cd /news/ai-agents/show-hn-ledgerline-local-first-finan… · home topics ai-agents article
[ARTICLE · art-29725] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Show HN: Ledgerline – local-first finance MCP server

Ledgerline, a local-first finance MCP server, gives AI agents read-only access to financial data via a single SQLite file without cloud storage or exposing credentials. The tool supports bank sync through SimpleFIN Bridge and CSV/OFX/QFX file imports, running entirely locally with no API keys required for core functionality.

read8 min views4 publishedJun 16, 2026

Give AI agents read-only access to your finances without giving anyone your data: one SQLite file on your machine, no cloud, exact integer-cent answers over MCP.

Everything runs locally. Bank access (via SimpleFIN Bridge) is read-only by construction — Ledgerline never sees your banking credentials and cannot move money. Account numbers are dropped at parse time, so the model can never see what the database never contains. Delete the one .db

file and every trace is gone.

No clone, no signup, no API key, no real financial data — the demo seeds six months of clearly fabricated transactions so you can evaluate everything before connecting anything. uv is the only prerequisite.

uvx --from ledgerline ledgerline demo
uvx --from ledgerline ledgerline summary    # income/outflow by category
uvx --from ledgerline ledgerline upcoming   # expected charges, next 30 days

demo

prints copy-paste one-liners that connect the MCP server to Codex or Claude Code; then ask things like "What recurring charges are coming up?" or "Why was last month so expensive?". When you're done evaluating, delete data/ledgerline.db

and start fresh with real data below. (demo

refuses to write into a database that already has transactions.)

git clone https://github.com/jeraldhu-yuan/ledgerline
cd ledgerline
uv sync

Then get transactions in. Both paths work, and they can be mixed freely — the importer deduplicates.

Bank sync. Sign up at https://bridge.simplefin.org (SimpleFIN Bridge, a small paid service that turns your bank logins into read-only transaction feeds — Ledgerline never sees your banking credentials), link your bank(s), and create a new app on your account page to get a one-time setup token. Then:

uv run ledgerline connect    # paste the setup token when prompted
uv run ledgerline sync       # pull your transactions

connect

stores the resulting access URL owner-only in ~/.config/ledgerline/simplefin.env

. The first sync

prompts to map each bank account to a local label; re-running is always safe, and a stale database catches up in provider-friendly 45-day windows. If an institution is missing from SimpleFIN's catalog, that account just stays on file import — mixing both paths is a supported steady state.

File import. Download a CSV/OFX/QFX export from your bank's website:

uv run ledgerline ingest export.csv --account "Checking"

The database lives at data/ledgerline.db

(gitignored); override with --db

or LEDGERLINE_DB

. No API key is needed for any of this — the two optional embedded LLM commands (categorize

, ask

) read ANTHROPIC_API_KEY

from the environment, and everything else runs keyless.

Ledgerline runs as a local stdio MCP server exposing read-only tools: data freshness, transaction search, spending summaries, period comparisons, account balances, upcoming payments, and constrained SQL — plus a small set of local-metadata writers so the everyday upkeep works from chat: "I just signed up for a $850/month course, track it", "I cancelled Netflix", "that charge is business travel, not dining", "mark this card as business". Your agent does the categorizing too, so no Anthropic API key is needed on the MCP path. Nothing can touch the bank. The contract is deliberately small and uniform — exact integer cents, totals always per currency and never combined, and limitations (staleness, uncategorized spend, unknown account purpose) reported as data rather than prescriptive workflow text. The reasoning is the client model's job; the server's job is exact, truthful primitives.

The one cache-writing tool, refresh_data

, pulls from SimpleFIN at most once an hour. A refresh that hits provider errors is recorded as an attempt but not a success, and data_status

discloses the difference.

codex mcp add ledgerline --env LEDGERLINE_DB=/absolute/path/to/ledgerline.db -- \
  uvx --from ledgerline ledgerline-mcp

claude mcp add ledgerline --scope user --transport stdio \
  --env LEDGERLINE_DB=/absolute/path/to/ledgerline.db -- \
  uvx --from ledgerline ledgerline-mcp

(From a repo checkout, point the command at /path/to/ledgerline/.venv/bin/ledgerline-mcp

instead of uvx

.) Restart the client, then ask things like "How much did I spend on dining in January?" or "What recurring charges are coming up?"

If an AI agent is doing the setup, every step is non-interactive except the ones that belong to the human:

Human only: signing up at SimpleFIN Bridge, linking banks, and creating the setup token all happen on SimpleFIN's website with real banking credentials. Never give an agent your bank login; the tool never sees it either.- The human hands the agent only the one-time setup token. Then:

ledgerline connect --token "$SETUP_TOKEN"   # claim it, no prompt
ledgerline sync --accept-default-labels     # map new accounts automatically
  • The setup token is single-use: it is consumed by connect

and worthless afterward, so it transiting an agent's context is bounded risk. The resulting access URL is stored owner-only on disk and never enters the model. - The demo path ( ledgerline demo

) and file import (ledgerline ingest

) have no prompts at all, anddemo

prints the exactcodex mcp add

/claude mcp add

commands to wire up the MCP server. When registering, pick a server name that doesn't collide with one the user already has.

uv run ledgerline summary --month 2026-06

uv run ledgerline categorize

uv run ledgerline review

uv run ledgerline recurring detect
uv run ledgerline recurring add --label "Course tuition installment" \
    --amount 850.00 --cadence monthly --day 21
uv run ledgerline upcoming --days 30

uv run ledgerline ask "why was June so expensive?"

uv run ledgerline export --month 2026-06 --out june.csv

uv run ledgerline accounts set-context "Chequing" --purpose mixed \
  --entity "Northwind Consulting" --business-use-percent 70 \
  --context "Business income plus personal spending"

Account context (personal

/business

/mixed

/unknown

, owning entity, business-use percentage, free-form note) persists in SQLite and rides along on every MCP result, so agents segment cash flow before judging it.

If your bank's CSV doesn't auto-detect, the fix is a ~10-line pull request: add one dict to PROFILES

in ledgerline/ingest/profiles.py. OFX/QFX needs no profile.

"us_checking": {
    "columns": {"date": "Posting Date", "amount": "Amount", "description": "Description"},
    "date_format": "%m/%d/%Y",
    "sign": 1,            # -1 if the export shows charges as positive
    "skip_rows": 0,
    "external_id_column": None,  # column with a bank-side unique id, if any
},

Include a small fabricated CSV fixture (invented merchants, never real account data) in tests/fixtures/

and a test asserting it ingests with the right sign convention — see test_sign_convention_profile

in tests/test_ingest.py for the pattern.

Re-importing a file, overlapping export ranges, and sync + file import of the same period all produce zero duplicates (tested in tests/test_ingest.py

and tests/test_sync.py

).

Design note — one deliberate deviation from the spec: the spec folds FITID into dedupe_hash

when present. Done literally, that would create duplicates in mixed mode: a CSV row (no FITID) and a SimpleFIN row (with id) for the same transaction would hash differently. Instead:

dedupe_hash = sha256(account_id | posted_date | amount_cents | merchant_raw | occurrence_index)

with occurrence counting — the Nth identical row in a batch is a duplicate only if the DB already holds more than N such rows. Two genuinely distinct same-day, same-amount, same-merchant transactions survive because they arrive in the same export with occurrence indexes 0 and 1.- Bank-side ids (OFX FITID, SimpleFIN txn id) are stored in external_id

with a unique per-account index, short-circuit re-imports, and are backfilled onto rows that originally arrived without one.

This satisfies every acceptance test, including both orders of mixed-mode. Caveat: cross-source dedupe matches on the raw description, so it works when both sources export the same description string (typical for OFX/SimpleFIN from the same institution).

data/

,*.db

,*.csv

,*.ofx

,*.qfx

,.env

,analysis/

, and*.ipynb

gitignored from the first commit; test fixtures anddemo

data are fabricated only.- Account numbers are never parsed: the OFX reader and SimpleFIN connector drop ACCTID

/BANKID

-class fields at parse time. Only short labels ("US Checking") identify accounts. Asserted intests/test_security.py

. - The model gets full transaction detail through run_sql

— by design. What it can never see is what the DB never contains: account numbers, credentials, raw export files. run_sql

: read-only connection (mode=ro

URI), single-statement SELECT/WITH only, keyword denylist, SQLite authorizer denying everything but reads, 200-row cap, 5-second time limit, statement/result size limits. Literals and comments are stripped before the keyword scan (a merchant named "UPDATE" is not a false positive); the authorizer and read-only mode are the real guards. Tested with hostile inputs.- SimpleFIN access URL from SIMPLEFIN_ACCESS_URL

or a0600

config file only — never the repo, the DB, or the LLM context.https

is required, HTTP redirects are refused (credentials are never replayed to another host), and loose file permissions produce a warning. - New database files are created owner-only ( 0600

). ANTHROPIC_API_KEY

from env only; LLM steps fail loudly without it, everything else runs keyless.

uv run pytest

The suite covers the acceptance checklist: mixed-mode dedupe in both orders, quarantine of malformed rows, integer-cents math, per-currency reporting, run_sql

hardening against hostile inputs, recurring detection with gap tolerance, the MCP tools, the demo seeder, and the security invariants above.

MIT — see LICENSE.

── more in #ai-agents 4 stories · sorted by recency
── more on @ledgerline 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-ledgerline-l…] indexed:0 read:8min 2026-06-16 ·