Nisa AI · DramaLabs
The AI co-worker that texts you first.
Founding Engineer · DramaLabs
Nisa is an AI chief-of-staff for small teams that lives where they already talk: Slack, WhatsApp and iMessage. She reads an authored company brain, runs scheduled and event-triggered work across the team's apps, and hands back drafts and decisions, never sending anything without an explicit approval. I took her from prototype to production-grade myself: the agent harness, the channels, the memory, the approval layer and the infra that keeps it honest.
- Agent harness: multi-model tool-calling loop with provider failover; script skills run in sandboxed microVMs, agent skills as tool-calling loops
- Skills: 40 of them (SKILL.md + references + scripts), run on a schedule, on a trigger, or mid-chat via /name, plus one that writes new skills from a plain-language request
- Company brain: versioned markdown per team and per person, typed facts with provenance, hybrid keyword + vector retrieval
- Approval-first autonomy: every outbound action is an Approve / Edit / Cancel card, one code path gates every surface, and standing "always allow" permissions are narrowly scoped
Technical detailsHide details
How it works: a message from Slack, WhatsApp or iMessage enters one agent loop. Nisa pulls context from the company brain, picks the right skill, calls the team's tools over MCP (sandboxed when it runs code), and answers on the same channel. Anything that would leave the building comes back as an approval card first, and the turn is distilled back into memory.
She's proactive by design: scheduled skills send the daily brief, flag stalled threads and deals, and prep meetings before anyone asks. That's the "texts you first" part.
- Self-maintaining memory: every conversation turn is distilled into memory; a nightly "dream" pass retires stale facts, and nothing merges without approval
- Channel-native: Slack Block Kit cards with live status and a stop button, WhatsApp interactive buttons, iMessage approvals, voice notes on every channel
- Tools & isolation: 3,000+ tools via MCP and Composio, behind multi-tenant, fail-closed isolation
- Observability & QA: Langfuse tracing, golden-set and LLM-judge evals; production deploys are blocked until a cross-channel QA simulation passes on the exact commit