What is an AI control plane?

An AI control plane is a single system that inventories every AI agent an organization runs, records how those agents call each other and which tools and data they reach, monitors their health, and attributes their cost. ThoughtShieldAI provides this without requiring an SDK inside each agent.

How does ThoughtShieldAI discover AI agents without instrumentation?

It reads the transcripts and logs that agent runtimes already write to disk, and queries read only platform APIs through CLI sessions you are already signed into. No SDK is added to your agents and no application code changes.

Does ThoughtShieldAI send our data to a vendor cloud?

No. It is local first. Data is stored in a SQLite database on your own infrastructure and the dashboard binds to loopback only. Connector credentials stay in a permission locked local config file and never enter the fleet database.

Which platforms does ThoughtShieldAI support?

Local agent runtimes Claude Code, Codex, OpenClaw, and Hermes, plus read only catalog connectors for Google Vertex AI, Amazon Bedrock, Cloudflare, and Azure AI Foundry. It also exposes 16 MCP tools so an assistant can operate the fleet.

How does ThoughtShieldAI find shadow AI agents?

It builds the agent inventory from what is actually running rather than from what was registered, so any agent producing transcripts or appearing in a connected platform catalog shows up even if it was never approved. Agents outside the approved inventory are surfaced with their platform, owner, and last run.

How does ThoughtShieldAI attribute AI agent cost?

It reads token usage per run from the agent transcripts and applies per model pricing to produce cost per agent and per model. Runs it cannot price are labelled as unpriced rather than estimated, so the totals are never inflated by a guess.

Is ThoughtShieldAI open source?

The core is open source under the MIT license and published on GitHub, so a security reviewer can read every line that touches agent transcripts before it runs in a client environment.

The AI control plane for enterprise agentic AI

Every AI agent in your enterprise, on one screen.

ThoughtShieldAI finds every agent running across your clouds, maps which agent calls which, and shows what each one costs. No SDK, no code changes, and nothing leaves your infrastructure.

Discover · Map · Observe · Attribute · Govern building

Request a demo See it in action

Complete AI visibility

One trusted inventory of every AI agent, model, tool, and MCP server across every cloud and business unit.

Shadow AI surfaced

Agents running outside the approved inventory are named, owned, and put in front of security review.

AI spend under control

Token usage, model cost, and duplicated agents attributed back to the agent and the team that caused them.

Your agent count went up
again last week

Nobody approved it, nobody logged it, and nobody owns it. Here is what that costs you.

01 Nobody has the list Agents pile up across clouds and teams with no inventory and no named owner.

02 Shadow AI ships anyway Anything deployed outside approved channels is invisible to security review and to audit.

03 The wiring is a guess No view of how agents, tools, APIs, MCP servers, and data actually reach each other.

04 Failures stay quiet A scheduled agent stops running and you find out from the work it did not do.

05 The bill has no owner Token usage, model costs, and duplicated agents grow with nothing attributing them back.

The platform

One intelligence layer
for your AI estate

One chain, in this order, everywhere it appears. Four of the five run today; the fifth is marked wherever it is named.

DiscoverShipped
MapShipped
ObserveShipped
AttributePartial
GovernBuilding

How many agents do we have, and who owns them?

Discover

Every agent, model, tool, and MCP server found automatically. No SDK to install, no code to change.

Which agent called which, and what did it reach?

Map

Which agent spawned which, what each one reaches, and how much context moved between them. The relationship graph is the part nobody else ships at this fidelity.

Which agents need attention right now?

Observe

Health, usage, and behaviour tracked continuously across every platform, with an alert when a scheduled agent goes quiet.

What is each agent costing us?

Attribute

Cost per agent and per model, read from real token counts. Runs we cannot price are labelled unpriced rather than estimated.

Are they inside policy?

Govern Building

Ownership, policy, and compliance to one standard instead of per team. This is the one question the platform cannot answer yet.

Connects to

Claude Code Codex Amazon Bedrock Google Vertex AI Azure AI Foundry Cloudflare MCP

See your AI estate
in action

Screenshots of ATLAS, the relationship graph, and the mindmap as they ship today, not concept mockups.

35: agents discovered
547: agent relationships
2,871: runs correlated
6: platforms connected

Figures from the demonstration fleet shown below, not a customer environment.

Atlas

One inventory that stays current

Atlas connects to each platform read only and builds the inventory itself, then keeps it current as agents appear, change, and retire.

Reads the logs your agents already write
Captures platform, owner, trigger, and staleness per agent
Surfaces agents running outside the approved inventory

ThoughtShieldAI Atlas dashboard listing discovered agents with alerts, stale agents, and estimated cost

Lineage

Trace a workflow instead of guessing at it

When an agent triggers another agent, that relationship is captured automatically, so you can follow a workflow end to end.

Agent to agent, agent to tool, and agent to data relationships
Volume of context passed along each edge
Active against stale subagents, surfaced per lineage root

ThoughtShieldAI Lineage view showing a graph of agents spawning subagents across several platforms

Mindmap

Unfold any agent, one level at a time

Double-click an agent and the agents it spawns unfold to its right, the agents that spawn it to its left. Unfold again to follow a chain of subagents as far as it goes, and fold a branch away once you have seen enough.

Every edge carries its spawn count and the volume of context passed each way
The detail drawer lists the same relationships, the spawn tree, and the agent's cost and tools
Large fan-outs are capped at twelve per agent on the map; the drawer always lists all of them

How the mindmap works

ThoughtShieldAI mindmap: invoice-processing-agent unfolded three levels deep, from approval router and OCR extraction to PO matching and ERP lookup, with spawn counts on every edge

All three screenshots show the same demonstration fleet: 35 agents, 547 parent to child spawn edges, and $26.20 of estimated 30 day cost with zero unpriced runs. It is a sample used to show the product, not a customer environment.

How it works

Read only in.
One correlated graph out.

ThoughtShieldAI reads what your agents already write, correlates it into one graph, and serves it back through three surfaces. Every part of it runs on infrastructure you control.

Read only in → one correlated graph → CLI, dashboard, and MCP tools out

Enterprise AI ecosystem OpenAIClaudeGeminiAzure AIAmazon BedrockVertex AI LangChainMCP servers and toolsCustom and internal agents

ThoughtShieldAI Enterprise agentic AI control plane Discover · Map · Observe · Attribute · Govern

Shipped ATLAS AI discovery and inventory

Discover AI agents
Enterprise inventory
Relationship mapping

Shipped VISION AI observability

Runtime monitoring
Subagent health
Silent failure alerts

Partial PULSE AI FinOps

Token and usage analytics
Cost per agent and model
Vendor wide FinOps building

Building PHOENIX AI security

Threat detection
Data protection
Risk prioritization

Building NEXUS AI governance

Trust and identity
Policy enforcement
Compliance monitoring

Unified telemetry and data layer Cloud platformsAPIs and servicesMCP servers and tools Databases and data storesCode reposLogs and eventsIdentity and access

Ingestion is strictly read only. Data lands in a local SQLite store on your own infrastructure, behind one shared query layer, so the CLI, the dashboard, and the MCP tools report identical numbers instead of three drifting answers.

ATLAS, VISION, and PULSE are running in the product today. PHOENIX and NEXUS are in active development and their sequencing is genuinely open. Underneath all five, no connector can mutate a source platform: Vertex, Bedrock, Cloudflare, and AI Foundry are catalog only reads, and we treat any code path that could write back as a security bug rather than a feature request.

What works today,
and what does not yet

We would rather lose a deal on scope than win one on a roadmap promise.

ShippedWorking today

Zero instrumentation discovery across Claude Code, Codex, OpenClaw, and Hermes

Read only cloud connectors for Vertex, Bedrock, Cloudflare, and Foundry

Agent to agent lineage graph with data sharing metrics per edge

Cost analytics that label what cannot be priced instead of guessing

Silent failure alerts on declared cadences and token spikes

CLI, dashboard, and 16 MCP tools over one shared query layer

BuildingIn active development

Pulse, full FinOps intelligence across every major model vendor

AI native security and threat detection

Risk scoring for shadow, orphaned, and over privileged agents

Compliance tracking against internal AI use policy

Department and team spend rollup

Multi tenant workspaces, RBAC, and SSO

Sequencing is genuinely open. Design partners get to decide what ships first.

Built for enterprise AI

Nothing leaves your
infrastructure

ThoughtShieldAI reads transcripts your agents already wrote and keeps every byte on hardware you control.

Self hosted end to end

Runs inside your environment on your own machine or infrastructure. There is no vendor cloud in the path and no account to create.

Secure by design

Discovery is read only. No connector can write back to a source platform, and the dashboard binds to loopback rather than a network interface.

Auditable before you run it

The core is MIT licensed and on GitHub, so your reviewers can read every line that touches agent transcripts before it reaches a client environment.

Loopback only dashboard

Binds to 127.0.0.1. State changing requests need a same origin check and are size capped.

Credentials stay out of the database

Connector credentials live in a local, permission locked config file and never enter the fleet store.

Fails safe on hostile input

Every parse is guarded, so a corrupt or malicious transcript cannot crash or hang the collector.

Watchdogs cannot be quietly disabled

Cadence grammar is bounds checked, so a declared alert window cannot be degraded by a bad interval.

Why ThoughtShieldAI

The status quo is not a competitor.
It is tribal knowledge and spreadsheets.

vs doing nothing

Today "who owns this agent" lives in someone's head. ThoughtShieldAI turns that into a queryable inventory that updates itself every time an agent runs.

vs generic observability

APM and logging tools were not built to understand agent to agent spawning. ThoughtShieldAI treats the parent to child relationship as a first class object, not a trace span.

vs platform native consoles

Each platform's own console only sees its own agents. ThoughtShieldAI is the one view that spans Claude Code, Codex, Vertex, Bedrock, and every other platform your teams run.

Partnership and system integrators

An open source core, built to
co-develop the governance roadmap

The core is MIT licensed, so a system integrator can stand up a pilot on a client's own infrastructure without a procurement cycle, read the code that touches client data, and contribute back what a real deployment surfaces.

01 Fast pilots Self hosted on the client's own machine or infrastructure. No vendor procurement gate before a first pilot can start.

02 Auditable by design Your technical reviewers can read every line that touches a client's agent transcripts before it goes near a client environment.

03 Roadmap, together The governance and compliance layer is scoped by real deployments, not a whiteboard. Partner feedback decides what ships next.

We have not fixed a pricing or revenue share model yet, and that is deliberate. We would rather design the engagement model with the first few partners than impose one nobody asked for.

Use cases

What can ThoughtShieldAI
tell you?

The questions that have no good answer today, and which capability answers each one.

How many AI agents do we have, and who owns them? Discover · ATLASShipped

Which agent called which, and what did it reach? Map · ATLASShipped

What are they accessing, and how much context moved? Map · ATLASShipped

Which agents need attention right now? Observe · VISIONShipped

What is each agent costing us? Attribute · PULSEPartial

Are they inside policy, and who approved them? Govern · NEXUSBuilding

Questions we get asked

The technical review,
answered up front

What is an AI control plane?: An AI control plane is a single system that inventories every AI agent an organization runs, records how those agents call each other and which tools and data they reach, monitors their health, and attributes their cost. ThoughtShieldAI provides this without requiring an SDK inside each agent.
How does ThoughtShieldAI discover AI agents without instrumentation?: It reads the transcripts and logs that agent runtimes already write to disk, and queries read only platform APIs through CLI sessions you are already signed into. No SDK is added to your agents and no application code changes.
Does ThoughtShieldAI send our data to a vendor cloud?: No. It is local first. Data is stored in a SQLite database on your own infrastructure and the dashboard binds to loopback only. Connector credentials stay in a permission locked local config file and never enter the fleet database.
Which platforms does ThoughtShieldAI support?: Local agent runtimes Claude Code, Codex, OpenClaw, and Hermes, plus read only catalog connectors for Google Vertex AI, Amazon Bedrock, Cloudflare, and Azure AI Foundry. It also exposes 16 MCP tools so an assistant can operate the fleet.
How does ThoughtShieldAI find shadow AI agents?: It builds the agent inventory from what is actually running rather than from what was registered, so any agent producing transcripts or appearing in a connected platform catalog shows up even if it was never approved. Agents outside the approved inventory are surfaced with their platform, owner, and last run.
How does ThoughtShieldAI attribute AI agent cost?: It reads token usage per run from the agent transcripts and applies per model pricing to produce cost per agent and per model. Runs it cannot price are labelled as unpriced rather than estimated, so the totals are never inflated by a guess.
Is ThoughtShieldAI open source?: The core is open source under the MIT license and published on GitHub, so a security reviewer can read every line that touches agent transcripts before it runs in a client environment.

See your real fleet in an hour

Point ThoughtShieldAI at your existing agent logs and get a live inventory, a lineage graph, and a cost view. Self hosted, and nothing leaves your infrastructure.

Run a pilot on your fleet Read the source

adarshk@thoughtshieldai.com