The Problem

Most enterprise Claude deployments stall — not because of the AI, but because of what's missing above it.

You get an API key. You wire up a chat loop. It works in staging. Then legal asks: "Where are the audit logs?" Compliance asks: "Are we redacting PII before it hits the API?" Finance asks: "How much is the dev team spending per day?" And the rollout team asks: "How do we gate Wave 2 access until Wave 1 hits 80% adoption?"

The Anthropic SDK is not built to answer those questions. enterprise-claude-kit is.

It's the production-grade layer that handles governance, cost control, audit trails, and adoption tracking — so you can focus on the application, not the infrastructure.

Architecture

flowchart TD
    APP(["🖥️  Your Application\n─────────────────────\nawait agent.run(prompt, user_id)"])

    subgraph ECK ["  enterprise-claude-kit  "]
        direction TD

        subgraph ORCH ["⚙️  AgentOrchestrator"]
            AGT["Named Agent · system_prompt · persona · tier · MCP connectors"]
        end

        subgraph PIPELINE ["  5-Stage Async Pipeline — wraps every Claude call  "]
            direction TB
            S1["🛡️  ① GovernanceLayer  —  Pre-flight\nPII detection · blocked keywords · prompt-length guard · persona allowlist"]
            S2["☁️  ② Claude API\nclaude-sonnet-4-6  ·  claude-haiku-4-5  ·  claude-opus-4-6"]
            S3["🛡️  ③ GovernanceLayer  —  Post-response\nPII in output · GxP citation check · pre/post hooks"]
            S4["💰  ④ TokenMonitor\ncost = tokens × price  ·  daily budget enforcement  ·  async alert callbacks"]
            S5["📋  ⑤ AuditLogger\nappend-only SQLite  ·  SHA-256 checksum  ·  user-ID hashed  ·  GxP-ready"]
            S1 --> S2 --> S3 --> S4 --> S5
        end

        subgraph OPS ["  Operational Layer  "]
            direction LR
            AT["📈  AdoptionTracker\nWave-gated rollout · Literacy scoring\nInactive-user detection"]
            MCP["🔌  MCP Registry\nGitHub · Jira · Slack · Confluence\nSharePoint · PostgreSQL · +5 more"]
            CLI["⌨️  ecl CLI\necl cost summary --days 7\necl audit query\necl waves list"]
        end

        subgraph DB ["  Zero-infra SQLite persistence  "]
            direction LR
            MD[("monitor.db\ncost records")]
            AD[("audit.db\nevent trail")]
            WD[("adoption.db\nwave state")]
        end
    end

    ERR(["⛔  GovernanceViolation\n    BudgetExceededError\n    WaveGateError"])
    OUT(["✅  RunResult\ncontent · cost_usd · input_tokens\noutput_tokens · governance_result"])

    APP --> ORCH
    ORCH --> S1
    S1 -- "❌ PII / blocked keyword" --> ERR
    S4 -- "❌ daily budget exceeded" --> ERR
    S5 --> OUT
    S4 --> MD
    S5 --> AD
    AT --> WD
    CLI -. "reads" .-> MD & AD & WD
    ORCH -. "wave-gated access" .-> AT
    ORCH -. "injects tool definitions" .-> MCP

    style APP  fill:#f0f9ff,stroke:#0284c7,color:#0c4a6e
    style ERR  fill:#fef2f2,stroke:#dc2626,color:#991b1b
    style OUT  fill:#f0fdf4,stroke:#16a34a,color:#14532d
    style S1   fill:#fefce8,stroke:#ca8a04,color:#713f12
    style S2   fill:#eff6ff,stroke:#3b82f6,color:#1e40af
    style S3   fill:#fefce8,stroke:#ca8a04,color:#713f12
    style S4   fill:#fff7ed,stroke:#ea580c,color:#7c2d12
    style S5   fill:#f0fdf4,stroke:#16a34a,color:#14532d
    style AGT  fill:#faf5ff,stroke:#7c3aed,color:#3b0764

Quick Start

pip install enterprise-claude-kit
cp .env.example .env   # add your ANTHROPIC_API_KEY

import asyncio
from enterprise_claude import AgentOrchestrator, GovernanceLayer, TokenMonitor

governance = GovernanceLayer(pii_filter=True, constitutional_ai=True)
monitor    = TokenMonitor(daily_budget_usd=50.0)

async def main():
    async with AgentOrchestrator(governance=governance, monitor=monitor) as orch:
        agent  = await orch.create_agent(
            name="analyst", system_prompt="You are a concise analyst.", persona="analyst"
        )
        result = await agent.run("Summarise Q3 risks in 3 bullets.", user_id="alice")
        print(result.content)        # the answer
        print(result.cost_usd)       # e.g. 0.000312
        print(result.governance_result.passed)  # True

asyncio.run(main())

Three lines of config. One await. Full governance, cost tracking, and audit trail included.

Modules

🛡️ GovernanceLayer

Every Claude call passes through the governance layer first. Configure it once; it runs everywhere.

from enterprise_claude import GovernanceLayer

governance = GovernanceLayer(
    pii_filter=True,                            # redact emails, SSNs, phone numbers
    blocked_keywords=["classified", "MNPI"],    # hard-stop on sensitive terms
    constitutional_ai=True,                     # self-critique harmful content
    gxp_mode=True,                              # require [SOURCE:] citations (pharma)
    max_prompt_length=50_000,                   # guard against prompt-stuffing
    pre_hooks=[lambda prompt, ctx: log(prompt)],
    post_hooks=[lambda resp, ctx: verify(resp)],
)

GovernanceResult tells you exactly what happened:

result = await agent.run("Who is the patient John Smith?", user_id="alice")

gov = result.governance_result
print(gov.passed)        # False — PII detected
print(gov.pii_detected)  # ["john smith"]
print(gov.flags)         # ["pii_in_prompt"]
print(gov.violations)    # ["pii_blocked"]

Exceptions give you clean programmatic control:

from enterprise_claude import GovernanceViolation, BudgetExceededError

try:
    result = await agent.run(prompt, user_id="alice")
except GovernanceViolation as exc:
    print(exc.violation_type)   # "pii_detected" | "blocked_keyword" | …
except BudgetExceededError as exc:
    print(exc.details)          # {"budget_usd": 50.0, "spent_usd": 50.003}

🤖 AgentOrchestrator

Register named agents with pre-configured system prompts, personas, and tool bindings. Deploy them to your whole org. Change their config in one place.

from enterprise_claude import AgentOrchestrator

async with AgentOrchestrator(governance=governance, monitor=monitor, audit_logger=audit) as orch:

    # Agents are reusable — create once, run many times
    code_reviewer = await orch.create_agent(
        name="Code Reviewer",
        system_prompt="You are a senior engineer performing security-focused code review.",
        persona="engineering",
        tier="default",                              # "default" | "batch" | "gated"
        mcp_connectors=["github", "jira"],           # inject tool definitions
    )

    # Run against any prompt, any user
    result = await code_reviewer.run(diff_text, user_id="bob@acme.com")
    print(f"Review: {result.content}")
    print(f"Cost:   ${result.cost_usd:.6f}  |  Tokens in/out: {result.input_tokens}/{result.output_tokens}")

💰 TokenMonitor

Real-time cost tracking, per-user budgets, and alerting — backed by SQLite so nothing is lost on restart.

from enterprise_claude import TokenMonitor
from datetime import UTC, datetime, timedelta

monitor = TokenMonitor(
    daily_budget_usd=100.0,
    alert_threshold_pct=0.80,   # fire alert at 80 % of budget
    db_path="monitor.db",
)

# Async alert callback — fires when the threshold is crossed
@monitor.on_alert
async def on_budget_alert(status) -> None:
    await slack.post(f"⚠️ Budget at {status.pct_used:.1f}% — ${status.used_usd:.2f} of ${status.budget_usd:.2f} used")

# Query spend for any time window
now     = datetime.now(tz=UTC)
summary = await monitor.get_cost_summary(start_date=now - timedelta(hours=24), end_date=now)

print(f"Total:      ${summary.total_usd:.4f}")
print(f"By model:   {summary.by_model}")    # {"claude-sonnet-4-6": 0.042, …}
print(f"By persona: {summary.by_persona}")  # {"analyst": 0.018, "engineer": 0.024}
print(f"Calls:      {summary.record_count}")

# Check the current budget gauge
budget = await monitor.check_budget()
print(f"{budget.pct_used:.1f}% used — ${budget.remaining_usd:.4f} remaining")

📋 AuditLogger

An immutable, append-only event log with SHA-256 tamper detection. Required for SOC 2, HIPAA, and 21 CFR Part 11 compliance programmes.

from enterprise_claude.audit import AuditLogger, AuditFilter

audit = AuditLogger(db_path="audit.db", gxp_mode=True, retention_days=365)

# Events are written automatically by AgentOrchestrator — no manual logging needed.
# Query them at any time:
page = await audit.query_events(AuditFilter(
    persona="clinical_ops",
    governance_result="pass",
    page=1,
    page_size=50,
))

for event in page["events"]:
    print(f"{event.timestamp}  {event.agent_id}  {event.governance_result}  {event.cost_usd:.6f}")

print(f"Total events: {page['total']}  (page {page['page']} of {page['pages']})")

# Tamper detection — verify every stored SHA-256 checksum
checksums = await audit.verify_checksums()
tampered  = [eid for eid, ok in checksums.items() if not ok]

if tampered:
    raise RuntimeError(f"Audit log tampered — {len(tampered)} event(s) modified: {tampered}")
else:
    print(f"✅ All {len(checksums)} events verified clean")

📈 AdoptionTracker

Wave-based rollout management. Gate Wave 2 access until Wave 1 hits 80%. Track literacy scores. Surface inactive users before your renewal conversation.

from enterprise_claude.adoption_tracker import AdoptionTracker, WaveGateError

tracker = AdoptionTracker(db_path="adoption.db")
await tracker.initialize()

# Define a gated rollout — Wave 2 requires 80 % of Wave 1
wave1 = await tracker.create_wave(name="Architects", target_count=50, order=1)
wave2 = await tracker.create_wave(
    name="Developers", target_count=200, order=2,
    gate_wave_id=wave1.wave_id, gate_threshold_pct=0.80,
)

await tracker.activate_wave(wave1.wave_id)

# Wave 2 is blocked until the gate is met
try:
    await tracker.activate_wave(wave2.wave_id)    # → WaveGateError
except WaveGateError as exc:
    print(exc)   # "Wave 'Developers' requires 'Architects' to reach 80.0% (currently 0.0%)"

# Record activations and call activity
await tracker.record_activation(wave1.wave_id, "alice@acme.com", persona="architect")
await tracker.record_call("alice@acme.com")

# Progress + literacy
progress = await tracker.get_wave_progress(wave1.wave_id)
print(f"Wave 1: {progress.activated}/{progress.target}  ({progress.completion_pct:.0%})")

score = await tracker.get_literacy_score("alice@acme.com")   # 0–100
print(f"Alice's literacy score: {score:.0f}")

# Find users who haven't called the API in 7 days — candidates for re-engagement
inactive = await tracker.get_inactive_users(days=7)
print(f"Inactive: {inactive}")

🔌 MCP Connector Registry

Ten pre-built connector configs for GitHub, Jira, Slack, Confluence, SharePoint, PostgreSQL, ServiceNow, Salesforce, Teams, and the local filesystem — with env-var validation built in.

from enterprise_claude import get_connector, list_connectors, MCPConnectorRegistry

# Fetch a config and check whether env vars are present
github = get_connector("github")
print(github.display_name)          # "GitHub"
print(github.auth_type.value)       # "bearer_token"
print(github.required_env_vars)     # ["GITHUB_TOKEN"]

# List everything available, with env-var readiness flag
for connector in list_connectors():
    status = "✅" if connector["env_configured"] else "❌ missing env vars"
    print(f"{connector['display_name']:<20}  {status}")

# Register a custom connector
registry = MCPConnectorRegistry()
registry.register("my-data-lake", ConnectorConfig(
    name="my-data-lake",
    display_name="Acme Data Lake",
    url="@acme/mcp-server-datalake",
    auth_type=AuthType.OAUTH2,
    required_env_vars=["DATALAKE_CLIENT_ID", "DATALAKE_SECRET"],
    description="Query the Acme internal data lake",
    documentation_url="https://internal.acme.com/data-lake/mcp",
    tags=["data", "analytics"],
))

⌨️ CLI — ecl

A full management CLI ships with the package. No code needed for day-to-day ops.

# Agents
ecl agents list                         # list all registered agents
ecl agents deploy --config agents.yaml  # deploy from YAML spec

# Cost & usage
ecl cost summary --days 7               # spend by model and persona, last 7 days
ecl cost export --format csv --out spend.csv

# Waves
ecl waves list                          # show all waves + progress bars
ecl waves activate <wave-id>            # open a wave

# MCP connectors
ecl connectors list                     # all connectors + env-var status
ecl connectors validate github          # check GITHUB_TOKEN is set

# Audit
ecl audit query --persona clinical_ops --result pass --limit 100
ecl audit export --format csv --out trail.csv

Inspired by Real Enterprise Deployments

This library is modelled on how Anthropic's largest partners deploy Claude at Fortune 500 scale.

The patterns here — wave-gated rollouts, per-user cost envelopes, GxP-mode audit trails, pre/post governance hooks — are drawn from production deployments in financial services, pharmaceuticals, and defence. The specific firms aren't named, but the problems are real:

• A global bank that needed per-trader cost caps before their compliance team would approve Claude access
• A top-10 pharma running clinical-trial summarisation that required 21 CFR Part 11–style audit trails
• A systems integrator rolling Claude to 8,000 engineers in waves, gated on adoption metrics

enterprise-claude-kit is the distillation of those patterns into a single, pip-installable library.

Comparison

Examples

The examples/ directory contains three runnable demos — each needs only ANTHROPIC_API_KEY in .env:

python examples/basic_governed_agent.py
python examples/clinical_trial_agent.py
python examples/sdlc_accelerator.py    # offline — no API key required

Installation

Minimal (governance + orchestrator only):

pip install enterprise-claude-kit

Full (CLI, YAML deploy, Rich terminal UI):

pip install "enterprise-claude-kit[cli]"

Development:

git clone https://github.com/sairajboddula/enterprise-claude-kit
cd enterprise-claude-kit
pip install -e ".[dev]"
pytest          # 66 tests
mypy enterprise_claude/
ruff check .

Environment variables

cp .env.example .env

# Required
ANTHROPIC_API_KEY=sk-ant-…

# Optional — override defaults
ECL_DEFAULT_MODEL=claude-sonnet-4-6
ECL_DAILY_BUDGET_USD=100.0
ECL_ALERT_THRESHOLD_PCT=0.80
ECL_AUDIT_DB_PATH=audit.db
ECL_MONITOR_DB_PATH=monitor.db
ECL_ADOPTION_DB_PATH=adoption.db

Project Structure

enterprise_claude/
├── governance.py        # GovernanceLayer — the policy engine
├── orchestrator.py      # AgentOrchestrator — agent lifecycle
├── token_monitor.py     # TokenMonitor — cost accounting
├── audit.py             # AuditLogger — immutable event log
├── adoption_tracker.py  # AdoptionTracker — wave rollout
├── mcp_connectors.py    # MCPConnectorRegistry — connector catalogue
└── cli.py               # `ecl` command — management CLI

examples/
├── basic_governed_agent.py     # ← start here
├── clinical_trial_agent.py     # GxP + audit
└── sdlc_accelerator.py         # wave simulation (offline)

tests/                          # 66 pytest tests, all passing

Contributing

Pull requests are welcome. Please:

1. Fork and create a feature branch
2. Write tests — the project maintains 100% coverage on core paths
3. Pass the quality gate: pytest && mypy enterprise_claude/ && ruff check .
4. Open a PR — describe the problem you're solving and link any related issues

For significant changes, open an issue first to discuss the design.

Development setup

pip install -e ".[dev]"
pre-commit install   # runs ruff + mypy on every commit

License

MIT — use freely in commercial products.