Open-Source Agent Workspaces: How Paperclip and ENZO Let You Run Agents on Your Own Keys and Infrastructure

I'm Denis Shokhirev — Agentic AI Systems Architect in Freiburg, Germany, running DennisCraft AI Studio. My production stack: Claude, Supabase, n8n, Doppler, and self-hosted Postgres. Last quarter, a fintech client hit a wall: they needed to deploy production-grade agents on their own API keys and infrastructure — no cloud vendors, no demo sandboxes, full auditability.

Why Open-Source Agent Workspaces Matter

Most open-source agent platforms either lock you into their cloud or top out at toy use cases. In regulated markets like Germany and Austria, running agents in public sandboxes is a non-starter: GDPR, internal data, and key management are mandatory. The real challenge isn't LLM selection — it's integration, access control, audit, and rotating secrets on your own stack.

In my experience, 80% of the engineering effort goes into infrastructure, not model tuning. Enterprise teams need open workspaces where agents run on their own keys, with full isolation, audit trails, and CI/CD integration — not just playground demos.

Paperclip and ENZO: Open Tools Proven in Production

After evaluating dozens of agent workspace projects, the only stack I’ve seen work in production — not just in notebooks — combines Paperclip and ENZO:

• Paperclip: a lightweight, open-source workspace for LLM agents. Supports custom API keys, self-hosted deployment, Supabase/Postgres integration, and direct CI/CD hooks. Lets teams run agent experiments securely and reproducibly.
• ENZO: an open-source framework for orchestrating multi-agent systems. Lets you define agent roles, entrypoints, execution pipelines, and permissions. Works with both local and cloud LLMs (Claude, OpenAI, Hugging Face, Anthropic SDK).

Both are community-supported, open, and actually ship in production for DACH B2B teams.

Running Agents on Your Own Keys: Patterns and Risks

API keys are the soft underbelly of any agent deployment. The OWASP Top 10 for LLM Apps (2023) highlights key leakage and prompt injection as major risks (OWASP, 2023). In practice, I’ve seen the same SQL-injection patterns show up in agent-generated code on three recent launches.

• Secret storage: Use Doppler or Vault. Never store keys in .env or source code.
• Key rotation & audit: Automate via n8n pipelines; log every agent action to Supabase audit tables.
• Static analysis: Run semgrep and bandit on all agent pipelines. 100% of my production agents are statically scanned before deploy.

import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

def get_agent_response(prompt):
    response = client.completions.create(
        model="claude-3-opus-20240229",
        prompt=prompt,
        max_tokens=1024
    )
    return response.completion

# API key is pulled securely from Doppler or environment, never hardcoded

Integrating Workspaces into Enterprise Infrastructure

CI/CD Pipeline Integration

Paperclip drops into GitHub Actions or GitLab CI. You can trigger agent runs on every main branch push, auto-test, and deploy to staging or prod — all self-hosted, no cloud roundtrips. ENZO lets you orchestrate multi-agent pipelines, so one agent can review or gate another's output.

name: agent-deploy
on: [push]
jobs:
  build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run Paperclip agent
        run: |
          docker-compose up -d paperclip
      - name: Run ENZO orchestrator
        run: |
          docker-compose up -d enzo

Supabase/Postgres for Logging and Audit

Supabase provides a full audit trail and role-based access. Every agent invocation is logged at the DB level, making incident response and compliance reporting feasible for regulated markets.

Limitations and Security Considerations

• Even self-hosted, LLM API calls can leak data if you’re not running local models.
• Paperclip is not a panacea: for sensitive data, use additional isolation (Docker, k8s), especially when agents touch customer records.
• ENZO does not natively support GDPR, but lets you build permissioned workflows and explicit audit trails.

On my recent launches, static analysis and database audit logs catch more issues than runtime sandboxes. Security is a process, not a checkbox.

FAQ

Can I fully isolate agents in a self-hosted setup?

Yes, especially if you run local LLMs (e.g., Llama.cpp). But any external API call remains a potential leak point.

Are there mature enterprise agent workspace solutions?

Most proprietary offerings are either closed-source or don’t support full key/audit control. The Paperclip + ENZO stack is more flexible for B2B teams.

How do I audit agent actions?

Log every action to Supabase/Postgres; use audit tables to detect anomalies or abuse patterns.

How do I trigger agents from n8n workflows?

Build pipelines where the agent is a workflow step, passing keys securely via n8n’s credentials store.

Is it GDPR-safe to use Claude via API in Germany?

Only if keys are not cloud-stored and all calls happen within your controlled infrastructure (see Anthropic docs: Anthropic, 2024).

Where in your LLM pipeline do you catch the most production bugs: key management, audit, or at runtime? I’d genuinely like to know.

I run a free 30-min stack audit for DACH founders building AI in regulated markets. DM me on LinkedIn or write to @ger_dennis_ai.