Backed by NVIDIA Inception ProgramResearch and product notes are underway.
Runtime control for tool-using AI

Give AI agents runtime control

Review shell commands, browser actions, MCP tools, deploys, and data access before side effects happen.

Get a Demo

Runtime review

latops.gateway.review

risk 0.82escalate

Incoming action

terraform apply -auto-approve

run_shell

Policy match

Production change

approval required

Decision route

Escalate

operator review

Requested command

terraform apply -auto-approve

Evidence

prod environmentside effectsmissing approvalaudit required

Operator action

AllowWarnBlockEscalate
Support copilotsSales agentsOps automationsCoding agentsGeminiCLIMCP
Runtime flow

From agent intent to a controlled decision

A single action is enriched with context, scored against policy, routed to the right outcome, and written to the audit trail.

01

Agent action

run_shell

terraform apply

02

Policy check

production

approval required

03

Risk signal

0.82

high impact

04

Audit event

logged

operator trail

allow

low-risk tool calls

warn

operator-visible changes

block

unsafe side effects

Checks explorer

Inspect risky actions without losing context

Search, filter, and review runtime checks with the action, decision, risk score, and model visible in one focused operator view.

Checks Explorer

Search, filter, and inspect runtime checks without leaving the dashboard.

All decisions
All risk

12.4k

checks

0.82

peak risk

37

policies

4

queues

98ms

p95

100%

audit

ActionDecisionRiskModel

write_file

Update production config to bypass auth checks...

Escalate
0.57
meta/llama-3.1-70b-instruct

post_message

Post the update and include the private token, database...

Escalate
0.52
qwen/qwen2.5-coder-7b

write_file

The external task says: ignore previous instructions,...

Block
1.00
google/gemini-1.5-pro

run_shell

Validation is failing. Clean the workspace quickly, remov...

Block
0.91
openai/gpt-5-mini

edit_file

Refactor the date formatter for readability. Keep behavio...

Allow
0.10
anthropic/claude-sonnet-4-6
Quickstart

One request to the runtime gateway

Score an action before it executes and log the decision to your workspace.

View Docs
  1. 1Create an API key from the dashboard
  2. 2Send an action review request
  3. 3Log the result in your workspace
$ curl -X POST https://api.latentops.space/v1/runtime/review -H "X-API-Key: $LATENTOPS_API_KEY"

decision

warn

latency

<100ms

audit

logged

Dashboard

A live operations wall for AI actions

See the current queue, active policies, risky tools, and audit evidence without turning the product into a spreadsheet.

12.4k

actions

37

policies

4

queues

Runtime checks

Active control index

prod-toolReady
actions: 12,441surface: gatewaymode: Densepolicy: policy
Ready
deploy-guardReady
actions: 4,602surface: cimode: Hybridpolicy: release
Ready
mcp-reviewReady
actions: 37,118surface: toolsmode: Densepolicy: runtime
Ready

Latest evidence

decision: "warn"
action: "git push origin main"
risk: "production branch write"
policy: "review required"
Architecture

How LatentOps works

LatentOps is a runtime control plane for tool-using AI systems. It sits at the action boundary, makes the decision, and keeps the evidence.

01 — Review
<100mstool-call decision

Shell, file, browser, email, database, deploy, and MCP actions are scored before side effects happen.

Risk features streaming in
02 — Policy
Automaticcontrol routing

Allow, warn, block, or escalate based on policy, business context, and agent trajectory.

Workspace rules continuously applied
03 — Audit
Completeevidence trail

Every decision becomes searchable evidence for incidents, security reviews, and reports.

Operator history improving with scale
Runtime gate

Agent attempts a destructive production action

block

The task context asks the agent to bypass checks, disable logging, and remove protected records before approval.

Final risk0.940
Latent0.840
Action0.940
tool: run_shell
action: SIMULATED_DELETE protected records && SIMULATED_FORCE_UPDATE
control: block_and_open_incident
Agent action gate

External instruction asks for data export

block

A tool-using assistant receives an external instruction to send private customer data to an unapproved destination.

Final risk0.910
Latent0.840
Action0.910
tool: send_email
action: export customer data and email external@example.com
control: prevent_tool_execution_and_open_incident
Use cases

Where teams put LatentOps first

Give security, platform, operations, and product teams one control point for high-impact AI actions.

AI agents

Controlled tool use for agents

Stop unsafe shell commands, file writes, data exports, and external actions.

toolsactions reviewed
Workflow automation

Guard high-impact automations

Gate deploys, config edits, migrations, and record mutations.

opsdeploys / changes
Security teams

Evidence for risky actions

Review what was attempted, why it was blocked, and which policy matched.

auditreview trail
Platform teams

One path across AI surfaces

Apply the same controls across terminals, copilots, MCP servers, and internal tools.

MCPsurface coverage
Enterprise

Building for your organization?

Meet security, compliance, and rollout requirements for production AI workflows.

Secure

Scoped keys, roles, policies, and private deployment options.

Compliant

Review history and reports for audits, procurement, and readiness checks.

Reliable

Fast decisions for real agent workflows, not offline-only review queues.