Roadmap — individuals → teams → orgs (partner mindset)

Status: the client (this llm-spendguard package) is production-ready and works standalone. The hosted team/org server (e.g. llmseg.ai) is a separate repo, currently in development — the URLs below describe where it will live, not a live service yet.

The guiding principle: partner, not supervisor. Every user sets their own limits and keeps their own ledger. Teams and orgs get visibility and pooled learnings by roll-up, not control. Nobody's caps are locked from above; the value is shared sight and shared knowledge, opt-in.

Tiers

Tier	Who	What they get	How
I (developer)	one person	local gate + caps they set, local ledger, local learnings	pip install llm-spendguard (this repo) — already shipped
We (team)	a team	see each other's usage, pooled + aggregated learnings, team rollup report	paste the server key (one key, server maps it to your team) + opt-in sync
Us (org)	an org	org-wide visibility, aggregate advice across teams, org rollup	members use keys the server maps to the org — no ids in client config

Roll-up is additive and opt-in: a user chooses how much to share (full usage / costs-only / scrubbed-learnings-only) in their own config. The default is private; sharing is a choice, and it's their choice. Greater visibility is something a user turns on — the partner mindset.

Two repos

• llm-spendguard (this repo, open): the client — gate, pricing, ledger, advisor, learnings. Each user has their own ledger here. It gains a small, optional sync capability: with a team_id/org_id + token in config, it pushes its ledger summary + scrubbed learnings to the SaaS for aggregation. No server code lives here.
• llm-spendguard-server (separate repo, the SaaS — e.g. llmseg.ai): login / create account / credit card / define team & org → get team/org ids to drop in your client config. It ingests members' pushed data and provides: cross-member visibility dashboards, aggregate learnings, aggregate advice, and rollup reporting. This is the open-core → hosted-collaboration layer; build it as its own repo.

The client never depends on the server (works fully standalone, fail-open). The server is purely additive visibility/aggregation over what users opt to send.

Client seam (BUILT — saas.py, spendguard saas)

The client is ready to connect the moment the server exists: - Config lives in ~/.spendguard/saas.json (gitignored; template saas.example.json): enabled, url (e.g. https://api.llmseg.ai), api_key (secret; or SPENDGUARD_SAAS_KEY), visibility (private | team | org), sync_interval (off | hourly | daily | weekly). Surfaced in spendguard config like every other knob. - One key = identity. The client holds NO team_id/org_id — the SERVER maps the Bearer key to the user→team→org hierarchy. Less to leak, nothing to keep in sync. - Contract (saas.py, versioned {url}/v1): GET /v1/health, POST /v1/ledger (per-day roll-up), POST /v1/insights (scrubbed abstracts), GET /v1/insights?scope= (pooled learnings). Bearer auth. - Cadence: sync_interval drives when the roll-up pushes. saas.sync(if_due=True) is wired into the daily report cron (and spendguard saas sync --if-due), so it pushes on schedule and no-ops otherwise. last_sync tracked in saas_state.json. - Fail-safe by design: every call degrades gracefully ("not connected") until the server is up; the client never depends on it. visibility=private = nothing leaves the machine. Reuses share.py's scrub (abstracts only — never prompts/keys/$). Each user's SQLite ledger stays the source of truth. - Status/test: spendguard saas status · spendguard saas ping · saas sync [--if-due] · push / pull.

Access — slash commands (shipped)

spendguard install-skills deploys /spend (quick status: totals + leak + top learnings) and /spendguard-learn (the advisor) into ~/.claude/skills/, so they work as slash-commands in Claude Code (CLI + the VS Code extension). A native VS Code panel is possible later but optional — the skills cover both surfaces today.

Build order

1. (done) Client: gate / pricing / ledger / advisor / learnings / reconcile-ledger / report / slash-commands.
2. (done) Client sync seam — saas.py + saas.json config + spendguard saas; speaks the /v1 contract, fail-safe until the server exists.
3. llm-spendguard-server (separate repo, NEXT): implement the /v1 contract — account/team/org, ingest, aggregate visibility + learnings + advice, billing. llmseg.ai. The key-holding proxy (the only true no-bypass guarantee) is the natural first milestone here.
4. Publish client to PyPI; orgs use private index / base image / install-hook for fleet rollout.