The accountability layer for AI

Get your
under control.

Gate cuts your spending, keeps you protected, and lets you monitor every request. Keep using the AI tools and subscriptions your team already pays for.

16 public benchmarks · arXiv:2606.02959

  • 95.4% of prompt injection attacks caught
  • Any harness ImageImage
  • Any model provider
  • Keep your existing subscriptions
  • Cut your AI spend 20%
  • Verifiable audit trail
01 / SECURITYRemove threats
02 / COSTSave tokens
03 / OBSERVABILITYAudit everything
Between your agent and the model Gate removes a threat, reduces token usage by twenty percent, and creates a verified record with cost and latency. AGENT REDACTED −20.6% 6,500 → 5,162 TOKENS $0.082 742 MS VERIFIED MODEL

Between your agent and the model

01 / SECURITY Remove threats REDACTED
02 / COST Cut your AI spend 20% 6,500 5,162 TOKENS −1,338
03 / OBSERVABILITY Audit everything $0.082 742 MS VERIFIED

01 / Security

Attacks arrive as text.
Gate catches them in transit.

Gate screens every request on the way in and every response on the way out.

  • OWASP LLM #1

    Prompt injection

    One line of attacker text can override your instructions, exfiltrate data, or call a tool.

    Blocked before the model sees it.

  • Secrets · PII

    Credential and PII leakage

    A customer's SSN, an API key, an internal note. One response is enough to leak it.

    Redacted before the response returns.

  • Indirect injection

    Hijacked tool calls

    Whatever a tool hands back becomes your agent's next set of orders.

    Stopped before your agent acts.

See how Gate blocks attacks
Prompt-injection detection Gate's rank on each public benchmark
DATASET#1#2#3#4#5#6#7#8FIELDDEEPSETOF 28SAFEGUARDOF 13WILDGUARD-BENIGNOF 10GENTEL-GOAL-HIJACKINGOF 9GENTEL-PROMPT-LEAKINGOF 9LLMAILOF 7SALAD-DATAOF 6ADVBENCHOF 3GANDALF-IGNOREOF 24GENTEL-JAILBREAKINGOF 9JACKHHAOOF 6ILION-BENCHOF 5NOTINJECTOF 10HARMBENCHOF 12BIPIAOF 16HACKAPROMPTEVALUATED — NO PUBLISHED THIRD-PARTY COMPARATORRANK #18/16DATASETS#2 ON 3 · #3 ON 1ONE GLOBAL THRESHOLDPOOLED FPR 1.0%
#1 ON 8 OF 16 #2 ON 3 · #3 ON 1

16 public datasets · one global threshold · 1.0% pooled FPR arXiv:2606.02959

Measured against Lakera Guard, the incumbent: 96.6% average F1 to their 83.7%, with Gate re-tuned to the false-positive rate each dataset published.

Read the benchmark report

02 / Cost

Same prompt, same answer,
20%+ fewer tokens.

Gate processes each request on its way to the provider, removing token waste and collapsing inefficient tool calls and duplicate content. The model receives the same content and returns the same answer. You are billed for less of it.

One request before and after Gate compression A 6,500 token request goes upstream as 5,162 tokens, a reduction of 1,338 tokens or 20.6 percent. The removed span is mostly content that was repeated in the original request and is now sent once. REPEATED CONTENT — NOW SENT ONCE BEFORE 6,500 TOK AFTER GATE 5,162 TOK −1,338 TOK · −20.6%
Example request · POST /v1/messages Duplicate content · Repeated reads · Tool-call overhead · Lossless
Before
6,500 tok
After Gate
5,162 tok
Saved
−1,338 tok · −20.6%
Transforms
Duplicate content · repeated reads · tool-call overhead · lossless

Request compression

Lossless and cache-aware. Gate drops what the model does not need to read twice: repeated file reads, duplicated context, envelope noise. What reaches the model is the same content.

20%+ fewer tokens per request

Repeat-request caching

When the exact same request comes in again, Gate returns the stored response instead of calling the provider. Agentic systems repeat themselves constantly: the same file read, the same retry, the same tool call around a loop. A hit costs nothing at all.

100% of tokens saved on a cache hit

Free tier, no card. Point one tool at Gate and watch the first bill fall.

Get started → See pricing

03 / Observability

See what your agents are doing.
Cap what they can spend.

Gate is already in the path of every request, so the record writes itself.

Workspace activity Live · 4 agents

Month to date $182.40 of $250.00

Example activity: five events with the time, what happened, the model, tokens billed and cost.
Time Activity Model Tokens Cost
15:42:08 Prompt screened, clean claude-sonnet-4-6 1,204 $0.011
15:42:11 Injection blocked · tool output — 0 $0.000
15:42:14 Response returned · 2 secrets redacted claude-sonnet-4-6 890 $0.014
15:44:02 Cache hit · response reused claude-sonnet-4-6 0 $0.000
15:51:30 Budget reached · dev-agents paused — — —
Sealed · batch #12,402 24 proofs verified, 0 failed

One place to look

Every request from every agent and every tool, filterable by agent, by key, or by what happened. No SDK, no traces to wire up.

Limits that hold

Spend caps and rate limits per key, per agent, or per workspace, enforced on the traffic. A runaway loop stops at your number, not on next month's invoice.

Proof you can hand over

Each entry is hashed, chained, and anchored to Constellation Digital Evidence. An auditor can check a record against the chain without going through Gate.

Audit an OpenClaw run Digital Evidence

04 / Setup

Keep the tools and subscriptions
you already pay for.

Gate goes in front of the AI tools your team already uses. Keep the Claude Code and Codex subscriptions you already pay for and run them through Gate, or keep your own provider keys. Either way the model side keeps billing you the way it does today.

The Gate Connect routing screen: 'Route through Gate' switched on, with Claude Code, Codex and OpenRouter enabled and Gemini off
Gate Connect · one toggle per tool

Install Gate Connect, sign in once, and it finds the tools you already run and routes each one through Gate. One click per tool. Most setups finish in under a minute, with no config files to edit.

  • Image Claude Code
  • Image Codex
  • Image OpenClaw
  • Image OpenCode
See how Gate Connect works

Prefer the command line?

Point your existing SDK at Gate. Same code, same provider key, same response shape. Every request after that is screened and recorded. 

# the only line that changes
client = OpenAI(
  base_url="https://gate.constellationgate.ai/v1",
  api_key=os.getenv("GATE_API_KEY"),
)

Pay as you go

No accounts to open. One endpoint, every model.

Top up a balance with Gate and run frontier and open-weight models side by side. You pay the same price you would pay the provider directly.

  • Image Claude Opus 4.7
  • Image DeepSeek V3.2
  • GPT-5
  • Image Qwen3 Max
  • Image Gemini 2.5 Pro
  • Image Llama 4 405B
  • Image Grok 4
  • Image Mistral Large 3

Routed through Anthropic, OpenAI, Gemini, Bedrock, OpenRouter and Alibaba Cloud. Plus hundreds more models from the same endpoint.

Learn about pricing

Start now

Put Gate in front of your agents today.

Attacks blocked, tokens saved, and a full record of what your agents did. Connect the tools you already use in a couple of clicks.

Get started → View pricing

Free tier included · No credit card required