An SFA is "something people fill in" — so it always rots
"Daily logs and deal updates get pushed back, and the fields stay empty…" "We adopted a feature-rich SFA, but no one can master all of it…" "Data entry for the sake of data entry eats the time we should spend selling…"
Sales tools fail for the same reason every time: they're built on the premise that a person does the input. The more features they add, the higher the learning cost — so the team avoids entering data, the data rots, and the system becomes a hollow shell.
Nexa Engineering questioned that premise itself. Move the one who inputs and masters the tool from the human to the AI. That is why we built Hashira SFA in-house — a sales system that AI operates.
The shift: let AI master the features
There is only one way to use Hashira: chat with Claude.ai, just like you always do. That's it.
Write "Met with Company A. Sending a quote by next week" into the conversation, and the AI identifies the account, updates the deal stage, creates the task, and records it on the timeline. Opening screens, filling forms, and hunting for buttons all move to the AI's side.
The human does exactly one thing: review what the AI has organized. Leave the mastery to AI; focus your attention on judgment. This is what AIX (AI Transformation) looks like to us.
Traditional SFA
AI is just one feature bolted onto the app. People operate the SFA and AI only handles a slice — its real capability goes unused.
AI takes the lead. You just talk to it, and it masters the entire SFA.
In a traditional SFA, AI is just one feature of the app (Human → SFA → AI): people operate the system and AI only handles a slice, so its real capability stays structurally unused. Hashira inverts this (Human → AI → SFA) — AI takes the lead, you just talk to it, and it masters the entire SFA (sales-force automation system).
How it works: just chat, and the SFA moves
Hashira connects directly to Claude.ai as an MCP (Model Context Protocol) server.
You ⇄ Claude.ai ⇄ MCP ⇄ Hashira SFA (production DB)
Claude.ai automatically picks from the 24 operation tools Hashira exposes, as the situation calls for it — searching, creating, and updating accounts, deals, tasks, and timelines that a person used to handle on screen, now inferred from the flow of conversation. The connection is protected by OAuth 2.1 (dynamic client registration, PKCE, refresh tokens), and you choose the target tenant at consent time.
The four domains AI manages automatically
| Domain | What the AI does for you |
|---|---|
| Deal pipeline | Judges the pre/post-close stage from the conversation and updates the kanban. Won = ongoing, closed = lost/completed, sorted automatically. |
| Tasks | Detects promises like "I'll send it by ~" and turns them into tasks — due date, owner, and deal linkage included. |
| Accounts & contacts | Resolves and registers the companies and people mentioned, with automatic tagging. |
| Timeline | Records every exchange in the account/deal chronology, so you can trace the history later. |
The design that lets people verify
Precisely because we hand work to AI, we made it impossible to run wild and always auditable — our top priority.
- Every write is logged — Any create, update, or delete by AI lands in
mcp_audit_log, without exception. You can always check when, through whose connection, and what was changed. - Full tenant isolation (RLS) — Row Level Security on every business table, deny by default. The connected user's JWT can only reach its own tenant's data.
- Humans give final approval — The AI proposes and executes; people hold the call on what's right. The workflow is built around review.
Combine "AI is fast" with "people can safely delegate." That is the heart of Hashira.
Where AI goes further: cross-search and CRM automation
- AI cross-search — The AI summarizes, classifies, and extracts text from attachments and activity records, so you can find things by meaning through semantic search powered by Voyage embeddings.
- CRM lifecycle — It manages contracts, recurring revenue, transaction history, LTV, and health scores, and a Cron job automatically detects renewal timing to prompt follow-up.
Tech stack
| Area | Technology |
|---|---|
| Frontend | Next.js 15.3 (App Router / Server Actions) / React 19 / TypeScript |
| Backend | Supabase (PostgreSQL + Auth + RLS, Tokyo region) / Zod (boundary validation) |
| AI integration | MCP server (@modelcontextprotocol/sdk) / Claude.ai connector / Voyage embeddings |
| Security | OAuth 2.1 (DCR / PKCE / refresh) / Cloudflare Turnstile / Upstash rate limiting |
| Infra | Vercel (Tokyo hnd1) / Resend (email) |
Live in production: sfa.nexa-eng.com — a multi-tenant design we use to run our own sales.
Hashira, too, was born by changing the build order
Hashira didn't aim for the finished form from day one. We shipped the foundation first — account and deal management — then built out timeline, tasks, dashboards, and finally the AI integration via MCP, step by step, while actually using it.
"Start from a working thing, before the spec document" — the very approach we propose to clients, practiced on our own core system.
Free your sales from "data entry," too
"I'd like to talk about whether this fits our sales flow" or "I want a business system that AI runs" — whatever it is, get in touch.