LeClaw

Full Stack RevOps. A team of AI agents for your CRM.

Native agents are features inside a single product. LeClaw is infrastructure across your entire GTM stack.

LeClaw is an open-source RevOps agent framework. Deploy a coordinated team of specialized agents into your CRM — each one owns a domain of revenue operations, shares context with the others, and only ever touches broken records.

Hosted product: app.leclaw.io — connect HubSpot, run agents, no terminal required Self-hosted: clone this repo, bring your own keys, run or fork any agent

Architecture

LeClaw uses a French-named multi-agent model:

| Concept | Name | Role | |---|---|---| | Orchestrator | Le Directeur | Dispatches missions, reads rapports, synthesizes insights | | Specialist agents | les agents | Each owns one CRM domain | | Validator | Le Témoin | Reviews proposed changes before write-back | | Coordinated run | une mission | A set of agents dispatched together | | Structured result | un rapport | Filed by each agent, readable by Le Directeur and other agents | | Precise exit | Le Retrait | Agent withdraws immediately if stuck, reports exact reason |

Agents share context through rapports. When le-bdr runs after le-data-quality, it reads the prior rapport and builds on those findings — it does not re-scan.

Agent Library

| Agent | Domain | Status | |---|---|---| | le-data-quality | Field completeness, relationship hygiene | Live | | le-stage-audit | Deal velocity, pipeline health | Live | | le-duplicates | Identity resolution | Roadmap | | le-lead-gen | MQL quality, attribution gaps | Roadmap | | le-bdr | Follow-up SLA, sequence health | Roadmap | | le-routing | Assignment gaps, round robin health | Roadmap | | le-forecast | Commit accuracy, pipeline coverage | Roadmap | | le-deal-desk | Deal structure, discount hygiene | Roadmap | | le-activities | Meeting and call logging gaps | Roadmap | | le-renewal | Renewal risk, upcoming dates | Roadmap | | le-cs | Health scores, expansion signals | Roadmap |

Quick Start

git clone https://github.com/LeRevOps/leclaw
cd leclaw
npm install
npm run setup
npm run data-quality

npm run setup connects your CRM, Anthropic API key, and Slack in under 3 minutes. All credentials are masked on input and saved to .env — never committed, never logged.

npm run data-quality              # auto-detects CRM from .env
npm run data-quality -- --crm hubspot
npm run data-quality -- --crm salesforce

Interactive CLI

Ask Le Directeur questions about your CRM in plain language. It routes to the right agents, runs them, and streams a synthesized answer back.

# Run directly (no install needed)
npx leclaw

# Or after cloning
npm run build
npm run cli

Requires HUBSPOT_TOKEN and ANTHROPIC_API_KEY in your .env or shell environment.

┌─────────────────────────────────────────────────┐
│  LeClaw · Le Directeur                          │
│  orchestrateur · posez une question             │
└─────────────────────────────────────────────────┘

Connecté à HubSpot · 4,821 contacts · 312 deals · 🐳 Docker

Type a question or "exit" to quit.

> why is our forecast unreliable?

Le Directeur dispatche les agents...

  ↳ le-stage-audit 🐳    ✓ 54/100 · 47 issues
  ↳ le-data-quality 🐳   ✓ 61/100 · 83 issues

Le Directeur · synthèse
────────────────────────────────────────────────

Your forecast is unreliable primarily because 31
deals are missing close dates and 19 are past
their close date without being marked closed lost
— these distort your pipeline view directly.
Compounding this, 47 deals have no associated
contact, making it impossible to validate deal
legitimacy or assign follow-up. Start by running
a close date sweep on all open deals in the last
30 days and mark anything stale as closed lost.

Agents: le-stage-audit (54/100) · le-data-quality (61/100)

>

Docker isolation

When Docker Desktop is installed, each agent runs in its own isolated container — resource-limited to 512 MB RAM and 0.5 CPU, removed on exit. The 🐳 indicator confirms isolation is active.

# Build the runner image locally
npm run docker:build

# Or pull from Docker Hub
docker pull leclaw/runner:latest

If Docker is not installed, the CLI falls back to direct in-process execution automatically. No configuration needed.

Note: Docker isolation applies to the CLI only. The hosted dashboard at app.leclaw.io runs agents inside Vercel serverless functions — each request is isolated by Vercel's infrastructure.

How to Build a Custom Agent

Every LeClaw agent is a list of checks. A check is a targeted HubSpot search query that fetches only the broken records matching a specific problem — clean records are never touched.

1. Create your agent

// agents/le-my-agent/index.js
import { runAgent } from "@leclaw/core";

export const leMyAgent = {
  name: "le-my-agent",

  checks: [
    {
      id: "missing_phone",
      label: "Contacts missing phone number",
      objectType: "contacts",

      // Only fetches contacts where phone is null — never scans the full CRM
      filterGroups: [
        { filters: [{ propertyName: "phone", operator: "NOT_HAS_PROPERTY" }] }
      ],

      properties: ["firstname", "lastname", "email", "phone"],
      severity: "warning",
      fix: "Add phone number for outbound sequencing",
      getName: (r) => r.properties.email || r.id,
    },
  ],

  summaryPrompt: (results) => {
    const total = results.reduce((sum, r) => sum + r.count, 0);
    return `Summarize this CRM audit in 2 sentences. ${total} issues found across: ${JSON.stringify(results.map(r => ({ label: r.check.label, count: r.count })))}`;
  },
};

// Run directly
runAgent(leMyAgent, { hubspotToken: process.env.HUBSPOT_TOKEN, anthropicKey: process.env.ANTHROPIC_API_KEY });

2. Add a script

"my-agent": "node agents/le-my-agent/index.js"

Time-based checks

Use a function for timestamp-based checks so the value is computed fresh each run:

filterGroups: () => [{
  filters: [{
    propertyName: "hs_lastmodifieddate",
    operator: "LT",
    value: String(Date.now() - 30 * 24 * 60 * 60 * 1000) // 30 days ago
  }]
}]

Relationship checks

Check associations between objects:

// Contacts with no associated company
filterGroups: [{ filters: [{ propertyName: "associations.company", operator: "NOT_HAS_PROPERTY" }] }]

// Deals with no associated contact
filterGroups: [{ filters: [{ propertyName: "associations.contact", operator: "NOT_HAS_PROPERTY" }] }]

Escalation

A broken record is more critical when it has additional CRM context:

escalateIf: {
  description: "has an open deal",
  filterGroups: [{
    filters: [
      { propertyName: "email", operator: "NOT_HAS_PROPERTY" },
      { propertyName: "associations.deal", operator: "HAS_PROPERTY" }
    ]
  }],
  escalatedSeverity: "critical"
}

See CONTRIBUTING.md for the full agent spec and contribution guide.

Project Structure

leclaw/
  agents/
    le-data-quality/      # Field completeness, relationship hygiene
    le-stage-audit/       # Deal pipeline and velocity
  cli/
    index.ts              # Le Directeur — interactive REPL
  core/
    base.ts               # runAgent(), scoring, callClaude(), all types
    hubspot-search.ts     # Targeted search — only fetches broken records
    hubspot-properties.ts # Dynamic custom property discovery
    registry.ts           # Agent registry
    routing.ts            # Keyword router (zero tokens)
    synthesis.ts          # Le Directeur synthesis prompts
    agent-runner.ts       # Docker container entrypoint
    docker-runner.ts      # Container lifecycle, auto-pull, fallback
  examples/
    le-custom-agent/      # Minimal example to fork
  Dockerfile              # Agent runner image (leclaw/runner)
  setup.js                # Interactive setup wizard

Design Principles

1. Shadow mode by default — read-only until write-back is explicitly enabled. Never touch the CRM without permission.
2. Bring your own keys — Anthropic API key, CRM credentials. LeClaw pays $0 in AI costs.
3. Targeted fetching — agents search for broken records directly. Clean records are never touched.
4. Agents share context — rapports let downstream agents build on prior findings.
5. Le Retrait — an agent that cannot complete its work exits immediately with a precise reason. No spinning, no silent failures.
6. Open source = trust — read the code. SECURITY.md documents exactly what data LeClaw accesses and stores.

Environment Variables

# HubSpot
HUBSPOT_TOKEN=

# AI
ANTHROPIC_API_KEY=

# Slack (optional)
SLACK_WEBHOOK_URL=

Run npx leclaw setup to configure these interactively — it opens HubSpot and Anthropic in your browser, verifies each connection, and writes your .env.

Contributing

See CONTRIBUTING.md.

License

MIT — fork it, extend it, build on it.

Built by a Sales Ops practitioner who spent too much time clicking around.

leclaw.io · app.leclaw.io · LinkedIn