The product
The control surface for AI agents.
Silvanexum is built for people who run fleets of agents in production — not a chatbot directory. Four primitives turn agent spend from an opaque blended bill into something you can attribute, cap, and audit at a glance.
Attribute
Spend, per agent and customer.
Govern
Budgets and policies that cap.
Trace
Every hop, with its cost.
Mesh
Attribution across agent calls.
Agents are cost centers, not endpoints.
Each agent gets a stable handle, a version history, and capability badges — the durable entity you attribute spend and authority to. When you publish, others can discover it; when you iterate, the profile and its track record carry forward instead of forking your memory of it.
- Stable handles like @org/agent, versioned over time.
- Spend and policy attach to the identity, not a loose key.
- Public or private — you decide what the network sees.
Reliability you can audit, not adjectives.
Each agent's track record is a calibrated blend of success rate, ratings, and run volume — a supporting signal on a fail → warn → ok ramp, so a glance tells you whether to trust an agent with a budget. No agent rates itself; reliability is earned one run at a time.
- 0–100 score with a transparent fail / warn / ok ramp.
- Five-star ratings from the people who actually ran it.
- Volume matters: a 99% over 12k runs beats 100% over 3.
The flight recorder for every dollar.
When runAgent executes, it emits a structured, multi-hop trace: each model turn, tool call, and sub-agent hand-off becomes a node with its own latency, token count, and cost — tagged with who called it on whose behalf. Cost attribution stops being archaeology — you read the timeline.
- Every hop captured, with cost rolled up — not just the answer.
- Per-step timing, tokens, and spend, aligned in tabular mono.
- Unauthorized or runaway calls render inline at the step.
- planner · gpt-4o-minimodel0.42s180+96 tok
- web.searchtool0.88s
- vector.recalltool0.21s
- fact-checkeragent1.10s
- compose · claude-sonnetmodel0.96s1.2k+420 tok
Attribution that survives every hop.
Agents call other agents, and good ones get forked. Silvanexum keeps both legible — the agent-to-agent spend graph at run time, and fork lineage over time — so cost and authority stay attributable across the whole mesh, not lost when one agent hands off to the next.
- Agent-to-agent (A2A) calls captured in the trace.
- Cost and delegation attributed up the chain.
- Fork lineage rendered as a navigable graph.
Your keys. Your tools. No middleman.
Silvanexum runs on the providers and tools you already use. Keys are encrypted with AES-256-GCM, validated the moment you paste them, and decrypted only server-side at run time — we never mark up your tokens.
Bring your own model keys
Validated on paste with a free probe, then locked away. A green dot means an agent can run; a red one tells you before it fails.
Connect the tools agents need
Authorize a connector once and it becomes a runtime tool any of your agents can call — merged automatically with agent-to-agent tools.
Collaboration runtime
A tool-calling loop with retries lets agents plan, call tools, and hand off to other agents — all recorded.
Durable memory
Agent memories are stored as vectors, so context persists across runs instead of evaporating with the session.
Readouts, not dashboards
Latency, success, and volume render like a control panel — legible at a glance.
Put your agent spend under control.
Attribute every dollar, cap runaway costs, and audit every call — on your own keys, no markup. Start free.