Laki Bits

Laki Bits is an AI-native Salesforce platform I built around two tools that share one account and one security model: LakiQL, an expanded query layer for your org's data, and Laki Agent, an AI technical architect for the org itself. LakiQL reads and answers. Laki Agent finds and fixes.

LakiQL: past the SOQL ceiling

SOQL can't join two objects that aren't linked by a lookup, and its aggregates give out fast. The usual escape hatch is a warehouse — an ETL job to babysit and a copy of your data that's always a little stale.

LakiQL takes ANSI-flavored SQL, a dialect I call SFQL, and compiles it down to whatever Salesforce — or a local engine — can actually run, with a cost-based planner picking the cheapest correct path for each query:

• SOQL — simple filters and relationship joins push straight down, nesting as native parent-child subqueries.

• Anonymous Apex — a join on a non-key field, like matching two objects on an email, compiles into generated Apex. Heavier ones run in two phases: extract each side, then join and aggregate server-side in a second pass.

• A local DuckDB extract — cross-org joins and the rest of what Salesforce won't do, under row and byte guardrails.

You can also just ask in English. LakiQL drafts the SFQL and shows it to you rather than running it behind your back.

Laki Agent: a browser that sees, a CLI that acts

Laki Agent is a macOS app that is a real Chromium browser with an agent living inside it. Salesforce is browser-first — Lightning, managed packages, and admin setup all render client-side — so the agent reads the same page you're looking at through the DevTools Protocol instead of guessing from raw metadata. It acts through the sf CLI you already have installed, which means it never holds your Salesforce tokens; the CLI owns auth and refresh.

Its Salesforce knowledge is grounded rather than improvised. A curated best-practice set is pulled into every run, and mid-task it can run semantic search across the full official Salesforce developer documentation to check an exact API name, signature, or governor limit instead of guessing at one.

The rule underneath all of it is that reading is free and changing waits for you. Every mutation becomes a plan showing the exact command and the target org, and command execution is deny-by-default: anything not positively allow-listed gets the same gate as a write. It ships as a signed and notarized disk image.

Where it stands

Both tools run their Salesforce work as a connected Salesforce user over OAuth, so that user's own permissions carry through — no god-mode integration user sitting behind everything. Models are your choice: bring your own API key, or point at a self-hosted endpoint.

LakiQL runs on the SQL-over-Salesforce engine I originally built as a standalone project, PoutyQL, which is why the two have folded into one. It's live at lakibits.com, invite-only behind a waitlist and free while early access lasts. The SFQL surface is deliberately narrow while it settles — when a query can't be expressed yet, the error tells you why.