Sterling CoWorker

Sterling CoWorker is a Slack-native AI operating partner for small businesses, agencies, and virtual assistants. You @mention it/DM it like a coworker and it gets to work inside your Slack Workspace.

Comment 2

Inspiration

Agentic AI works — but only if you're the kind of person who enjoys terminals, MacMinis, VPS boxes, and API keys. I'm a fractional CAIO, so I am that kind of person. The business owners I support are not. They don't want a Mac Mini running an agent in a closet. They want the work done.

Their teams already live in Slack all day. So the question became: can a real agentic AI employee — one that connects to your actual business systems, respects roles & responsibilities per team member, and can actually execute on mission critical items — live entirely inside Slack, installed in one click?

That's Sterling CoWorker.

What it does

Sterling is a Slack-native AI operating partner for small businesses, agencies, virtual assistants. You @mention it or DM it like a coworker, and it:

  • Searches your workspace's memory in real time. Ask "what did we decide about pricing last month?" and Sterling uses Slack's Real-Time Search API to search your workspace's messages and files, then answers with clickable permalinks to the exact messages — so every claim is one click from its source.
  • Works across your business stack. Through secure per-workspace OAuth connections, Sterling can pull reports and take actions in Stripe, Shopify, Klaviyo, Google Workspace, HubSpot, Meta Ads, Social accounts, and 3000+ other tools — no tab-switching, no copy-paste relay between systems.
  • Asks before it acts. Write actions (sending an email, changing a campaign) go through human-in-the-loop approval buttons in Slack. Reads are instant; writes wait for a click.
  • Respects your org chart. Workspace admins control which members can talk to the bot at all, and per-user access levels control which connected tools each person can use — ie. the intern can use the Instagram tools but never sees QuickBooks.
  • Feels alive. Responses stream into Slack token-by-token with live status lines ("Searching your Slack workspace…", "Working in Shopify…") instead of a silent spinner.

How we built it

Slack events (mentions, DMs, App Home, interactive buttons) hit an Express + TypeScript service on Google Cloud Run over signature-verified HTTPS. Each request runs an agentic loop on Gemini 3.5 Flash (via the @google/genai SDK) — up to 18 tool turns — with three tool paths: Slack's Real-Time Search API (assistant.search.context, called with the workspace's own bot token under the granular search:read.* scopes), Composio for OAuth'd business integrations, and approval-gated write actions. Firestore holds workspace state, thread history, approvals, per-user access maps, and a full audit log; secrets live in GCP Secret Manager; usage is metered to Stripe with free trial credits. Onboarding is a one-click install from sterlingcoworker.com with a CSRF-signed OAuth flow that stores a bot token per workspace.

The agent runs as a single Cloud Run service shipped from Docker via Cloud Build CI/CD; the marketing site and one-click onboarding flow are a Next.js / React / Tailwind app on Vercel.

Security was a first-class requirement, not a patch: every piece of user content and every tool result is wrapped in untrusted-data envelopes to blunt prompt injection, secrets are redacted before anything reaches the model or the logs, all actions are audit-logged, and uninstalling triggers a full data purge after a 30-day grace window.

Challenges we ran into

  • Making it fast enough to feel like a coworker. The first cut routed every message through a Hermes CLI subprocess and was too slow. We tore it out and called the Gemini SDK directly, pulling round-trips down to ~3–8s.
  • Reliable in-chat OAuth across 900+ apps. Authorizing Composio connections from inside Slack — without a brittle SDK session — took several rewrites; we ended up driving the connect flow over REST and passing exact connect URLs back into the thread.
  • Multi-tenant token isolation. Every workspace's bot token, integration connections, and per-user access map has to stay walled off from every other tenant. We scoped connections to a teamId entity so one workspace can never see another's data or tools.
  • Prompt injection on untrusted content. Slack messages, file contents, and tool results are all attacker-controllable, so we wrap every one in untrusted-data envelopes and redact secrets before anything reaches the model or the logs.
  • Cost and token burn. An always-on agent with large tool schemas burns tokens fast; stabilizing the prompt prefix for caching and scoping the tool-router to only connected toolkits kept margins sane.
  • Metering money correctly. Usage-based billing means every token has to map to Stripe meter events and trial credits without double-charging or drift.
  • Making Slack feel alive. Streaming token-by-token into a Slack message — within Slack's message-update rate limits — while showing honest status lines took real plumbing.
  • Marketplace-grade compliance. Uninstall data purge, AI disclosures, and signed-cookie / OAuth-state hardening were the last mile to a listable app, and each was its own project.

Accomplishments that we're proud of

  • It's live and multi-tenant, not a demo — any workspace can install it in one click from sterlingcoworker.com.
  • 900+ real integrations with per-user access control — the intern gets Instagram, never QuickBooks.
  • Human-in-the-loop approvals on every write, so an autonomous agent never sends an email or changes a campaign without a click.
  • Real-Time Search answers that cite their sources — every claim is one permalink from the exact Slack message.
  • Security as a feature, not a patch: untrusted-data envelopes, secret redaction, a full audit log, uninstall purge, and signed cookies + OAuth-state.
  • Live token streaming with honest status lines instead of a spinner.
  • Usage-metered billing with free trial credits — a real business model, wired end-to-end.
  • Built solo in ~5–6 weeks, from PRD to a production Cloud Run service.

What we learned

"I'll prompt this over a weekend" turned into two months of real engineering. An AI employee that businesses can trust needs four things — easy setup, real integrations, member-level access control, and security — and each one is its own project. The hardest lessons were the invisible ones.

What's next

Sterling is live and continuing past the hackathon: Slack Marketplace distribution, a customer web dashboard for admins, and deeper proactive workflows (scheduled pulses, anomaly alerts) so Sterling doesn't just answer — it notices. The goal stays the same: take work off people's plates so they can spend time on the things that matter.

We'd love for you to try it: https://sterlingcoworker.com/

And of Course - Join our Slack Community here: https://join.slack.com/t/sterling-coworker/shared_invite/zt-4171rp4ap-AkFY44_7mOZnLV0telpnNw

Built With

Share this project:

Updates