Prove what your agents did, spent, and accessed — before your customers ask
KeyForge is the security control plane for production AI agents. Give them vk_ virtual keys instead of raw credentials, cap spend before it reaches the provider, and seal every request into a tamper-evident audit chain — deployable inside your own infrastructure.
One virtual key, every major model provider
0
Raw provider keys exposed to agents
0-bit
AES-GCM encryption at rest
0
AI providers unified behind one key
0%
Requests sealed in the HMAC audit chain
Sign up and cut your token bill, no code changes
Security is only half the story. KeyForge sits between your agents and every provider, so it can spend less of your budget on the same work: routing cheap prompts to cheap models, reusing deterministic answers, and capping how far any key can escalate.
Cost-aware Auto Router
Send model:"auto" and KeyForge scores each prompt's complexity, then routes trivial calls to the cheapest capable model and only escalates to a frontier model when the work genuinely needs it. Most agent traffic is simple, so stop paying GPT-4o prices for a yes/no classification.
- Deterministic complexity classification
- Cheapest capable model first
- Escalates only when the prompt demands it
Per-key cost-tier ceiling
Cap each virtual key to a Low, Medium or High cost band. A background or bulk agent physically cannot escalate to premium models: the router is clamped to the tier you set, so a single mislabeled prompt can never quietly 10x your bill.
- Low / Medium / High spend bands per key
- Hard ceiling on model escalation
- Pair with prevent-overrides to lock the model
Deterministic response cache
Agents repeat themselves constantly through retries, idempotent tool loops, and re-runs of the same deterministic prompt. Identical temperature-0 requests are served from a stored prior response instead of billing the provider again. Tenant-isolated and time-boxed, so you never read another account's answer or a stale one.
- Exact-match, deterministic-only (temp 0)
- Per-account isolation, 24h freshness window
- Zero provider charge on a cache hit
Free & cheapest-first routing
Choose a price-first provider sort and the router prefers free rate-limited and lowest-cost models within the tier before anything pricier. Combined with the Auto Router, a large share of everyday agent calls can land on a $0 or near-$0 model without you rewriting a line of agent code.
- Price-first provider sort preference
- Free rate-limited models routed first
- No agent code changes required
Savings depend on your traffic mix: how much of it is simple, repeated, or deterministic. KeyForge gives you the routing controls and the per-request cost metering to see it for yourself.
Four pillars of agent-safe infrastructure
Routing alone is not security. KeyForge is the layer between your agents and every provider they touch.
Unified LLM API
One OpenAI-compatible endpoint for every model and provider. Swap GPT for Claude or Gemini by changing a string; your agent code never changes.
- OpenAI-compatible /v1/chat/completions
- Streaming, JSON mode, tool calls
- Provider-agnostic model routing
Virtual Keys
Agents authenticate with vk_ keys. Real provider credentials stay encrypted in the vault and are injected server-side, so a compromised agent leaks nothing.
- Per-key quotas, spend caps, expiry
- Instant revocation, zero redeploys
- Real keys never touch agent code
Rate-Limit Resilience
When a provider returns 429, KeyForge auto-shuffles to a fresh key from your pool: same provider, same model, no dropped requests, no stalled agents.
- Key pools per provider
- Automatic 429 rotation
- Stay on the model you chose
Tamper-Evident Audit Chain
Every call is HMAC-signed and hash-chained to the previous entry. Any insertion, deletion, or edit breaks the chain: cryptographic proof of what your agents did.
- HMAC-SHA256 hash chaining
- One-click chain verification
- Shareable signed audit reports
What a payments company that owns the router can't give you
When the company that routes your AI traffic also bills the model vendors and processes the payments, neutrality, residency and hard spend limits become promises you can't check. KeyForge makes each one provable.
Provable routing neutrality
We hold no financial stake in any provider and earn nothing from your provider spend. When a router is owned by the company that also bills the model vendors, “best route” is unverifiable. Every KeyForge route can be minted as a signed neutrality certificate.
- No markup on provider spend
- No vendor billing relationships to steer routing
- Signed, auditable neutrality certificate per call
Enforceable data residency
Pin each key to the provider origins you allow and block the ones you don’t, including China-based models subject to foreign intelligence law. Requests to a disallowed origin are refused at the gateway, and you get a certificate proving where your traffic did and did not go.
- Per-key origin allowlist / blocklist
- China-origin models blocked by policy
- Residency certificate for auditors
Hard budget stops & portable proof
Per-key spend caps that actually halt spend, a progressive loop-guard that suspends runaway keys, a published SLA and status page, not Discord-only support, and an attestation ledger you can hand to compliance long after the call.
- Hard spend caps, not soft warnings
- Loop-guard suspends repeated-request storms
- Independently verifiable attestation ledger
Every request, cryptographically chained
Each audit entry carries an HMAC of its contents plus the previous entry’s hash. Tamper with one record and every hash after it breaks, instantly detectable.
The questions every serious buyer asks
How is KeyForge different from Portkey, Helicone, or LiteLLM?
Those are excellent tools for routing and observability. What they do not have is a cryptographically tamper-evident audit chain you can hand to an auditor with a single verify command, or bounded agent identity with pre-authorized payment mandates. KeyForge is a security control plane first: credential isolation, hard spend enforcement, and provable evidence are the foundation, not add-ons.
Could we just build this ourselves?
Some teams do. The simple part — virtual keys — takes a sprint. The hard part — HMAC chain integrity, concurrent spend-cap correctness, key-pool rotation, loop-guard, and PII redaction without false positives — takes months and a dedicated engineer, and still ships no agent-identity or payment-rails layer. KeyForge gives you all of it on day one.
OpenRouter has 400+ models. Why would I use KeyForge instead?
You should not choose between them — put KeyForge in front. Your agents authenticate with vk_ keys through KeyForge, and KeyForge forwards to OpenRouter as one of its providers. You keep every model and add hard spend caps, credential isolation, and a tamper-evident audit chain that a payments-owned router cannot credibly provide on its own.
We are not in a regulated industry — do we need audit trails?
Your customers might be. The moment you close a deal with a healthcare, fintech, or legal team, they will ask how you govern your agents. The audit chain is also your forensic tool when an agent does something unexpected. It is better to have evidence you never needed than to need evidence you do not have.
Does a gateway add latency to every request?
The credential resolution, spend check, and PII scan run inline and are engineered to be a small fraction of a model call that already takes hundreds of milliseconds to seconds. Same-provider key-pool rotation can even reduce total latency by avoiding 429 back-off stalls. You can measure the per-request overhead yourself with built-in cost and timing metering.
Can we deploy KeyForge inside our own infrastructure?
Yes. KeyForge is self-hostable in your own VPC or fully air-gapped, so secrets, audit evidence, and traffic never leave your boundary. Data residency can be pinned per key, and disallowed provider origins are refused at the gateway — with a certificate proving where your traffic did and did not go.
Stop handing agents your real keys
Free tier includes 3 virtual keys, 1,000 requests a month, and the full HMAC audit chain. No card required.