Autonomous model routing for agents
Buyer agents discover verified model capabilities. Provider nodes publish GPT, Claude, DeepSeek, Kimi, GLM, Qwen or local models. Crucible routes by price, latency, policy and trust.
Buyer Agent
Sends intent and policy
Crucible Router
Price / latency / trust
Provider Node
Cross-region model capability
Receipt
Metered audit record
# Any OpenAI-compatible agent changes one base_url
export OPENAI_BASE_URL=
"https://api.a2a-crucible.xyz/v1"
model =
"best:coding.high_reasoning"
returns route + receipt
Routing policy
quality-first
Candidates
provider nodes
Receipt
rcpt_*
Failover
< 30s
Provider nodes can expose these model capabilities
Model families
Routing signals
Entry points
Failover target
Three Autonomous Loops
Buyer agents discover, provider nodes list, and the platform routes and supervises
Intent-Based Discovery
Buyer agents ask for capabilities, quality, latency, privacy and budget instead of choosing a fixed vendor
{
"capability": "coding.high_reasoning",
"max_price": 0.05,
"privacy": "no_retention"
}Autonomous Listing
Provider nodes publish model families, price policy, current load, health and service constraints
{
"models": ["claude", "deepseek"],
"load": 0.42,
"region": "asia"
}Policy-Aware Routing
Crucible matches requests to nodes by price, latency, region, trust score and failure history
{
"route": "node_8f2a",
"score": 0.94,
"fallbacks": 2
}Meter and Score
Every invocation produces usage receipts, audit logs, health updates and quality feedback
{
"receipt": "rcpt_2dbb9a53",
"tokens": 1842,
"latency_ms": 912
}A2A-Crucible Core Protocol
intent → discover → route → invoke → meter → score
Agents choose capabilities, not vendors
Crucible abstracts model providers into routable capabilities and selects provider nodes by task type, price, latency, region and trust score.
High-Reasoning Coding
best:coding.high_reasoning
For review, debugging, architecture and hard refactors.
Low-Cost Reasoning
cheap:reasoning
For batch analysis, agent background work and retryable tasks.
Long-Context Chinese
best:zh.long_context
For Chinese documents, knowledge work and large context windows.
Local Private Node
private:local
For internal agents that need data locality and no-retention policy.
Price
budget caps
Latency
SLA targets
Region
cross-border routes
Trust
health and receipts
client.invoke("best:coding.high_reasoning", prompt, policy={"max_price": 0.05, "privacy": "no_retention"})Turn idle model capacity into agent-discoverable nodes
The early goal is not payout promises. It is a small, stable Provider Node network with observable models, regions, price, heartbeat, readiness and receipts.
3-5
Seed nodes
15s
Heartbeat
rcpt_*
Receipts
CRUCIBLE_AUTO_REGISTER=true \
CRUCIBLE_API_BASE_URL=https://api.a2a-crucible.xyz/api \
CRUCIBLE_NODE_PUBLIC_URL=https://your-node.example.com \
PROVIDER_NODE_MODELS=best:coding.high_reasoning,cheap:reasoning \
node scripts/provider-node.mjsConfigure upstream key or local model
Auto-register and report heartbeat
Enter cross-region routing candidates
Expose upstream or local models
Provider Node can relay OpenAI-compatible upstreams or wrap Ollama, vLLM and custom services.
Auto registration and heartbeat
Nodes list capabilities on startup and report latency, success rate, queue and memory signals.
Signed invocation endpoint
Optional HMAC verification protects Provider Node capacity from non-platform calls.
Receipts build trust
Every routed invocation creates a receipt for trust score, audit and future billing plugins.
Discovery, routing, metering and trust first
Payments can arrive later as integrations. The core product is reliable discovery, invocation and evaluation of model provider nodes.
OpenAI-compatible Gateway
Connect existing agents through a base_url change
Agent Cards
Describe provider node models, capabilities and policies
Heartbeat
Live health checks, load reporting and delisting
Routing Engine
Route by price, latency, region and trust score
Metering Receipts
Record invocations, usage, latency, cost and audit trails
Trust Score
Turn success rate, SLA, feedback and failures into scores