Every AI action, governed.
A policy enforcement proxy that sits between your AI agents and every tool they touch. Write deterministic rules in a human-readable DSL, hold high-risk actions for M-of-N human approval, and get a tamper-evident audit trail anyone can verify — no SDK required.
Forensic ledger — BLAKE3 chained
Works with any agent framework — keyless mode needs no SDK
See it in 30 seconds
An agent tries a risky action. Clausgap holds it for approval, records the decision in a chain anyone can verify, and takes four steps to set up.
The ungoverned AI problem
average cost of a data breach — IBM Cost of a Data Breach Report 2024
Everything you need to govern AI
Write rules, not code
Write governance rules in a human-readable DSL. Version-controlled and reviewable — and with approval required, no change reaches your agents until a second person approves that exact version.
policy "support_bot" {
allow call tool "web_search"
deny call tool "delete_records"
context_deny time_of_day outside "09:00-17:00"
}Skip the blank policy page
Turn on Discovery Mode and nothing is blocked — Clausgap just watches what each agent role actually calls, then drafts a policy from what it saw. Review the draft, edit it like any other policy, and turn enforcement on when you're ready. No DSL to learn before you get your first policy.
Discovery Mode is on
Hold high-risk actions for human approval
When an action is too risky, the proxy holds it and notifies your approvers — in Slack, Microsoft Teams, or email — with Approve and Deny buttons that open the request in your dashboard. Require M-of-N votes before release — any rejection immediately cancels. The approved action then runs once, exactly as approved, and only for the agent that asked. No polling, no SDK, no custom webhooks to build.
Action held for escrow
submit_order · $48,200 · procurement_bot
Prove every decision
Every decision is hashed and chained. Export a signed evidence package and your auditor can check it offline with clausgap-verify — no access to your Clausgap needed. Tamper-evident.
BLAKE3-chained · Ed25519-signed · exportable
One proxy, every protocol
HTTP, PostgreSQL, and MCP — one binary, one config, one policy engine. Multi-agent systems touching multiple data stores deploy a single container instead of three.
clausgap-proxy · single binary · shared policy engine
No SDK required
Deploy the proxy in front of any existing agent framework — LangChain, AutoGen, CrewAI, custom tool loops — with zero code changes. Clausgap resolves identity from JWT Bearer tokens, API keys, or source IP. Add the SDK later only if you need verified confidence scores.
keyless:
enabled: true
mode: "strict"
api_key_sources:
- header: "Authorization"
prefix: "Bearer "
static_roles:
- source_ip: "10.0.0.0/8"
agent_role: "internal-agent"Stop prompt injection at the gate
Detect and block prompt injection before it reaches your tools. Pattern scanning, session sequencing, and velocity spike detection run inline, inside the proxy.
injection_scan action = "deny"
# flag calls triggered by env, not user
allow call tool "send_email" when
trigger_source = "user"
sequence_deny tools ["exfil"]
after ["read_db"] within 60Up and running in minutes
Deploy the proxy — no agent changes needed
Sign up, open the Deploy Proxy wizard, and pick your protocols. The wizard generates routes.yaml, proxy.yaml, Docker Compose, and Helm values pre-filled. Enable keyless mode to drop in front of an existing agent with zero code changes.
Write a policy
Declare allow, deny, and escrow rules in the human-readable DSL. The inline linter flags errors as you type.
Monitor decisions
Watch every agent action in the dashboard. Export signed evidence packages that anyone can verify offline.
Free plan · No credit card
Governance without the tax
Route matching, identity resolution, policy evaluation and forwarding through the HTTP proxy, measured against a direct call to the same upstream.
p50 overhead
median added per request
p95 overhead
added per request
p99 overhead
added per request
| Percentile | Direct | Through Clausgap | Overhead |
|---|---|---|---|
| p50 | 0.05ms | 0.11ms | +0.06ms |
| p95 | 0.06ms | 0.26ms | +0.20ms |
| p99 | 0.10ms | 0.44ms | +0.34ms |
Policy evaluation alone takes under 2 µs at p99, including a 20-policy bundle. Audit records are written off the request path and are not part of these figures. Measured October 2026 with the benchmark tool in the Clausgap source.
Start governing your AI agents today
No credit card. Policy engine, M-of-N approvals, and evidence anyone can verify offline — free to get started.