The wow moment · In three acts

Built. Queried. Taught.

The same Pat at Acme, all the way through. Act 01: the profile assembles itself from a real chat. Act 02: an operator pulls it back through MCP. Act 03: the admin teaches the schema the company's own vocabulary, and the whole history answers to it.

Act 01
Built from conversation
A messenger thread on the left. The Ediccio profile materializing on the right. Every record traces back to the message that produced it.
Messenger · Pat ↔ Nimbus
online · today, 09:14
end_user_id pat@acme.com
profile · building · async
0 obs
awaiting first signal
0.0s / 16.0s
loop
Pain
Value and severity, with the evidence.
Lifted from a four-word sentence. No taxonomy mapping, no manual tag.
Sentiment
Polarity & intensity, not a score.
The full timestamped series lives in the profile. Confidence travels with the value.
Intent
A verb-driven label, plus polarity.
Free-form, not an enum. The agent reads it and acts.
Traditional fields
Role, company, size, location.
No form ever filled. Pulled from one sentence in line three.
Preferences
Timing & follow-up channel.
The data that turns a stalled lead into a booked call.
Conversation lands
Slack, email, support chat, sales call, anywhere your customer speaks.
↓
Profile assembles
Five dimensions, each with the message that produced it as evidence.
↓
Any agent can read it
Any MCP-compatible host: your own agent loop, a chat client, a scheduled job. Same record, same MCP tool.
Act 02
Queried by any agent
Rae in RevOps opens her chat client. Asks one question. The MCP tool fetches the profile from Act 01. The synthesis writes itself.
Nimbus RevOpsworkspace
MCP: ediccio · connectedRS
Acme · Pat
3 turns · started 14:08
RS
Rae Suresh · RevOps

What's the latest on Pat at Acme? Are they close to ready, and what should I do this week?

Ask a follow-up…
⌘ + ↵ to send
0.0s / 18.0s
loop
One prompt
No retrieval prompt, no RAG glue.
The MCP tool list lives in the chat client. Pat's record is one tool call away.
Real tool call
Visible, expandable, auditable.
Arguments in. Structured JSON out. Receipt of what the agent actually saw.
Synthesis
Talks to the operator as a peer.
Reads the record, draws the conclusion, suggests next moves. No "I'm here to help."
Citations
Every claim, traceable.
The dimensions cited link back to the records, and through them, the messages.
Acts 01 & 02 used the agent key
It writes signals and reads profiles. It cannot touch what they mean.
↓
Act 03 switches to the admin key
It can change what profiles mean. It can never read one.
Act 03
Taught your language
The admin, in their own chat client: define an objection dimension and a prospect journey in plain English, dry-run the proposal on a real ticket, confirm it, then replay the history under the new schema with the cost shown first. The same sequence ran on our own tenant.
Nimbus · adminschema & replay
MCP: ediccio · admin keyPK
Teach it our language
admin credential · schema tools only
PK
Priya Kaur · founding engineer

Our profiles should speak our language. We track objections, and prospects move subscriber → engaged → qualified. Set that up, test it on a real ticket, and re-read our history.

PK
Priya Kaur · founding engineer

Run it.

Refine a definition, add a stage…
⌘ + ↵ to send
0.0s / 28.0s
loop
Propose
A diff and a token. Nothing active.
The definition is data, written in your words. You read exactly what would change before anything does.
Dry run
One real extraction. Nothing stored.
A real ticket through the proposed schema. You see what would land, including what gets dropped as ungrounded.
Confirm
Forward-only from here.
New signals project under v2 immediately. History stays untouched until you explicitly ask.
Replay
History answers to the new words.
Cost estimate first, one subject per transaction, meters the whole way. Never automatic.
Three acts, one record

Built once. Read by every agent. Taught by you.

The same profile that powers your customer-facing AI is the one your RevOps team queries Monday morning and the one you re-teach in an afternoon when your vocabulary outgrows ours. One record, every surface.

Provenance, the whole way

The evidence quote that lit up in Act 01 is what your operator cited in Act 02, and the dry run in Act 03 shows its receipts before anything is stored. Truth travels.

Same MCP tools, every audience

Customer-facing agent, RevOps chat, batch job: the tool list is identical. The admin surface is the same protocol on a separate key.

One pricing meter

Signals: each conversation input we extract from. Reads are never metered, and profiles are never metered, no matter who does the reading.

Try it against your own conversation.