← All work

Open source, MIT · 2026

RevLoop

One current, sourced view of every account for GTM teams and their agents, plus a report that shows whether agent actions moved pipeline.

5MCP tools exposed to agents

The problem

What a team knows about a deal sits in calls, emails and the CRM, and it changes every week. The CRM says the deal is in negotiation; Tuesday’s call says it is dead. An agent drafting from the CRM gets it wrong, and nobody can tell whether the agents help at all.

Account memory · zsh
# Northwind Robotics · account context, with history

next_step   Send security questionnaire responses by Friday
            emails/2026-09-15-northwind-robotics.eml:12   2026-09-15
  superseded  Legal review of the MSA
              crm/salesforce-opportunities.csv:3          2026-09-10

stage       negotiation      crm/salesforce-opportunities.csv:3
competitor  Vantage          calls/2026-09-08-northwind-robotics.txt:18
objection   Procurement needs SOC 2 before signing

Proposed actionstalled-deal

Send the SOC 2 report with the questionnaire answers.

From objection · calls/…northwind-robotics.txt:18

ApproveReject

The agent proposes. A person approves. Nothing is sent automatically.

My role

Designer and engineer

View the code on GitHub ↗

Results

7
fact types extracted from calls, email and CRM
MIT
licensed and public on GitHub

Stack

  • Python 3.12
  • SQLite
  • MCP
  • Claude
  • YAML playbooks
  • uv
  • pytest

What changes

Account context

WithoutCalls, emails and the CRM disagree, and agents draft from stale CRM fields.

With RevLoopOne current view per account. Every fact points to the file and line it came from.

Agent actions

WithoutAgents propose follow-ups that nobody signs off.

With RevLoopAgents propose through MCP. A person approves or rejects every action.

Measuring impact

WithoutNo way to tell whether agent actions moved pipeline.

With RevLoopApproved accounts compared against a holdout, with the selection-bias warning printed.

Ownership

WithoutContext locked in a vendor’s cloud.

With RevLoopSQLite and plain files on the team’s own machine.

How it works

At a high level. A detailed walkthrough is available on a call.

  1. 01Sources
  2. 02Extract
  3. 03Remember
  4. 04Propose
  5. 05Approve
  6. 06Measure
  7. 07Serve
  1. Decision 01

    Newer stage, amount and next step supersede the old value. Pains, objections, competitors and stakeholders accumulate and dedupe.

    The current view stays current, and the old fact is kept as history.

  2. Decision 02

    Agents can propose actions through MCP but cannot approve them.

    An agent should not sign off on its own work.

  3. Decision 03

    Offline extraction rules by default, Claude only when asked for.

    It runs without an API key and gives the same answer every run.

  4. Decision 04

    The outcome report compares against a holdout and prints its own selection-bias warning.

    A human chose which actions to approve, so the uplift is not a randomised result, and the report says so.

  5. Decision 05

    SQLite and plain files on the team's own machine.

    Teams can inspect, back up and own their context layer.