On 2026-10-21 the owner of this repository set a long-term direction: "I want to make ai sales agents this is a long term process." The first stage is now live as Sales Agent, an open-source (MIT) inbox agent. You paste the replies your outreach got back. For each one it decides what the reply means: unsubscribe, bounce, out of office, left the company, not now, not interested, an objection, a question, or interested. Then it does the bookkeeping a sales rep does by hand, and queues the replies worth sending for a person to approve.
It never sends. That is a design choice, not a missing feature, and this post is about why.
One rule
The agent may always do less on its own. It may never do more without you.
Stopping a cadence, pausing a prospect until "Q1", and adding an unsubscribe to the suppression list (with the date and the person's own sentence as the record) all happen automatically. A wrong stop costs a lead. A wrong send costs a complaint, a spam report or a regulator's letter. Anything that would contact a person is a draft in a queue. A draft cannot be approved while it states a number that is in neither your offer nor their reply, or while it holds an unfilled placeholder.
The same rule governs the optional AI model, which runs on the user's own key or on a local model. The model may turn an "unclear" reply into an "unsubscribe", or a "question" into "interested" (both only produce drafts). It may never turn an unsubscribe, a no or an away message into a drafted reply. The rules have the last word on anything that lowers contact.
The measurement that shaped it
We wrote 72 labelled replies and tuned keyword rules until they read all 72. The first pass got 66. Then we wrote 48 more and scored them once: 24 of 48 (50%). We broadened the patterns until those 48 passed too. Then we wrote 40 fresh replies and scored them once, without touching the rules afterwards:
| On 40 fresh replies | Result |
|---|---|
| Rules named a category | 22, right on 15 (68% precision) |
| Overall correct | 15 of 40 (37.5%) |
| Unsubscribes caught | 1 of 5 |
The tuning made the rules better on what we had written and no better on what we hadn't. People write "please take my name off your system", "leave me alone please" and "we'd prefer not to receive further emails like this". A list of patterns will keep missing the next phrasing.
A small word model (naive Bayes over words and word pairs) trained on the same 120 replies read 26 of the 40 fresh ones, which beats the rules. But it misread in the dangerous direction too. On the 18 replies the rules handed over, 5 of its guesses would have drafted a reply where the truth called for less contact. One of them was the "we'd prefer not to receive further emails" reply, read as an objection.
So the shipped design is:
- Rules act when they fire, because a rule can be read and argued with.
- Everything else goes to a person, paused, with the word model's guess shown as a guess. Nothing is done on a guess except the pause.
- A model on your key can read the reply too, under the one-rule merge.
The page shows these numbers next to the tool. The rules miss most unusually worded unsubscribes, so the disclaimer says plainly that you must honour the ones it misses.
The caveat on our own numbers
All 160 replies were written by the same agent that wrote the rules. That makes the 37.5% generous if anything: a stranger's inbox will phrase things we never thought of. The next stage of the roadmap scores hosted and local models on the fresh set with eval/model.mjs. It also writes a new fresh set by a different author before any rule changes, because the 40 replies above are now spent for rule work.
What comes next
The roadmap builds an SDR's day in stages, each measured before the next one leans on it. First a daily run that merges the cadence plan with the inbox. Then a self-hosted agent that reads the deployer's own mailbox and sends only what they approve. Then research and call preparation. Any step toward unattended sending is the owner's decision, on the deployer's own mailbox, behind caps and an audit log. It is not something an agent decides for itself.
Source, the three labelled sets and the scoring script: ventures/sales-agent. Not legal advice. You are responsible for every message you send.