Fleak joins the Databricks startup accelerator. See the announcement

Every agent session, on the record.

Fleak collects telemetry from AI coding tools, AI gateways, and your own agents, normalizes it to one schema, masks prompts and secrets, and routes it to your SIEM and lake. Security sees what agents did. Finance sees what they cost.

Book a demo See how it works
  • Claude Code
  • OpenAI Codex
  • Cursor
  • GitHub Copilot
  • OpenTelemetry
  • Your own agents

~40%

Lower token cost

when agents reason over normalized data (Gruve)

1

Schema for every AI tool

sessions, prompts, tool calls, and tokens

0

Raw secrets downstream

masked before any destination sees them

Agents act on your systems.
Their logs don't agree on anything.

Every AI tool logs something different, and most ship with content capture switched off. When someone asks what an agent did, or what it cost, nobody can say.

Every tool, its own format

Claude Code, Codex, Cursor, and Copilot emit different events, fields, and identifiers. Following one session across them is manual work.

Policy changes, not sessions

Audit trails record who changed a setting, not which commands an agent ran, which files it read, or which data it sent out.

Prompts carry secrets

Prompts and tool calls include API keys, customer data, and source code. Shipping them raw to the SIEM creates a new exposure.

Usage, cost, and risk,
for every model and agent.

Fleak treats AI telemetry like any other security source: collected, normalized, masked, enriched, and routed by what each team needs.

Collect agent telemetry

Ingest OpenTelemetry traces, audit logs, and compliance APIs from Claude, OpenAI, Cursor, GitHub Copilot, Gemini, AI gateways, and your own agent frameworks.

One schema for every agent

Normalize sessions, prompts, responses, tool calls, and token usage into one schema, so one query works across every tool.

Mask prompts and secrets

Mask API keys, credentials, and personal data in prompts, responses, and tool parameters before they reach any destination.

Cost per user, team, and model

Enrich every call with the user and team from your directory, track token spend against budgets, and alert before a team overruns.

Flag risky actions

Detect shell commands, credential access, sensitive file reads, and data egress in agent sessions as they stream.

Govern the gateway

Route model calls through your AI gateway with budgets, fallbacks, and blocked requests, and keep an audit record of every call and routing decision.

Find shadow AI

Spot AI tools and models nobody approved, from egress, identity, and gateway telemetry.

Prompt and response audit

Keep a searchable, masked record of prompts, responses, and tool calls for investigations and compliance reviews.

Agent access, governed

Pair telemetry with Fleak's governed MCP gateway, so every agent's access to your systems is scoped, time-limited, and audited.

Account for every agent.

What it did, what it touched, what it cost. Answered with a query, not a guess.

Read the docs Book a demo

Agents generate telemetry.
Fleak makes it accountable.

The Brain maps each AI tool's events to one schema. The Muscle masks, enriches, and routes them deterministically: risky actions to the SIEM in real time, full sessions to the lake, usage to the teams who pay for it.

Sources

  • Claude Code
  • OpenAI Codex
  • Cursor
  • GitHub Copilot
  • Gemini
  • Custom agents

A coding agent runs a command that reads ~/.aws/credentials.
Who finds out, and when?

Without Fleak

Invisible

The command sits in a local trace export nobody collects. The audit trail shows nothing.

With Fleak

Flagged

The session event is normalized, the secret masked, and the command flagged to the SIEM in real time.

Accountability

Attributed

Enriched with the developer, the team, and the token cost, so the follow-up is a conversation, not a forensics project.

When someone asks what an agent did or what it cost, the answer is a query, not a guess.

"Agents you can't account for are agents you can't scale. The telemetry exists. It needs a pipeline."

AI observability · Fleak Platform

Bring the AI tools your developers already use.

30-minute working session. We'll show their sessions normalized, masked, and searchable.

Read the docs Book a demo

Explore Related

Collection

Every source, one ingestion layer

Receivers, API connectors, and cloud log discovery.

See Detail →
Edge

Collect and filter at the source

One agent, managed centrally, cutting volume before it moves.

See Detail →
OCSF

Every source in OCSF, long tail included

211 templates, plus AI-generated mappings for the rest.

See Detail →