Straza governs what your AI agents do.
Your identity governance system decides who. Straza enforces what their agents may do, at every action.
-
Identity
Your IGA sends your people, your AI agents and a sponsor for each agent: the person accountable for it.
-
RBAC
You make roles from the tools of MCP servers. Nothing is reachable until a role says so.
-
MCP gateway
Your MCP servers sit behind one gateway. An MCP server offers tools to an agent, such as restarting a service.
-
Delegated MCP administration
Every MCP server gets its own admin role. The team that owns a server runs it and no other, and your IGA decides who gets the role.
-
Human in the loop
Fine-grained approvals for MCP tools and shell commands. You choose which calls need a person’s yes. No answer is a deny.
-
Policy
Every tool call is decided before it runs. It is allowed, denied, or it waits for a person.
-
Credentials
The gateway adds a server’s credential after the decision, so the agent never has it.
-
Audit
Every decision, allowed or denied, goes into a hash chain, and a sink copies it to your SIEM.
On this page6 sections
One path from every agent to every tool
Every agent action passes through Straza. Identity comes in from your IGA, and the audit chain streams out to your SIEM.
Agents
- Claude Code
- Codex
- Gemini CLI
- Python agents
- Any other LLM
your IGA
SCIM 2.0 or REST
Audit sinks
file
webhook
your SIEM
MCP servers
- billing
- deployer
- helpdesk
- crm
- + any MCP server
straza exec is the only shell they get.Your IGA grants the roles that expose tools.
You create roles in Straza and bind tools to them. Your IGA imports the roles and decides who gets each one.
Approvals are signed on the device
Every decision on the phone is signed by a hardware key that is generated there and never leaves it. If the phone is lost, an administrator revokes it server-side, its token stops working, and the pairing is gone.
On iPhone and Android today. The console, a browser page and Slack are approver channels alongside the phone.
What the video shows, step by step
joe-java-developer-agentis asked in a terminal to restart the checkout service on prod withdeployer.- Straza stops the call before it runs, and the agent waits for a person.
- The approver app lists
restart service in deployer. The approver reads the tool, the requester and the exact arguments, taps Approve and enters the device PIN. - The app shows "Recorded: approved" and the agent reports the checkout service restarted and healthy.
One MCP gateway, one policy, one audit chain
RBAC for every tool. Nothing is reachable until a role says so.
What the video shows, step by step
- The console's MCP servers page reads "0 servers". A person clicks Add MCP server, types the address
https://learn.microsoft.com/api/mcpand changes the suggested namelearntomicrosoft-learn. - The credential step stays on None. The review reads "Nobody can call it until a role gives access to its tools." The person clicks Save and publish and confirms in "Publish microsoft-learn?".
- The console reports "microsoft-learn is live. Straza reached it and read 3 tools."
- Give access to it opens a new role, which the person names
microsoft-learn-docs. Under Choose per tool the person ticksmicrosoft_docs_searchandmicrosoft_docs_fetch. The third tool,microsoft_code_sample_search, stays unticked and reads "no access". - The person sets
microsoft_docs_fetchto Needs approval, decided by "The agent's sponsor" through "A ticket, granted within" one day, then one hour to run. The summary reads "1 tool is allowed. 1 needs a ticket the agent's sponsor grants within a day, after which the same call is allowed for an hour." The role is published and the console reports "microsoft-learn-docs is live." - On the business role
java-developerthe person clicks Compose a role, picksmicrosoft-learn-docs, types the role's name to confirm and publishes. The role's page then listsmicrosoft-learnwith "2 tools via microsoft-learn-docs". - In the agent's terminal,
joe-java-developer-agentis asked "Which microsoft-learn tools can you call? Then search the docs for Java on Azure Container Apps, one line." It listsmicrosoft_docs_searchandmicrosoft_docs_fetch, runs the search and answers with a link to Microsoft Learn. The status line reads "2 roles" and "2 MCP tools".
- Fail-closed
- Platform unreachable: an approval-gated call denies within two seconds.
- Self-hosted
- One server binary, laptop to Kubernetes. No callout to us at decision time.
- Open source
- Core AGPL-3.0-only. SDKs, spec, plugins and the approver app Apache-2.0.
- No AI inside
- Enforcement is deterministic policy evaluation.
- Agents can draft
- With the drafting role, an agent proposes a change to servers, roles or policy. Straza checks it, and only a person publishes it.
Releases
- v2.2.0
An AI agent can call an MCP server as its own Microsoft Entra agent identity. The console writes rules for MCP servers on people's own machines.
- v2.1.0
A tool that tells an agent what it may do. Pin a server's objects in the console.
- v2.0.0
Every credential on a person's machine moves into a local service that agents running as that person cannot read. Roles that cover some objects of a server, git through Straza, rules by tool name.
Full changelog · How Straza compares with five other open-source projects
Run it yourself
The whole stack on one machine
Straza with midPoint as the IGA, Keycloak and a working AI agent. Needs Docker with Compose 2.24 or later, 4 cores, 8 GB of memory and 20 GB of disk. Run it on a trusted network only.
git clone https://github.com/strazahq/straza.git && cd straza
deploy/compose/eval-stack/up.sh
docker compose -p straza-eval logs -f --no-log-prefix eval-seedStandalone
No Docker. Get strazad and straza from the releases page and follow the guide. It starts the server, enrolls your machine and connects Claude Code. A starter rule denies rm -rf, and a rule you add makes git push wait for your approval in the browser.
Or hand it to your coding agent
Read https://straza.ai/llms.txt and https://docs.straza.ai/evaluate/try-it-on-one-laptop/. Then set up Straza on this machine in standalone mode, enroll it, and show me one command that policy denies.