FAQ
Common questions, answered precisely.
Does Spineforge add latency to my agent calls?
No. Spineforge is never in the data path. Your agent calls LLM and tool providers directly. Spineforge observes via the OTel span pipeline after the fact, asynchronously. The APISink batches and sends telemetry in a background thread — zero impact on your agent's response time.
What happens if the Spineforge backend is unreachable?
Agents keep running. The APISink falls back to the local FileSink and buffers events to .spineforge/logs/actions.jsonl. Once connectivity is restored, you can replay buffered events. Nothing is lost.
How is revocation different from an instant kill-switch?
Revoking an agent in Spineforge blocks all future credential re-leases immediately. However, a lease copy already in the agent's memory remains valid until its TTL expires. Most LLM providers don't support natively short-lived scoped keys, so there is no mechanism to invalidate an already-issued copy. To force immediate access loss, rotate the underlying provider key in your Spineforge vault and in the provider's dashboard.
Why is cost stored and not recomputed, but latency is live?
Cost is a fact about a past event — it happened at a specific rate, and that rate should be frozen. A run that cost $0.03 three months ago should still show $0.03 today, even if rates have changed since. Latency, by contrast, is a property of a window and a dataset — adding a new action today meaningfully changes the P95. These have different semantics, so we treat them differently.
Do I have to change my agent code to use Spineforge?
Minimal changes. The init() call and the spine.run() context manager are required. If you're using LangChain, LangGraph, CrewAI, or AutoGen, that's it — no other changes. If you're using the raw OpenAI SDK with custom tool functions, add @track_tool to those functions.
Why is there a migration requirement for credential leasing?
Credential custody only works if raw provider keys exist exclusively in Spineforge. If a key also lives in an env file or shared vault, any process that can read that env file can use the key — registered or not. Removing keys from other locations is not optional fine print; it's the step that makes the security guarantee hold.
Can I use Spineforge without the registry backend (offline mode)?
Yes. In offline mode (no SPINEFORGE_REGISTRY_URL), identity is stored in .spineforge/registry.json and telemetry goes to .spineforge/logs/actions.jsonl. No data leaves your machine. Credential leasing and the backend dashboard are unavailable in offline mode.
Is Spineforge open source?
Not currently. The SDK is not open source. The architecture, roadmap, metric calculations, and pricing formula are published openly because you should be able to understand exactly what you're running and paying for.
Does Spineforge support multi-tenant isolation?
Spine IDs are scoped to your organization. Agents in different organizations cannot lease each other's credentials or access each other's telemetry. Agent-to-agent delegation is org-scoped by design.
What frameworks does Spineforge support today?
LangChain, LangGraph, CrewAI, AutoGen, and the raw OpenAI SDK — all auto-instrumented via OpenTelemetry. Node.js SDK is on the roadmap. From-scratch framework support (building a new agent framework) is explicitly out of scope.
How do I trace an action's cost back to the rate that produced it?
Every action row stores rate_card_id. You can join actions to rate_cards on that field to see the exact provider, model, and cost-per-token that produced any cost figure. This is queryable from the database directly or visible in the action detail view in the dashboard.
Is there a free plan?
Yes. The Free tier supports up to 3 agents and 10,000 actions per month with no payment method required. Credential brokering and the backend dashboard are not available on Free — they require the Startup tier ($49/mo) or higher.
Don't see your question? Contact us — we'll answer and add it here.