The AI agent control plane: discover, govern, secure, assure and prove.
Praesidia is an AI agent control plane: one place to find the agents and MCP servers you run, map what each AI system is built from, decide what it may do before it acts, test it before release, and hand an auditor proof they can check themselves. Every capability it ships is on this page.
Unmarked capabilities are in every plan, including the free Developer plan. Individual Advanced Enterprise mark the plan a capability starts on; see pricing.
Know what is running and what it can reach
Start with an inventory you can trust, built from the traffic that reaches Praesidia and the sources you connect.
-
Agent discovery
An inventory of the agents that reach Praesidia, plus the agents, MCP servers, tools and prompts defined in the GitHub and GitLab repositories you connect.
-
MCP inventory
The MCP servers registered with Praesidia, and the tools each one exposes, each server with its own identity.
-
MCP and skill scans
Scan an MCP server or skill manifest before you adopt it, against the permission policy you set for that server or for the whole organization.
-
AI intake
Teams submit a new AI system for review, and reviewers work through a queue before it is registered.
-
Data assets and data flow
Register the data your AI systems touch, trace each system's data flow down to classified data, and record a cross-border verdict for every flow.
Know what each AI system is made of, what it puts at risk and who owns it
An agent is a model, prompts, tools, data and permissions. Keep each AI system on record with its owner: the parts and how they connect, the risks it carries and the controls that hold them, the access it holds against the access it uses, and what it costs.
-
AI System Graph
What each AI system is built from, how its parts connect, and what is exposed when one part fails.
-
AI bill of materials
A signed, versioned list of every model, prompt, tool, data source and policy an AI system uses, compared release to release and exported as CycloneDX.
-
Risk register
The risks you carry per AI system, the controls that hold each one down, and when each is due for another look.
-
Least privilege
Compare the permissions each agent holds with the ones it actually uses, and turn the difference into a proposal to reduce them.
-
Cost and business value
Model spend attributed per agent, set next to the business KPIs you record for each AI system.
AI FinOps guide · per-agent spend Individual
Decide what each agent may do, before it acts
Every call an agent makes through Praesidia is checked against who it is and what it is allowed to do, then allowed, denied or sent to a person.
-
Runtime policy and identity
Each agent gets its own identity, and each tool call is decided against the rules for that identity before it runs.
-
Guardrails
24 ready-made guardrail templates, or your own rules, on what goes into and comes out of an agent.
-
Agent policies
Rate limits and spend caps for a group of agents, rolled out in stages like every other rule.
Individual -
MCP tool permissions
Decide which MCP tools each connection may call; Praesidia records the decisions it makes on the calls that pass through it.
-
Model gateway and routing
One gateway to ten model providers, with the health of each provider in view. Separately, routing policies pick among the model configurations you give an agent, by data classification, residency and cost.
-
Staged rollout and policy replay
Move a rule from simulate to observe to alert to enforce, with optional automatic rollback. Replay a rule against past traffic before you turn it on.
Replay Advanced -
Sequence rules
Catch one action followed by another within a time window, such as a customer-list read followed by an external email.
Advanced -
Human approval
Send a risky tool call to a named person, who approves or refuses it before it runs.
Multi-step workflows Enterprise -
Emergency freeze
Stop agent traffic for the whole organization in one step, after previewing every AI system that will stop.
In observe mode the rule is evaluated and recorded and the call goes through unchanged; in enforce mode the same rule can deny it or send it to a person.
Test it before it ships, and keep fixing what you find
Evaluations, red-team campaigns and release gates tell you whether a release meets your bar; findings and remediation close the gap.
-
Evaluations and release gates
Test each agent release against your own cases and your own bar, and hold the release until it passes.
-
Red team
Attack campaigns against the agents, connections and MCP servers you opt in, on demand or on a schedule.
Assurance Advanced -
Decision explanations
Open any policy decision and see its outcome, the rule that matched, the data classification, the identity behind it and its EU AI Act basis.
-
Findings
Findings raised automatically across the platform, in one list a person reviews, triages and accepts.
-
Remediation planner
Proposed fixes that are simulated, tested, approved, applied and monitored, and rolled back if they cause trouble.
Show afterwards what happened, and let others check it
Evidence an auditor, regulator or buyer can verify without trusting our console. A signed record shows it was not changed afterwards. It does not show that every action was captured.
-
Signed audit trail
Governed decisions recorded, signed and chained under the default configuration, and checked offline by your auditor.
-
Audit package
One export for a period: executive summary, inventory, risk register, controls, evaluations, incidents and the signed evidence bundle.
-
Trust passport
A public page, badge and PDF for the AI systems you choose to publish, and a check for the passports your vendors send you.
-
Incident response
Related signals grouped into one incident, a nine-step response, a signed evidence bundle and a regulator report.
-
Compliance mappings
Controls and evidence mapped to ISO/IEC 42001, the NIST AI RMF and the EU AI Act, with evidence coverage per framework and an ISO/IEC 42001 gap view.
Coverage and gap view Enterprise -
Reports and forensic search
A weekly executive report for organization owners, and search across audit records when you need to reconstruct what happened.
Forensic search Advanced
It fits the stack you already run
-
Integrations
Agent frameworks, model providers and MCP servers on the way in; on the Advanced tier and Enterprise, security events to your SIEM and tickets to your ITSM on the way out.
SIEM and ITSM Advanced -
Observability
OTLP traces in; organization metrics in Prometheus format for the monitoring stack you already run.
-
Governance packs
Ready-made sets of policies and controls to start from instead of a blank page.
Docs Advanced
Common questions about the AI agent control plane
An AI agent control plane is one place to see the AI agents you run, set what each may do, and keep a record of what was decided, instead of rebuilding those controls inside every agent. Praesidia is an AI agent control plane for the agents, models and MCP tools you connect to it. Calls that do not pass through Praesidia are not recorded.
No. New policies run in observe mode first: the verdict is recorded without changing the outcome, until you turn enforcement on. In enforce mode the same rule can deny a call or send it to a person.
Each capability on this page is marked with the plan it starts on; unmarked ones are in every plan, including the free Developer plan. Approving a held call is in every plan; multi-step approval workflows are Enterprise. Compare plans.
Praesidia records the decisions it makes in an append-only, hash-chained audit trail, signed under the default configuration; an operator can turn signing off. A signed record shows it was not changed afterwards. It does not show that every action was captured. External anchoring is optional; without it, removal of the most recent records cannot be detected. An exported bundle can be checked offline, without trusting our console.
No. Controls and the evidence they produce are mapped to ISO/IEC 42001, the NIST AI RMF and the EU AI Act. A mapping is not a certification, and Praesidia does not certify you against any of them.
Go deeper
- The AI control plane guide — what a control plane does, and how it differs from an API gateway.
- The AI governance guide — control points, human oversight, audit trails and regulation for AI agents.
- AI control plane vs API gateway — where each one stops.
- The AI agent audit trail guide — what a record of agent activity must contain, and how it is checked.
- AI control plane and AI agent governance, defined in the glossary.
Put one agent under the control plane.
Create your workspace, connect one agent, and inspect the decisions it produces.