
A mind that listens before it remembers.
Rangabot follows the current request first, carries useful conversation context, and uses only relevant memories you have approved.
Explore the path →02AvailableScholarYour books become a private classroom.
The Knowledge Vault retrieves from local documents, shows its evidence, and separates sourced claims from model background.
Explore the path →03ExperimentalAnalystQuestions become verified calculations.
With explicit access, Rangabot plans and runs bounded read-only analysis, then exposes the query behind its explanation.
Explore the path →04TrainingBuilderLocal context for useful creation.
Approve a repository, attach scoped code, and turn the conversation into explanations, plans, or validated Word documents.
Explore the path →Privacy is the architectureLocal first is a boundary,
Local first is a boundary,
not a badge.
Rangabot binds locally, rejects remote Ollama configuration, and keeps private data out of the public website. External help remains a future, permission-gated path.
Read the privacy model →You
RangabotLocal control planeLocal modelKnowledge vaultApproved data
External service Locked by default
Local intelligence
Private books become clear, cited teaching.
See what Rangabot retrieved, where it came from and which parts rely on model background.
Open the product tour →
Proof before promises
Every claim has a trail.
Scores name the suite, model and limitations. A merged feature is not automatically a mastered capability.
Critical trust cases
One complete candidate run. Intermittent failures still count during repeated release review.
Memory precision & recall
Synthetic memory-selection audit; no real conversations or saved memories were inspected.
Analytical transfer
Frozen astronomy holdout. Useful progress, still below the 90% release target.
Built by the pack
Help a local companion grow.
Choose a bounded capability, discuss the acceptance criteria, merge evidence and keep the achievement attached to the people who earned it.