An accounting agent that can run on your own hardware. A local-first operating layer for accounting and finance teams: an AI agent that drafts, reconciles and explains its work, running on a hosted model or entirely on machines you control.
Snapshot
In designWhat we are designing
01
One interface to hosted models or to local runtimes such as Ollama, llama.cpp, LM Studio and vLLM, which all expose OpenAI-compatible APIs. Changing models is configuration, not a rewrite.
02
Documents, prompts and client data processed on hardware the firm controls, with open-weight models. In local mode, no third-party model provider receives client data.
03
The agent drafts entries, reconciliations and client questions. Posting, filing and sending always need a person, and every step lands in an audit trail.
04
Each suggestion links back to the source document and transaction behind it, so a reviewer checks the work instead of trusting it.
05
Local models are benchmarked on our own accounting test sets before they are offered. We will publish accuracy by task, including where a model falls short.
06
Designed to build on Cayde Ledger's approval-gated agent, so month-end work, documents and client questions share one workspace.
Planned architecture
Progress
1 of 5 complete as of September 15, 2026. Design stage; no release date announced. Work in progress is shown but never counted.
Approval-gated agent pattern
Status: Sep 15, 2026
Proven in production in Cayde Ledger's AI Bookkeeper: proposals, an activity log and spending caps.
Discovery and requirements
Status: In progress
Mapping month-end close, reconciliation and client-communication workflows from day-to-day practice.
Architecture and threat model
Status: Next
Local-first data model, model gateway, audit trail and a privacy impact assessment template for adopting firms.
Local model evaluation
Status: Next
Benchmark open-weight models on classification and reconciliation tasks against our own labelled test sets.
Private pilot
Status: Next
A small pilot with Canadian firms before any general release.
Context
Accounting work runs on confidential client information. Sending it to a hosted AI model means sending it to another company's servers, often outside Canada.
PIPEDA governs how private-sector organizations handle personal information across much of Canada, alongside provincial laws in Quebec, Alberta and British Columbia. Quebec's private-sector privacy law, as amended by Law 25, requires a privacy impact assessment before personal information is communicated outside Quebec.
Federal Bill C-36, tabled on June 15, 2026 to replace PIPEDA's private-sector rules, would require a privacy impact assessment before personal information is transferred outside Canada. It is still before Parliament and may change.
Running the model on hardware the firm controls keeps the AI step out of that question entirely. This is general information, not legal advice.
Elune Finance OS is in design. The capabilities on this page describe what we are building, not a product you can buy today.
Ollama, llama.cpp, LM Studio and vLLM are independent open-source or third-party projects, named here for compatibility only.
More from EluneDev
Bookkeeping workspace
Live · Invite onlyThe calm ledger for Canadian books.
A focused workspace for Canadian bookkeepers. Clients, fiscal years and accounting periods live in one place, and an AI bookkeeper drafts the work while a person approves it.
Perpetual inventory and supplier pricing
In designLive inventory, from raw materials to finished goods.
A perpetual inventory system for organizations that buy raw materials and make finished goods. It keeps quantities and costs current with every movement and brings in prices from your suppliers and brokers, so you know what is running low and what it will cost to reorder.
Fantasy basketball toolkit
Private alphaWin category leagues with the math shown.
A toolkit for Yahoo head-to-head, nine-category fantasy basketball: rankings, projections, a draft-night board and trade and waiver analysis, all built on our own calculations.