obsify

obsify

[CI] [PyPI] [Python] [License: MIT]

Let an AI assistant work on sensitive files without their raw values ever entering the model's context.

obsify is a local, deterministic MCP server. The frontier model reasons over shape — schemas, synthetic twins, masked feedback — while deterministic local code touches the substance and returns only masked, aggregated results. No LLM calls, no network at runtime: detection is regex + checksums + dictionaries + Presidio's local NER.

It ships with Australian entity support (ABN / ACN / TFN, checksum-validated), credential/secret detection (cloud keys, API tokens, private keys, DB connection strings), and a label-driven routing layer that makes "when should the assistant avoid raw data" a deterministic, enforced decision rather than a judgement call.

Honest scope: run_on_real executes model-written code in a best-effort local sandbox and masks its output best-effort. It is not a jail. Read SECURITY.md before pointing it at anything you cannot afford to leak. Return aggregates.

Why

Feeding confidential documents to a hosted LLM means the substance leaves your perimeter. The usual answers are "don't use the LLM" or "trust the provider." obsify takes a third path — compute-to-data: bring the code to the data, not the data to the model.

• The model sees the schema of a spreadsheet, not its rows.
• The model develops against a synthetic twin (faked values, real structure).
• The model's analysis code runs locally; only masked, aggregated output returns.

The frontier model's reasoning is preserved. Only its eyes on raw values are removed.

Tools

Supported documents: PDF (text + tables; complex-table fallback via obsify[tables]), Excel .xlsx/.xlsm, and Word .docx (paragraphs + tables). Unreadable or unsupported files are surfaced as explicit notes/blind spots, never silently dropped. (No OCR yet — scanned/image pages are flagged as low-coverage, not transcribed.)

Known-entity masking (optional). Supply a local .obsify.entities list of names to hide; scan_pii / redact_text deterministically catch them — and the suffix/abbreviation variants NER misses (BRIGHTWATER HLDGS P/L for Brightwater Holdings Pty Ltd) — as KNOWN_ENTITY. The list stays local and never enters the model's context. See docs/known_entities.md.

Demo

Poke at all five tools live against synthetic data with the official MCP Inspector:

python -m obsify.make_corpus --out ./corpus_demonpx @modelcontextprotocol/inspector obsify-mcp

Call scan_pii on ./corpus_demo/ledger.xlsx and confirm it returns types / counts / locations only — never values. See docs/verifying.md.

Try it — synthetic corpus

Generate a fake-but-realistic corpus (all synthetic; ABN/ACN/TFN are checksum-valid) spanning all three formats, then point a tool at it:

pip install "obsify[demo]"                 # reportlab, for the sample PDFspython -m obsify.make_corpus --out ./corpus_demo

It writes a multi-sheet Excel ledger (a numeric false-positive minefield), a PDF engagement letter (prose + trial-balance table), and a DOCX audit memo (paragraphs + vendor table). Great for kicking the tyres on scan_pii / make_synthetic_twin without touching real data.

Install & run as an MCP server

Requires Python 3.11+. obsify speaks MCP over stdio — the client launches it as a local subprocess; nothing is hosted remotely. Register it with any MCP-capable client (Claude Desktop, Claude Code, Cursor, VS Code, …) by adding one block to that client's config.

Recommended — zero-install via uvx:

{ "mcpServers": { "obsify": { "command": "uvx", "args": ["--from", "obsify", "obsify-mcp"] } } }

uvx fetches obsify from PyPI and runs it on demand — no permanent install. On first run, obsify downloads the spaCy NER model (en_core_web_lg, ~560 MB) once and caches it; this fetches a public model and sends no user data (set OBSIFY_AUTO_DOWNLOAD=0 to forbid it and install the model yourself). Later runs are instant and fully offline.

Or install it (pip / pipx):

pipx install obsify        # isolated, on PATH  (or: pip install obsify)

Then point the client at the installed command:

{ "mcpServers": { "obsify": { "command": "obsify-mcp" } } }

Restart the client and the tools appear. Optional extras: obsify[tables] (complex-table PDF fallback via camelot + Ghostscript), obsify[compute] (pandas, handy inside run_on_real code).

PATH gotcha (the #1 cause of "server won't connect"): the command must resolve on the PATH the client sees. A GUI client may not share your venv's PATH. Fixes: use uvx/pipx (globally resolvable), or give an absolute path — "/path/to/.venv/bin/obsify-mcp" (macOS/Linux) or "C:\\path\\to\\.venv\\Scripts\\obsify-mcp.exe" (Windows).

From this repo (before it's on PyPI):

pip install "git+https://github.com/Formative-Sum41/obsify.git"   # gets `obsify-mcp` + `obsify`

The routing layer — deterministic, not a judgement call

The hard part of "help me, but don't read the confidential file" is deciding when to protect. obsify moves that decision out of the model and into the environment:

1. .obsify.json — a label manifest classifying paths (public / confidential / restricted).
2. obsify.guard (run as python -m obsify.guard) — a PreToolUse guard that blocks a direct read of a labelled file (exit 2) and redirects the assistant to scan_pii / make_synthetic_twin / run_on_real.
3. A convention (in CLAUDE.md) so the assistant prefers obsify before it even hits the guard.

Set it up with one command:

obsify init [--dir PATH] [--with-claude-md]

obsify init is non-destructive by design — it owns exactly one file and hands you snippets for the rest:

• .obsify.json — obsify owns this; init writes it (never overwritten without --force).
• .claude/settings.json — your file: init prints the PreToolUse hook block to paste, never edits it (it runs code, so registering it is your call).
• CLAUDE.md — your file: the convention is opt-in. Default prints it; --with-claude-md appends a marker-wrapped, idempotent block that never clobbers your content.

Full convention: docs/obsify_routing.md.

How detection stays precise

• Checksum-validated identifiers. ABN/ACN/TFN candidates are proposed by regex and confirmed by their official checksums, so a random number is never reported as an identifier.
• Context-required IDs. A bare number is only accepted as an ABN/ACN/TFN when a label word ("TFN", "ABN", "BSB", …) is nearby — this kills the sequential-journal-ID false-positive flood on numeric ledgers.
• Letterless / NER-with-digit suppression. Pure numbers, amounts, dates and alnum codes are not flagged as names/orgs; real names, emails and addresses (which carry letters) are unaffected. Validated letterless PII stays exempt: checksum IDs (ABN/ACN/TFN/Medicare), Luhn cards, valid IPs, BSB-adjacent accounts, and phones (via context or phone shape) — while a decimal point still marks an amount, not a phone.
• Credentials, not just PII. Cloud keys (AWS/GitHub/Google/Slack/Stripe), JWTs, private-key blocks and DB connection strings are flagged as CREDENTIAL by anchored patterns — vendor prefixes (AKIA…, ghp_…) or a keyword-gated secret = <value>, never entropy heuristics (which would flood on hex/base64 ledger columns). The whole BEGIN…END private-key block is masked, not just its header, so no key body is left behind.

Measured accuracy

obsify ships a scored evaluation harness (eval/ — labelled synthetic corpus + answer key + scorer against the shipping detector, plus an independent third-party cross-check). Headline on the synthetic corpus: 100% recall on expected-detect items, 0 false positives on a numeric FP-torture sheet (with a grouped-number guard), bare context-gated IDs correctly suppressed. Independent cross-check vs Microsoft presidio-research: EMAIL/IBAN 100%, PERSON 94%.

The harness earned its keep — it found real defects, which were then fixed: credit cards and phone numbers were being silently suppressed by the numeric-noise filter (now exempt via checksum validation / phone shape), and Medicare, IP, date-of-birth, AU passport and driver-licence had no recognizer (now added, checksum- or context-gated). Full method, numbers, and remaining documented gaps (SWIFT/BIC, non-DOB dates): eval/README.md.

Tests

pip install -e ".[dev]"pytest tests/            # or run any file directly: python tests/test_obsify.py

Thirteen suites (88 tests), run in CI on Linux + Windows / Python 3.11 + 3.12:

• mcp-protocol — launches the real server over stdio and speaks MCP to it (the same path a client like Claude uses): confirms all five tools register with valid schemas and that calls round-trip through JSON-RPC — including scan_pii returning shape only, end to end.
• checksums — anchored to externally-published ABN/ACN/TFN worked examples (valid and corrupted), which breaks the generator↔validator circularity.
• obsify / twin / redaction — the privacy invariants: shape-only output, leak-free twins, and a fail-closed self-check.
• precision — the false-positive suppressors kill numeric-ledger noise while keeping real names.
• credentials — the anchored secret patterns catch cloud keys / tokens / JWTs / private-key blocks / connection strings, while keyword-anchored generics stay precise on prose (no entropy).
• routing — the guard's block/allow classification and obsify init's non-destructive contract.
• corpus — the synthetic PDF+Excel+DOCX corpus end to end: per-format detection, DOCX paragraph+table extraction, and shape-only output across every format.
• evaluation — the scored harness as a regression gate (recall, suppression, FP-torture, gaps).
• robustness — graceful degradation: corrupt/oversized/empty/nested/unsupported inputs never crash and are always surfaced as notes.
• model / variants — first-run model auto-download logic; variant normalization behind verify_value_free.

For interactive verification (MCP Inspector) and the live-client last-mile check, see docs/verifying.md.

Related work

obsify is one of several MCP servers tackling "let an AI touch sensitive data safely" — they're mostly complementary, solving the same problem from different ends. Worth knowing where each fits:

Where obsify is distinct: it's the only one of these where the model gets neither raw values nor a full mirror to operate on — just shape + masked aggregates — combined with checksum-validated identifiers, credential detection, a deterministic routing guard, and a hard no-network / no-LLM guarantee. That's the strictest-isolation end of the spectrum, tuned for confidential financial documents.

Honest trade: obsify optimizes isolation of the values over utility on the data. If you need referentially-intact analytics on a clean copy (cloakbox), reversible round-tripping (redact-mcp), or a multi-language hosted service (cms-ai), those are the better fit — and pair well with obsify rather than competing with it.

Contributing

PRs welcome — see CONTRIBUTING.md for setup, the merge bar, and the non-negotiable invariants (no LLM calls in the library, no runtime network, no real data, shape-not-substance). Security issues: SECURITY.md (report privately).

License

MIT — see LICENSE.