SME cash-flow planning, from evidence to reviewed action.
Home · Open demo workspace · Take the judge tour · Read the docs · Run locally
Fluxora helps a small-business owner understand when cash becomes tight, which facts support the forecast, and what can be prepared for review. Deterministic Python tools calculate the financial results; an optional constrained model helps interpret supported requests.
Built for Topic E — SME Finance Copilot of the 2026 Shenzhen International FinTech Competition. Current application source: 0.26.4. All bundled examples are synthetic. No real bank connection or payment execution is provided.
Latest evidence: Judge walkthrough and live-model checks · submission review checklist. Existing presentation assets are historical drafts; final slides/video remain deferred. Home and workspace routing records the current entry flow.
The application review records the broader 90-page audit, deployed Telegram-practice flow and local bot-settings concurrency correction. The subsequent rendered review covers ten principal pages at 360px, 390px and desktop, with focused menu, assistant and conversation checks. This is browser emulation, not a real-device or all-operation usability certification.
Use Send for a one-submit report or DRAFT/pending preparation after the inline provider/data disclosure. Options → Preview before sending retains the full outbound and interpretation previews. Final approval stays separate; missing facts still require clarification. Inspect the selected task and exact terms in the result. For immediate answers without a provider, use the direct report links.
Explore: Why Fluxora · Product tour · Features · Quickstart · AI & Telegram · Architecture · Verification · Documentation · Contributing
An invoice can be profitable and still arrive too late for payroll. A forecast is useful only when the owner can inspect its assumptions and connect an option to the right decision.
Fluxora brings that work into one multipage workspace:
- See the position. Read recorded cash, upcoming obligations and each currency's lowest projected balance.
- Trace the facts. Follow the source references, missing dates, restrictions and unsettled observations behind the result.
- Explore the options. Preview timing changes and compare bounded plans while protecting payroll and consent requirements.
- Prepare the next step. Create a DRAFT or pending workflow, then review its exact terms through a separate owner decision.
For judges: start with the five-minute guided tour. It combines direct financial tools with a separate Telegram practice lesson. No Telegram account or API key is needed. The hosted operator can enable bounded live guided task translation. Direct reports and Telegram browser practice need no model.
Earlier local application screenshot, source 0.25.9, captured with the bundled synthetic Atlas example. Current hosted captures are in the latest walkthrough. Decorative hero artwork is illustrative; it is not a financial result.
Open the workspace screenshots — forecasts, guided tasks and Telegram practice
Recorded cash, conditional collections, floor checks and source references remain distinct. USD and EUR are not pooled.
Direct reports work without a model. Optional interpretation requires a qualified configuration and provider/data disclosure. Send prepares a result in one submission; Options → Preview before sending retains exact interpretation confirmation. This earlier screenshot predates the one-submit controls.
This is a separate disposable lesson with simulated delivery and no live AI. It uses the deterministic tools; a practice DRAFT and a later snapshot approval are separate operations.
See screenshot provenance for the capture scope and limitations.
Actual hosted Chrome capture, 20 October 2026: the model prepared exact USD 100 terms; owner approval remains separate. The tested snapshot approval saved a different forecast workflow and did not approve this DRAFT. See the bounded walkthrough results.
| Workspace | What you can do | What stays explicit |
|---|---|---|
| Overview & daily review | Inspect recorded cash, priorities, upcoming obligations and forecast risk | Snapshot date, missing facts and conditional assumptions |
| Forecast & account funding | Read daily/weekly outlooks, account availability and protected-payroll funding | Currency separation and cash-floor constraints |
| Plans & scenarios | Preview invoice delays, partial collections, affordability and bounded plan alternatives | Hypothetical changes, consent requirements and no-feasible-plan results |
| Evidence & ledger | Stage imports; inspect sources, invoices, bills, loans and attachments | Evidence supplies facts, never authorization |
| Actions & approvals | Prepare DRAFTs, inspect exact terms and review pending local work | The model cannot approve itself |
| Settlement & recovery | Record source-backed synthetic observations, reconcile, correct, export and back up | Acceptance, observed payment and reconciliation are separate |
| Agent & skills | Use direct reports or six bundled scopes for supported task preparation | Model output is untrusted; financial results come from code |
| Telegram companion | Try browser practice or configure an allowlisted private bot locally | Simulation versus real opt-in private messages |
Browse the 64 dedicated workspace pages. Page counts and supported-operation counts are different; neither means every operation is proven through unrestricted live language.
Requirements: Python 3.10+ with SQLite and venv. The application runtime uses the Python standard library. Direct reports need no provider key or paid service.
Use a short writable folder, such as C:/Fluxora, to avoid the tested Windows SQLite long-path backup limitation:
git clone https://github.com/ShahadatTest/fluxora.git C:/Fluxora
cd C:/Fluxora
python -m tools.local_setup --check
python -m tools.local_setup --start --open-browserYou can also extract the repository and double-click START_FLUXORA.cmd.
macOS / Linux commands
git clone https://github.com/ShahadatTest/fluxora.git
cd fluxora
python3 -m tools.local_setup --check
python3 -m tools.local_setup --start --open-browserInstall your operating system's matching venv package if it is missing. Local Windows setup was exercised; a successful Linux CI run is a separate automated result, not a real-device macOS/Linux usability test.
| Destination | Local URL |
|---|---|
| Home / product website | http://127.0.0.1:8765/ |
| Business workspace | http://127.0.0.1:8765/app |
| Guided introduction | http://127.0.0.1:8765/start |
| Documentation | http://127.0.0.1:8765/docs |
| API & Telegram settings | http://127.0.0.1:8765/settings/connections |
The launcher creates a no-pip runtime and binds to loopback. Data persists in .runtime/local/state.sqlite; stop with Ctrl+C and restart with the same database path to retain it. An existing database is not reset by startup. Read local setup for custom ports, existing data, backup and troubleshooting.
| Mode | Available behavior | Access |
|---|---|---|
| Direct local tools | Deterministic reports, scenarios and reviewed owner forms | No model key required |
| Public hosted demo | Separate expiring synthetic visitor workspace and browser Telegram practice | No login; optional operator-enabled guided live AI with a daily allowance |
| Private local AI | Supported request interpretation with disclosure, one-submit preparation and optional exact preview | Configure and verify a compatible model privately |
| Private Telegram companion | Allowlisted commands, opt-in messages and exact reviewed buttons | Configure your own bot and pair your workspace |
In a private local installation, open Settings & help → API & Telegram. Multiple compatible model profiles can be saved, with one active profile. Save, activation, potentially billed verification and synthetic qualification are separate actions. Restart starts verification unverified. Hosted connection settings remain a locked preview.
Environment configuration is also supported:
CASHFLOW_MODEL_API_KEY
CASHFLOW_MODEL_BASE_URL
CASHFLOW_MODEL_NAME
These are variable names, not values. Enter credentials only through private settings, process environment or the documented hidden prompt. For a local Telegram bot, use a token that is not already polling elsewhere. See model setup, Telegram setup and local credential storage.
- The model can read supported reports and prepare DRAFT/pending work. It cannot self-approve, invent receipts or move money.
- A separate owner approval applies only to the reviewed local scope. It does not establish counterparty acceptance or observed payment.
- Documents are evidence, not authorization. Reconciliation is separate from recording an observation.
- Payroll is protected. Currencies stay separate. Display-only FX creates no spendable cash.
- Deterministic financial cards are authoritative for the computed result. Advanced model prose remains explicitly unverified; even a valid structured task can misunderstand intent.
Autopilot watches recorded financial inputs, runs deterministic analysis, prepares a bounded eligible DRAFT batch and a separately approved forecast snapshot, and records follow-up/digest jobs. It makes zero model calls for scheduled checks. Optional qualified models still interpret explicitly submitted tasks.
Policies have separate currency limits, saved owner opt-in and a pause control. Private Telegram delivery needs the workspace’s current owner pairing and notification consent. Exact DRAFT approval/rejection and snapshot decisions are supported; no real sending/payment, invented consent or receipt is added. Public visitors have isolated policies and cannot message the operator’s private chat. Read setup, limits and recovery.
flowchart LR
E[Evidence and recorded facts] --> V[Local scope and input validation]
B[Browser or private Telegram] --> V
M[Optional external model] -->|Send: untrusted instruction| V
M -->|Options → Preview before sending| R[Exact interpretation review]
R -->|Read or prepare only| V
V --> F[Deterministic financial tools]
F --> D[DRAFT or pending work]
D --> O[Separate owner decision]
O --> X[Executor for approved local scope]
F --> S[(SQLite and audit history)]
X --> S
Python and SQLite preserve integer minor-unit calculations, transactional saves and durable request identities. Stale exact reviews are rejected; accepted replay does not create another effect. Provider deadlines, response bounds and sanitized failures sit at the adapter/runner boundary. Read the technical design →
The badge at the top reports the actual GitHub Actions workflow. Local measured evidence is recorded separately:
| Evidence | Recorded outcome | Scope |
|---|---|---|
| Complete source suite (0.26.4) | 2,219 passed in 429.90s on 20 October 2026 | Automated synthetic tests, local HTTP stubs and mocked providers; separate from live-model evaluation |
| Clean build & install | Source package, wheel and fresh offline installation passed | Assets, fixture examples and bundled skills present |
| Focused browser checks | Latest judge journey: 24 checks across eight surfaces at desktop, 360px and 390px; no document overflow | Chrome emulation and selected interactions; earlier ten-page review remains separate |
| Public live guided AI | Latest two tasks passed: forecast and exact USD 100 DRAFT; neither showed 429/timeout | OpenRouter Qwen; 14,016 provider-reported tokens; two samples, not general language accuracy or future availability |
| Telegram checks | Latest rendered browser practice plus fresh HTTPS replay/rejection checks passed | Simulated transport, zero model/API calls in the HTTPS lesson; earlier real API acknowledgments are historical |
| Fresh-user walkthrough | User reports completion with no problems found | Detailed participant/device/task/timing records were not supplied |
The earlier deep-path Windows test failure, its reproduction and the short-path workaround are retained in validation. The subsequent rendered review resolved the browser-access blocker. Broader live-language evaluation and a real-business security/operations review remain open. The user-reported walkthrough is useful feedback, rather than a documented usability study.
Run the tests yourself
Create a separate development environment with pip; the minimal runtime does not contain pytest.
python -m venv .venv-dev
.\.venv-dev\Scripts\python.exe -m pip install -e ".[test]"
.\.venv-dev\Scripts\python.exe -m pytest -qOn macOS/Linux use .venv-dev/bin/python. For Windows CI or deeply nested workspaces, put pytest's temporary data in a short external directory. Explicit live-evaluation utilities under tools/ require separately configured credentials and can incur charges; normal tests do not establish live provider behavior.
Repository map
cashflow_copilot/ Financial engine, SQLite, web pages, assets and bundled skills
tests/ Synthetic domain, transport, security and regression tests
tools/ Local setup and explicit evaluation utilities
examples/ Synthetic import and workflow examples
docs/ Current guides, architecture, validation and screenshots
deploy/vps/ Optional container deployment, backup and rollback
.github/ Automated tests and contribution templates
Read CONTRIBUTING.md before proposing a change. Use invented data in a bug report, preserve financial/permission boundaries and include relevant verification. The security policy explains what must stay out of public issues and screenshots.
Application source is 0.26.4; recent changes are summarized in the changelog. No distribution license has been selected automatically. Review usage/distribution rights before reuse; there is no implied open-source license grant. Generated artwork and trademark limitations are documented in asset provenance.
Demonstration boundary: no real bank feed, payment rail, lender integration or production identity system. Recorded sources can be incomplete or wrong; forecasts remain conditional. Publishing this repository does not complete competition registration or submission. Competition context · Known risks