Flagship project · GTM AgentOS

Revenue operations that can explain every decision.

An AI-native platform that qualifies leads, routes agent workflows, retrieves grounded internal knowledge, exposes controlled tools, and requires human approval before sensitive external actions.

Watch 84s demo View on GitHub
PositioningProduction-style engineering
ValidationMajor flows tested end-to-end
Test suite118 automated tests
GTM AgentOS operations console showing synthetic portfolio data

Problem

GTM automation becomes risky when decisions disappear inside disconnected tools.

Qualification, research, tools, and outbound actions need shared state, clear boundaries, and a durable audit trail.

Solution

One controlled system from lead intake to operator decision.

GTM AgentOS combines structured AI outputs, explicit graph routing, grounded evidence, allowlisted MCP tools, approval gates, and production-style observability.

End-to-end demo

See the complete operating model.

Portfolio Mode uses synthetic data and does not call external providers during capture.

Architecture

A visible path from input to accountable action.

Each boundary has a narrow responsibility and leaves evidence for the next one.

  1. 01
    Lead ingestion

    Idempotent intake creates one durable lead record.

  2. 02
    Claude qualification

    Structured output produces score, classification, and next action.

  3. 03
    LangGraph orchestration

    Explicit state transitions route HOT, WARM, and COLD paths.

  4. 04
    Grounded retrieval

    Voyage embeddings and pgvector return ranked internal knowledge.

  5. 05
    Controlled MCP tools

    Allowlisted, validated calls are audited before reaching a handler.

  6. 06
    Human approval

    Sensitive external actions remain pending until an operator decides.

  7. 07
    Observability

    Runs, transitions, evidence, latency, usage, and actions form one timeline.

Key engineering challenges

The hard parts live at the boundaries.

01

Structured LLM output validation

Qualification stays inside a typed contract before it can affect routing.

02

Lead and action idempotency

Repeated requests do not create duplicate leads or duplicate provider calls.

03

LangGraph state auditing

Every routing decision is persisted from START through END.

04

Grounded retrieval

Research is built from ranked internal evidence, with a safe no-context fallback.

05

MCP allowlisting

Only six explicit read-oriented tools are discoverable; unknown tools are rejected and audited.

06

Human approval gates

Sensitive actions cannot dispatch while they remain pending.

07

HMAC and replay protection

n8n requests are signed, time-bounded, and guarded against reuse.

08

Operator session security

The console is authenticated without exposing service credentials to the browser.

PythonFastAPISupabase / PostgreSQLClaudeLangGraphVoyage AIpgvectorMCPn8n

Operations console

Evidence for the operator, not just output from the model.

The console connects portfolio-safe metrics, run inspection, RAG sources, tool controls, approval state, and the lead timeline.

GTM AgentOS operations overview with synthetic metrics
Operations overview — synthetic portfolio data
GTM AgentOS run inspector showing qualification, RAG evidence, MCP tools, and an external action
Run inspector — qualification, evidence, tools, and action state
GTM AgentOS lead timeline showing LangGraph transitions through END
Lead timeline — state transitions and complete audit trail

End-to-end validation

Validated as a system, not a collection of screenshots.

Major flows were validated across the real Supabase, Claude, Voyage AI, pgvector, MCP, n8n, HubSpot, and Resend integrations during the documented project phases.

Public demos use synthetic data. This case study claims production-style engineering and end-to-end validation — not a production deployment.

Inspect the implementation

The complete source, architecture, migrations, workflow, and tests are public.

View GTM AgentOS on GitHub