AI-based finance management where the intelligence comes to your data — your data never goes to the intelligence.
Ledgerly is a private, self-hosted finance dashboard that runs in your own Cloudflare account: your transactions in your D1 database, your receipts in your R2 bucket, no vendor in between. The vision is to grow it into the privacy-first alternative for AI-assisted money management — first for individuals, then for freelancers and small businesses.
In mid-2026 the "connect every account and let AI analyze it" lane was claimed by the biggest player possible: OpenAI shipped bank-linked personal-finance features inside ChatGPT (Plaid, 12,000+ institutions, an Intuit partnership). Competing with that on connectivity would be pointless. But that product has a structural ceiling it can never cross: your complete financial life lives on someone else's servers and feeds someone else's models.
Meanwhile, the open-source finance world (Actual Budget, Firefly III, Ghostfolio, Maybe/Sure) proved lasting demand for self-hosted money tools — tens of thousands of GitHub stars, real communities — yet none of them ship native AI, and their communities' loudest complaints are import friction and missing intelligence.
That gap is the vision: the trust model of self-hosted open source, with the intelligence of a modern AI finance app.
A complete, tested v1 — deterministic core plus the first rung of the AI ladder (receipt extraction, shipped August 2026):
- Nine pages: dashboard, transactions, recurring, subscriptions, budgets, goals, documents, rules, settings — with a first-run onboarding wizard and 30-currency display support.
- Deterministic engines with real test suites: duplicate-fingerprint imports, an auditable recurring/subscription detection algorithm, a rules engine, strict never-guess CSV parsing (ambiguity always asks the user).
- Documents vault: original receipt/statement bytes stored in R2.
- AI receipt extraction (off by default): Workers AI in the user's own Cloudflare account, or bring-your-own-key Anthropic. Suggestion-only — per-field confidence, a review form, and no transaction until the user confirms.
- A full data-erase flow, empty-start guarantee (no sample data, ever), and a
/api/drive-synccontract ready for scheduled import runners.
- Your data stays in your infrastructure. Default deployment is the user's own Cloudflare account. Nothing phones home.
- Deterministic core, AI assist. Money math — totals, dedupe, recurring detection, budgets — stays deterministic and auditable. AI suggests; it never silently writes.
- Never guess. Ambiguous imports ask. Uncertain extractions go to review. Confirm-before-write everywhere.
- No sample data, no dark patterns, no invented numbers. Empty states are honest; trends render only when real data supports them.
License: AGPL-3.0. Every comparable app with a hosted arm chose it (Firefly III, Ghostfolio, Maybe); the one permissive-licensed peer (Actual, MIT) watched third parties capture its hosting revenue. AGPL keeps a future first-party cloud viable while staying genuinely OSI-approved open source — which fair-source licenses are not, at real cost to directory listings and launch trust.
Launch checklist: LICENSE + CONTRIBUTING + code of conduct, one-command deploy docs ("deploy to your own Cloudflare in 10 minutes"), screenshots and a public demo instance, then a Show HN positioned as the open-source, privacy-first alternative to AI finance apps. Growth follows the proven playbook (Plausible, Cal.com): "open-source alternative to X" content, not star-chasing.
What the market data says is genuinely sticky is invisible AI judged by friction reduction — not chat gimmicks (only ~18% of consumers are comfortable letting AI make financial decisions autonomously).
- Categorization that learns from corrections. 🚧 In progress. AI fallback where deterministic rules don't match; user corrections get promoted into new rules. The existing rules engine stays authoritative.
- Receipt extraction. ✅ Shipped. The bytes were already in R2; a schema-constrained LLM reads one document into a suggested transaction (merchant, date, total, category) with per-field confidence, and the user confirms it in a review form. Two paths from one feature: Workers AI (inference inside the user's own Cloudflare account — the privacy-first default) and bring-your-own-key frontier models for maximum accuracy. Cents per document or less.
- PDF statement extraction — many transactions from one document. ✅ Shipped August 2026. Receipt extraction handles the one-merchant-one-total shape; a bank or card statement is a table of dozens of rows, so it needs its own pipeline: page-aware reading, row-level extraction, and a review screen that triages a whole batch at once rather than a single form. CSV statements already import deterministically and stay the recommended path — this closes the gap for banks that only hand out PDFs. Non-negotiables carried over: nothing is inserted without confirmation, the existing duplicate fingerprint applies per row, unreadable rows are reported and skipped rather than guessed, and per-row confidence drives what the review screen puts in front of the user first. Feasibility note: this is the first feature where per-document cost and context limits matter (a 12-page statement is not a receipt) — expect chunking by page and a hard row cap per run.
- Cash-flow forecasting on top of the recurring-detection engine — the numeric projection stays deterministic and auditable; AI narrates scenarios in plain language. ✅ Deterministic projection shipped August 2026 (30/60/90-day horizons on the Recurring page, income + expense series); AI narration still to come.
- Natural-language search / MCP. ✅ Shipped August 2026. Read-only. Expose the user's data as an MCP server so they can point their own AI client at their own finances — "chat with your money" without the data ever leaving their control. Six deterministic tools (search, summary, recurring, forecast, budgets, briefing) behind a bearer token; read-only is structural — no mutating tool exists. Setup in docs/MCP.md.
- Proactive briefings & anomaly alerts — last, once 1–4 generate the signal that makes alerts useful instead of noisy. ✅ Briefings shipped August 2026: a deterministic weekly/daily digest (last-7-days activity, next-7-days forecast, items awaiting review) delivered over WhatsApp via the user's own Meta Business Cloud API credentials. Anomaly alerts still to come.
The structural bet the deployment model makes possible: your bank's own alert emails, e-receipts, and statements become a credential-free, real-time feed — routed to an address you control and processed entirely inside your own account. No aggregator, no shared bank login, any bank that can send an email. A cloud vendor cannot copy this without reading your email on their servers; Ledgerly has no servers to read it on. Bank formats are community-contributed parser packs — the intelligence ships as open code, never as pooled data. Ingestion is suggestion-only, allow-listed, and never triggers AI, so a spoofed email can at worst propose a row the user rejects.
- Ledgerly Cloud: first-party hosting for people who don't want to self-host. Same codebase, no feature gap for individuals — the paid product sells convenience, not ransomed features.
- Freelancer / micro-business tier (Cal.com-style: a separately-licensed enterprise directory, core stays AGPL): business-expense tagging, receipt- to-ledger automation, multi-user. The wedge is not "cheaper QuickBooks" — free incumbents own that — it's the same data-ownership thesis plus AI automation. The segment's 2026 lesson (Botkeeper's shutdown) is that funding-dependent platforms die and take client data workflows with them; a tool that runs in the customer's own account structurally cannot.
- Solo-maintainer burnout is the #1 documented killer of projects like this. Scope must stay ruthless; the deterministic core must stay small.
- Stars are not revenue. Maybe Finance had 54k stars, VC funding, and a beautiful product — it was archived in 2025 and survives only as a community fork. The models that worked (Plausible: ~$3.5M ARR, bootstrapped) took years of content-driven growth.
- AI trust is fragile in finance. Cleo's $17M FTC settlement shows how fast AI + money + dark patterns destroys trust. Our principles above are the moat; violating them once forfeits it.
Research basis: competitive and licensing landscape surveyed August 2026 (OSS finance apps, AI finance incumbents, open-core case studies). Detailed findings with sources live in the project's research notes.