OpenRouter routes.
RelayPlane observes,
governs, and learns.
OpenClaw is Exploding
155K+ GitHub stars. Agents that write code, debug systems, and ship features autonomously. A single session can make 100+ API calls — and you have no idea what's happening inside.
The problem isn't routing. That's solved.
The problem is everything else:
OpenRouter solved routing. RelayPlane solves ops.
Three Pillars. One Platform.
Learning Ledger
Captures everything. Every run, every decision, every outcome.
Query your runs. Export for audit. Replay decisions. All local, all queryable.
┌─────────────────────────────────────────────────────────────┐ │ RUN #a3f7c │ ├─────────────────────────────────────────────────────────────┤ │ Provider anthropic │ │ Model claude-opus-4-5 │ │ Auth api_key │ │ Tokens 12,847 in / 2,341 out │ │ Cost $0.47 │ │ Latency 1,203ms (TTFT: 412ms) │ │ Tools read_file, write_file, exec │ │ Routing capability:tool_use → opus (fallback: sonnet) │ │ Outcome ✓ completed │ └─────────────────────────────────────────────────────────────┘
Policy Engine
Enforces the rules before requests hit providers.
- •Budget caps — Per-run, per-day, per-agent, per-workspace
- •Model restrictions — Allowlists, denylists, approval gates
- •Tool restrictions — Control what agents can do
- •Context caps — Limit token usage per request
Policies evaluated. Violations blocked. Everything logged.
# relayplane.policies.yaml
- name: "Daily Budget Cap"
type: budget.per_day
limit_usd: 50
action: deny
- name: "Opus Approval Gate"
type: model.allowlist
models: [claude-opus-*]
action: require_approval
approvers: [eng-leads]
- name: "No External Tools in CI"
type: tool.denylist
scope: { agent_tag: "ci-*" }
tools: [browser, web_fetch]
action: denyPattern Detection
Turns run history into actionable intelligence.
- •Detect failure patterns before they cost you
- •Surface optimization opportunities automatically
- •Promote successful configurations across your team
- •Keep the learning local — your patterns, your data
┌─────────────────────────────────────────────────────────────┐ │ 💡 SUGGESTED RULE │ ├─────────────────────────────────────────────────────────────┤ │ │ │ "Code review tasks succeed 94% with Sonnet, 96% with Opus. │ │ Cost difference: 10x. Suggest: route code_review → sonnet" │ │ │ │ Evidence: 847 runs analyzed │ │ Confidence: 0.91 │ │ │ │ [Approve] [Reject] [Review Runs] │ └─────────────────────────────────────────────────────────────┘
Flexible Auth. Your Choice.
RelayPlane doesn't dictate how you work. It adapts to how you already work.
Safe Defaults. Override When You Know Better.
// Default: recommended mode (safe) workspace.auth_enforcement_mode = "recommended" // Power users: permissive mode workspace.auth_enforcement_mode = "permissive" // ⚠️ Consumer auth now allowed for automation // All overrides logged with auth_risk=true
Sub-Agents with Different Profiles
Need your main agent on Max but sub-agents on API keys? That's the pattern.
Human Session (Max) ──spawns──▶ Sub-Agent (API Key)
│ │
└── interactive run └── background run
(consumer auth) (API auth)Each run = one agent = one auth profile. No mixing. Full traceability.
The Flow
Every request passes through four stages. Every decision is logged.
Your Agent
OpenClaw • Claude Code • Custom
Auth Gate
Validate credentials
Policy Engine
Enforce rules
Router
Select provider
Learning
Detect patterns
Learning Ledger
Every run recorded
PROVIDERS
Anthropic
Claude
OpenAI
GPT-4
OpenRouter
100+ LLMs
Local
Ollama
Every decision logged. Every run explainable. Every pattern learnable.
Roadmap
We're building in public. Here's what's ahead.
Join the waitlist to get early access as each phase ships.
Join the Waitlist
RelayPlane is coming soon. Local-first. Open source. Your data stays yours.
Coming soonnpx clawhub@latest install relayplane