CFO Navigator

CFO Navigator is Brian Weisberg's personal site, and the home for the tools he's built for the finance leader community. It's live at bmweis.com.

It started as a creative exploration, with tools built for his own use. Easy access to the writing he reads every week. Somewhere to keep the piece he'll want to reference six months from now, when the question finally comes up.

That grew into a toolbox. Which software to look at when the close process needs replacing, which community is worth the membership fee, which benchmark to trust. Those were questions he was working through for himself anyway. Other finance leaders turned out to be working through the same ones, so the whole thing is open now.

The other half of the exercise was seeing what's possible building this way, with AI doing real work alongside him rather than around the edges. The site is honest about where that holds up and where it doesn't.

If you're here for the tools, start at the site. This repo is the code behind it.

If you're here as a developer: one FastAPI app, one SQLite database with FTS5 full-text search as the spine, no separate backend or frontend build. Public pages, a login-gated back office, and a remote MCP server all live in the same process.

What it does

• CFO Toolbox (/tools): the working reference. A curated directory of finance software, spanning ERP and close management through billing, tax and spend, plus a second directory of finance communities. Both carry full profile pages and a side-by-side Compare view. Alongside them: the benchmarking sources worth using, and a short reading list. Vendor profiles add an AI-assisted read on how much of the work the product really automates. Anyone can submit a vendor or a community, and Brian reviews and approves them. More sections get added as they earn their place.
• FP&A Buddy (/tools/fpa-buddy): retrieval-augmented Q&A. It answers from Brian's saved article archive and current feed (RSS items plus a search of a curated list of trusted finance sites), optionally the open web, cites every claim, and remembers the thread so you can follow up. Member-gated; every answer carries a per-user monthly dollar cap.
• The Reader (/read, admin-only): a personal reader and archive for online content, pulling the saved archive, an RSS feed and a read-later list into one three-pane view.
• Thought leadership (/thought-leadership): talks, podcasts, press, writing, and a handful of long-form original pieces with their own pages.
• The public site: a bio homepage, a contact form, and /how-this-is-built, which explains where AI actually does work on this site and where a human still signs off.

Site map

Public

• /: bio homepage
• /about: the fuller bio
• /thought-leadership: talks, podcasts, press, writing, and the featured long-form pieces
• /thought-leadership/{slug}: an individual piece (Growth Engine Ratio, the AI hackathon playbook, connecting Claude to NetSuite, and any new one added through the admin) plus a standalone /thought-leadership/growth-engine-calculator
• /tools: the CFO Toolbox landing page, linking to /tools/software and /tools/communities (each with its own directory, profile pages, and a Compare view) and /tools/resources (benchmarking and book recommendations)
• /current-feed: the writers and publications Brian actually subscribes to, as a mixtape tracklist
• /how-this-is-built and /how-this-is-built/{slug}: where AI does real work on this site, and where it doesn't
• /tools/fpa-buddy/how-it-works: the FP&A Buddy explainer
• /play and /play/leaderboard: a small arcade game
• /contact, /privacy

Login-gated

• /tools/fpa-buddy and /ask/history: FP&A Buddy and your own question history (member)
• /read and /read/{article_id}: the merged Feed/Archive/Read Later reader (admin)
• /admin: the back office: article enrichment, dedupe, tag cleanup, the Toolbox admin, user management, brand/voice standards, /admin/checks, backups, and more, grouped by area

API

• /mcp: a remote MCP server (bearer-token auth), covering read access to the Toolbox directory and Compare, the article archive and feed, FP&A Buddy and the Toolbox matchmakers, and read-only database introspection. Separately, scripts/mcp_server.py is a small stdio MCP server that wraps GET /api/search for Claude Desktop/Code.
• /save and /save-later: the bookmarklet's save endpoints (token-only, since the bookmarklet runs cross-origin on someone else's page)

Quick start (local)

pip install -r requirements.txt
cp .env.example .env        # every knob is documented in this file
uvicorn webapp.app:app --reload     # http://localhost:8000

Set ANTHROPIC_API_KEY to enable enrichment and FP&A Buddy. With no LINKLIB_SAVE_TOKEN/LINKLIB_PASSWORD set, the private section is open—local-dev convenience.

The library pipeline

One-time import from Feedly (Preferences → Privacy & Personal Data → "Download your data"):

python -m scripts.import_archive --zip feedly-archive.zip --db library.db

Board names become tags; re-running merges rather than duplicates (URL is the natural key).

Ongoing capture: the /save endpoint plus a bookmarklet (grab it at /bookmarklet when logged in) or an iOS Share Sheet shortcut, and the CLI (python -m scripts.add_link <url>).

Enrichment. Claude writes a clean summary and consistent tags for every article. This is the material FP&A Buddy reads from:

python -m scripts.enrich_backfill --db library.db

Defaults to claude-opus-5 (claude-opus-5-5 is selectable), chosen for summary quality over per-article cost; override with LINKLIB_ENRICH_MODEL, or change it live from /admin/system/ai without a redeploy. To judge whether a different model is worth it before a full re-enrich, compare models side by side on one article:

python -m scripts.enrich_compare --url https://example.com/some-article

FP&A Buddy (Q&A)

Web UI at /tools/fpa-buddy, CLI at python -m scripts.ask "your question". Answers are grounded in the saved article archive and the current feed (RSS items plus fresh web results restricted to the domains in preferred_sites.opml), and, when Open web is turned on, an unrestricted web search, with every citation linked. FP&A Buddy takes a Quick, Standard or Deep effort choice, which maps to a model and token budget internally. Model pickers used elsewhere (re-enrich, backfill) are dynamic (linklib/models.py), reconciled against the live Anthropic Models API so a new Claude model surfaces on its own; the enrichment picker stays curated so a whole-archive re-enrich can't land on an unexpectedly pricey model by accident.

MCP server

Two separate MCP surfaces, for two different jobs:

• scripts/mcp_server.py: a stdio server (pip install mcp) wrapping GET /api/search, for a local Claude Desktop or Claude Code session to search the archive.
• /mcp: a remote MCP server mounted in the same FastAPI process, authenticated per-user with a bearer token (scripts/mint_api_token.py mints one). It's a real tool surface, not a search wrapper: database introspection (list_tables, describe_table, sample_rows, get_rows), the Toolbox directory and Compare (search_software, get_software, compare_software, plus the communities and resources/books equivalents), the article archive and RSS feed (search_archive, get_article, browse_feed, search_feed), and two proxy tools onto FP&A Buddy and the Toolbox matchmakers (ask_fpa_buddy, ask_matchmaker) that respect the same per-user dollar caps the web UI does, and four admin-only benchmark-figure tools (find_figures, coverage, get_digest, metric_definition). Three of the introspection tools are admin-only; the rest follow whatever access level the equivalent web page has.

Other CLI tools

The full, current inventory: purpose, cadence, required env vars, exact invocation, lives at /admin/system/scripts, the live reference this repo keeps in sync with scripts/. A couple of highlights:

scripts/voice_review.py    check any text against the writing-voice standards
scripts/mcp_server.py      stdio MCP server wrapping /api/search (see above)

Project layout

linklib/    core library: db.py (SQLite+FTS5 spine), agent.py (FP&A Buddy),
            matchmaker.py (Toolbox matchmakers), compare.py/gates.py
            (Compare + review-state gating), enrich/extract/pipeline
            (ingest), feed.py (RSS), dedupe.py, tagstyle.py, models.py,
            pricing.py, backup.py (Drive backup), email_utils.py (Gmail API
            email), passwords.py, authcheck.py, brand_check.py,
            voice_review.py/voice_db_scan.py/voice_mechanics.py (voice
            enforcement), sources.py, archive.py, and more
scripts/    CLI entry points (see above); scripts/archive/ holds retired
            one-time migrations, kept for history
webapp/     app.py (nearly every route, HTML/CSS/JS inline), mcp_server.py/
            mcp_toolbox.py/mcp_library.py/mcp_qa.py (the /mcp server, see
            above), checks.py (/admin/checks), tasks.py (admin badges),
            static/
tests/      pytest suite, runs in CI alongside a secret scan

library.db is deliberately not in the repo (personal reading history); CLAUDE.md holds the full architecture notes, auth model, and contributor conventions; BRAND.md holds the visual/voice standards, enforced by CI and /admin/checks.

Hosting

Railway, building from the Dockerfile, auto-deploying from main. If it isn't on main, it isn't live. Changes ship via PR: a branch ruleset on main requires a PR, requires the tests and secret-scan checks, and blocks force-pushes and branch deletion. The database lives on a Railway volume; daily off-site backups go to Google Drive when the GOOGLE_OAUTH_* vars are set (the same OAuth client also powers outbound email—setup steps in .env.example).

License

AGPL-3.0. You're free to use, study, modify, and redistribute this code, including commercially. The one real obligation: if you run a modified version of this app as a network service, you have to offer that modified source to the people using it, not just to whoever you hand a copy to directly. That's the "affero" part, and it's the reason this license was picked over a plain GPL for a web app.