Engineering · Jan 2026
Firebase + Genkit: the best stack for agentic AI in 2026
Agent frameworks come and go. What clients need is stable: auth, data, scheduled jobs, and agents that call Gmail and Sheets without a separate microservice zoo. For nonprofits, small businesses, and ops teams we serve, Firebase plus Genkit is the default — not because it's trendy, but because we already run their volunteer hubs and enrollment flows on the same project.
Why not a standalone agent platform?
Dedicated agent SaaS (or self-hosted LangChain servers) adds another vendor, another login, another bill. Your volunteer roster is already in Firestore. Your digests already run on Cloud Functions. Adding agents as Functions in the same Firebase project means shared auth, shared secrets, shared monitoring — and one handoff document for the client.
What Genkit gives you
- Structured tool definitions — Firestore reads, Sheet appends, Gmail send (post-approval) as typed tools
- Model plugins — swap Groq, OpenAI, Google AI without rewriting orchestration
- Flows you can test — run agent paths locally before deploying Functions
- Telemetry hooks — trace which tool fired and what the model returned
Firebase pieces that matter for agents
Cloud Functions (2nd gen)
HTTP triggers for review UI · Pub/Sub for queue workers · scheduled triggers for batch agent runs
Firestore
Agent run logs, approval queue, roster data the tools read — real-time updates for coordinator dashboards
Firebase Auth
Only approved staff trigger agent runs or confirm outbound actions
Secret Manager
LLM API keys and Gmail service accounts — not in client-side code
Reference architecture
Event (form submit, schedule, manual button) → Cloud Function invokes Genkit flow → agent calls tools (read Firestore, draft content) → writes proposal to approvals collection → coordinator approves in hub UI → second Function executes Gmail/Sheets/publish tools → audit log updated. Same pattern on volunteer matching and enrollment triage.
RAG on Firebase
For SOPs and FAQs, chunk documents into Firestore or Cloud Storage, embed with your chosen model, retrieve at agent runtime. Keep corpora small and scoped — "academy enrollment FAQ" not "every PDF since 2019." Grounded answers reduce hallucinated policy advice.
When we'd choose something else
Heavy Python ML pipelines, strict on-prem requirements, or massive relational analytics might point to Cloud Run + Postgres or a dedicated data warehouse. For web apps with ops automation — the stack we ship every quarter — Firebase + Genkit stays the default.
Related: Volunteer matching agent · Firebase for volunteer platforms · Production-ready agents guide
Building on Firebase already?
We add agent layers to existing projects — typical add-on $4K–$9K.