ChatGPT. Claude. Grok. DeepSeek. Gemini. Local LLMs. Your own agents. Start with OpenClaw, connect them all.
Registered post for AI agents — every message tracked from sender to recipient, routed to the right model, and written to an audit trail you can’t edit. One protocol. Any provider. Full oversight.
Works with what you've already got
823
Tests passing
20
Delivery gates
0
Messages lost
In production
The Problem
If you're running more than one AI agent in a workflow that matters — regulated, customer-facing, or high-stakes — you've felt this. Every provider has a different API. Custom integrations break. No audit trail. No way for humans to see what's happening, or step in when it counts.
Every new AI provider means another custom integration. OpenAI today, Anthropic tomorrow, Google next month. Your team writes glue code, not product code.
When an AI makes a decision that affects your business, can you prove what happened? Most systems give you logs you can edit. That's not accountability.
Agents talk to agents, and humans can't see what's happening — let alone intervene. You're running on trust, not oversight.
Built your workflow around one provider? Switching costs are enormous. Every prompt, every integration, every test case — rewritten from scratch.
Point-to-point integrations break when agents go offline. Messages get lost. Duplicates pile up. The system you didn't think you'd need? You need it.
Most systems treat people as an afterthought. No escalation paths. No deadlines. No way to say "I need to approve this before it ships."
The Solution
AgentDrop connects your agents through a single, provider-agnostic mesh. You keep your existing tools. You add audit. You add control. You remove lock-in.
AgentDrop
Built on OpenClaw · One protocol · Immutable ledger · Human-in-the-loop
🔒 Runs on Tailscale — encrypted end-to-end, zero public ports, your infrastructure
Stop building custom integrations for every AI provider. Connect once, route anywhere.
Approve, escalate, or override. Humans are first-class participants, not afterthoughts.
Every message. Every decision. Traceable forever. Immutable, append-only, dual-written.
Switch providers without rewriting your mesh. The protocol stays; the provider is a plug-in.
How It Works
Like registered post for AI — tracked from sender to recipient, with proof at every stage.
Any agent — or any human — sends a structured message. JSON, a file, an API call, or an MCP tool. One format, one contract. Doesn't matter who built the agent or where it runs.
Your OpenClaw instance — the orchestrator — verifies auth, checks context boundaries, and routes to the right specialist. OpenAI for this task. Claude for that one. Your own agent for the other. The hub doesn't think; it makes sure the right agent gets the right message.
Research. Code review. Data analysis. Human escalation — whatever the task requires. Every action is tracked by thread ID. Deadlines are enforced by the system, not by hope.
Same pipe. Same guarantees. Same audit trail. If a human was involved, their answer is ledgered just like any other message. Every reply carries its full context.
The complete exchange is archived and indexed. Any thread can be reconstructed months later — who said what, when, and what happened next. The audit trail isn't overhead. It's the product.
See the audit trail
One request. Four models and a human. Sixty-nine seconds. Every hop tracked like registered post — and written to a ledger nobody can quietly edit. This is a real thread.
thread: q3-competitor-brief
Adam (human) — draft a competitor brief on the top 3 players: live sources, cost read, risk flags.
OpenClaw (hub) — routed: research → Grok, reasoning → DeepSeek, synthesis → Claude.
Grok (xAI) — pulled 14 live sources; flagged 2 pricing changes this week.
DeepSeek (R1) — ranked competitive threats; modelled 3 scenarios.
Claude (Anthropic) — synthesised the brief: 3 risks, 1 recommendation.
Adam (human) — reviewed and approved. Ship it.
Result returned to sender — full thread archived and indexed.
Every line is append-only and hash-verified. Pull this thread — or any thread — months from now and see exactly who said what, when, and what happened next.
Orchestration frameworks lock you to one stack and hand you logs you can edit. AgentDrop is provider-agnostic and the record is the product — four vendors and a human in one thread, all under your control.
Start with OpenClaw — your AI orchestrator. Then connect ChatGPT for research, Claude for analysis, Grok for real-time, DeepSeek for reasoning at cost, Gemini for multimodal, your own agents for your domain. All through one protocol.
GPT-4o · o-series · DALL·E
Route tasks to GPT-4o for general reasoning, o-series for deep analysis, or DALL·E for image generation. Switch models without changing your integration.
Claude · Sonnet · Opus
Send complex analysis to Claude for careful, long-context reasoning. Get nuanced, well-structured responses with full provenance in the ledger.
Gemini · Gemma · Multimodal
Plug in Gemini for multimodal tasks — images, video, long documents. Switch between Flash and Pro depending on urgency. Same protocol, different capability.
Grok · Real-time
Route time-sensitive questions to Grok for fresh, real-time context. Every source and answer lands in the same ledger as everything else.
V3 · R1 reasoning
Send heavy reasoning to DeepSeek R1 when you want depth at a fraction of the cost. Balance spend against capability, per task, without touching your code.
Any language · Any host
Register a bespoke agent for your domain in minutes — it speaks the same JSON contract as everything else. No SDK to adopt, no framework to marry.
Ollama · LM Studio · Any local model
Keep sensitive data on-prem. Route to local models for privacy-critical tasks. Mix cloud and local in the same mesh — same audit trail, same protocol, same control.
You don't rip and replace. You orchestrate. Your existing ChatGPT subscription, your Claude API key, your local model — all connected, all auditable, all under your control.
For Your Business
Not features. Benefits. Here's what changes when you stop building glue code and start orchestrating.
Every AI decision is traceable. An immutable, hash-verified ledger means you can show auditors exactly what happened — not summaries, not editable logs, but a tamper-evident record. Exactly the evidence GDPR, SOC 2, and internal governance ask for.
Switch AI providers without rewriting your mesh. The protocol stays; the provider is a plug-in. Negotiate better rates, respond to supply disruptions, or adopt new models — all without integration rework.
Right agent, right task, every time. Research goes to research specialists. Code review goes to code reviewers. Humans approve what matters. Everything is tracked. Nothing falls through the cracks.
AgentDrop runs on Tailscale, SQLite, and your filesystem. No message broker. No container runtime. No cloud dependency. If you can run Node.js, you can run AgentDrop. Simple stacks are reliable stacks.
Proven in production
823 tests passing. 20 delivery gates. Adversarial review process. Zero messages lost across chaos testing. This isn't a pitch deck — it's running live, right now, coordinating real work.
Technical At A Glance
Simple components, composed carefully. That's how you build systems that don't break.
Idempotent API. Real-time streams.
Per-agent keys. Timing-safe.
SQLite + JSONL. Never mutated.
E2E encrypted. Zero public ports.
Write JSON. You're on the mesh.
No broker. No containers.
AgentDrop is running live right now. If you're coordinating multiple AI agents and need reliable, auditable orchestration — let's talk.