GeminiClaw: The Local-First Multi-Agent Orchestrator We Built to Operate in the Real World
No more chat-only assistants. GeminiClaw is the AI operations platform Autenticare built from scratch: multi-agent orchestration, autonomous execution with guardrails, native Telegram and Google Ads integration, and a full control panel. All local-first, auditable, and with real governance.
Fabiano Brito
CEO & Founder
Most "AI agents" you see are, in practice, glorified chatbots: they receive a message, call an LLM, return text. GeminiClaw was built to do the opposite — execute real operations, with traceability, security, and configurable autonomy. It's the project born inside Autenticare that now has its own address: geminiclaw.dev.
Why build yet another agent framework?
It's a fair question. There are dozens of frameworks available — LangChain, LlamaIndex, CrewAI, AutoGen. The short answer: none of them were good enough to operate in production without constant supervision.
Recurring issues we identified in real projects:
- No Loop Guard: agents entered infinite loops or long sequences without a controlled stop point.
- Excessive permissions: a "read-only" agent ended up with write tools through configuration inheritance.
- Memory without boundaries: contexts from different sessions mixed together, generating incoherent responses or data leaks.
- No audit surface: when something went wrong, there was no way to reconstruct what the agent did and why.
GeminiClaw was built on top of Building Applications with AI Agents (Michael Albada, O'Reilly) — one of the most technically solid sources on agent engineering — and then adapted for real operational needs found in the field.
The architecture in terms that matter for operations
| Concept | What it means in GeminiClaw |
|---|---|
| Platform | The complete product: Studio, runtime, channels, governance. |
| Runtime | The execution layer that manages loops, approvals, replay, and tool routing. |
| Agent | Cognitive/operational entity hosted by the runtime, with isolated memory and tools. |
| Skill | First-class capability pack: metadata, dependencies, policy, and reusable instructions. |
| Tool | Executable capability with an explicit contract — the real execution primitive. |
| Channel | Delivery surface: Telegram, CLI, WhatsApp — where agents receive and deliver events. |
| Control Plane | The administrative and observability surface for operating the system. |
What GeminiClaw does differently in practice
1. Multi-agent orchestration with clear ownership
GeminiClaw operates with a main orchestrator agent + specialized subordinate agents. The Request Ownership rule is explicit in the runtime: every task remains with the agent to whom the user addressed it. Handoffs work for scoped delegation — not for silent reassignment.
2. Autonomous execution with safety gates
- Planning Gate: no code mutation occurs without an approved plan.
- Risk Gate: high-impact tools require explicit approval before execution.
- Loop Guard: long operations have adjustable autonomy count to prevent infinite loops.
- Checkpoint/Replay: durable execution recovery — if something fails midway, the runtime can resume from the last checkpoint.
3. Real Google Ads B2B operations
Real API-level operations — not "generate ad copy":
- Campaign diagnostics and mutation via Google Ads API
- Batch optimization workflows (
optimize_campaign_batch) - Dayparting, budgets, RSA optimization, negatives, geo targeting
- Demographic guardrail automation
- Offline conversion management + lead quality feedback loop
4. GeminiClaw Studio — the dashboard that was missing
- Board: operational Kanban for real-time task management and agent monitoring.
- Unified Timeline: task-level audit, tool call history, and message traces.
- Memory Governance: inspection, expiration, redaction, and purge of structured and semantic memories — with explicit audit trail.
- Governance Inbox: memory governance recommendations grouped by urgency, with direct actions.
Real risks and where GeminiClaw still has friction
- Configuration curve: the local runtime requires Node 20+, Telegram bot token, Gemini API key, and properly configured environment variables.
- Token cost in long operations: autonomous loops consume context quickly. Loop Guard helps, but cost still requires monitoring.
- MCP integration: MCP is supported as an integration layer, but A2A is still future work.
- Proprietary license: the current repository is proprietary/internal. Public access is on the maturity roadmap.
Start with GeminiClaw
The official site with full documentation is at geminiclaw.dev. Prerequisites: Node.js 20+, Telegram bot token, Gemini API key. Start command: npm run watchdog.
If you're building agent systems for production and want to see what governance, observability, and real execution look like when designed together from the start — start with GeminiClaw.