LIMEN13

Platform

What LIMEN13 does, capability by capability

For each capability: the problem, how it works, what you receive, its constraints and what you can see today. Statuses refer to release 0.10.1.

In plain words. LIMEN13 does ten jobs. Together they answer four questions: which AI agents do we have, what is each one allowed to do, who stops it when something goes wrong, and how do we prove what happened. Each job below says what works today and what does not.

Status labels

How to read the labels

  • AVAILABLEIn release 0.10.1 and exercised by its automated checks.
  • PILOTIn release 0.10.1, but it depends on your own systems (identity provider, SIEM, vault, MCP servers, network rules) and has not yet been validated in customer environments. We validate it with you during a pilot.
  • PLANNEDNot built. On the roadmap, with no committed date.

The whole product is a pilot release

Deployments are assisted and start on non-production workflows. An independent penetration test is required before production.

Integration

Three ways to connect an agent

The agent's own code and framework stay as they are.

Model SDK

Point an OpenAI-compatible or Anthropic SDK at LIMEN13's address. The provider key moves into LIMEN13.

REST tools

Register each tool service with its address. LIMEN13 pins the destination and calls it on the agent's behalf.

MCP

Point the agent's MCP client at LIMEN13's MCP endpoint, or register your MCP servers as tool services.

  • OpenAI-compatible APIs
  • Anthropic API
  • MCP servers
  • REST tools
  • Slack and Teams
  • SIEM · syslog CEF/JSON
  • OpenID Connect SSO
  • Vault · OpenBao
  • Docker · systemd

Capabilities

Ten capabilities

01

Agent discovery and inventory

The problem
Teams start agents faster than security hears about them. You cannot control an agent nobody listed.
How it works
An inventory records each agent with a named human owner. Discovery looks for AI programs LIMEN13 does not know yet: sign-in attempts it refused, callers of the model gateway that are not enrolled, a scan of the host it runs on, and report-only collectors on other machines. It recognises cloud model callers, local model servers, agent frameworks and MCP servers.
What you receive
A list of candidates to enrol or dismiss with a reason, and a coverage level for each agent: declared by you, or verified by recorded controls.
Constraints
Discovery never grants authority and does not prove complete visibility. Host scanning works on Linux.
See it today
The Discovery view of the interactive demo, simulated.
  • AVAILABLEInventory with owners, coverage levels, host scan, refused sign-ins, collectors.
  • PLANNEDDiscovery through the API of a SaaS or agent platform.
02

Agent identity and authorization

The problem
Agents often share one powerful API key, so nobody can tell which agent did what, or limit one of them.
How it works
Each agent has its own key, stored only as a SHA-256 digest, and each run gets a short-lived token. An agent's role lists the tools it may call. Anything not listed is refused.
What you receive
The agent's identity in every record, key rotation, and a role you can read on one page.
Constraints
A role that is too broad still allows harmful actions inside it. Roles need an owner who reviews them.
See it today
The Agents and Policies views of the demo.
  • AVAILABLEPer-agent keys, run tokens, deny-by-default roles, named owners.
  • PLANNEDTwo-person review of configuration changes.
03

Policy enforcement

The problem
A model can be talked into anything. Asking the model to behave is not a control.
How it works
Code, not the model, checks every call: the agent's role, the argument schema, content rules (secrets, blocked terms, unapproved links, look-alike domains), whether the run has read outside content, and a risk budget. Repeated refusals contain the run. An observation mode records what would be refused without blocking.
What you receive
A decision for every call (allowed, held or refused), with the rule and the reason.
Constraints
Content detection is heuristic and comes second to permissions. No product stops every prompt injection. LIMEN13 limits what a hijacked agent can do through it.
See it today
The scenarios in the demo's attack lab.
  • AVAILABLERoles, schemas, content rules, taint tracking, risk budget, automatic containment, observation mode.
04

MCP and API security

The problem
MCP servers and REST tools are a supply chain of their own: a tool description can change after you approved it, and a tool server can be swapped.
How it works
Tool services are registered with a destination pinned at enrolment. For MCP servers, the agent sees the tool names, descriptions and schemas approved at enrolment, never the server's live list, and a check compares the server with what was approved. MCP clients can also connect to LIMEN13's own MCP endpoint.
What you receive
Approved, pinned tools for each agent, and a high-severity alert when an approved tool has changed or disappeared.
Constraints
Tools only: no sampling, elicitation, prompts or resources. Streamable HTTP only. Servers with a bearer token or no authentication; MCP servers protected by OAuth are not supported. The change check runs on demand, not on a schedule.
See it today
The unapproved MCP tool scenario in the demo.
  • AVAILABLEREST and MCP tool services, pinned destinations and definitions, change check, MCP endpoint.
  • PILOTValidation against your own MCP servers.
05

Human-in-the-loop approvals

The problem
Some actions are fine for an agent to propose and wrong for it to decide: refunds, payments, messages to customers.
How it works
Policy marks the actions that need a person, for example above an amount or after the run has read outside content. The action waits. A named approver decides in the console or in the simple view. Policy is checked again at the moment of the decision, and requests expire.
What you receive
A queue of pending actions showing who asked, what and why, and a record of who decided.
Constraints
An approver can approve a harmful request. Separation of duties (the owner cannot approve, two-person tools) exists but has to be switched on. After a restart, pending requests cannot execute.
See it today
The Approvals view of the demo: approve or reject the $940 refund.
  • AVAILABLEHeld actions, named approvers, re-check at decision time, expiry, simple view.
06

Agent kill switch and containment

The problem
When an agent misbehaves, minutes matter, and you may not control the agent's own process.
How it works
Disabling an agent refuses its next call, contains its runs and cancels its pending approvals, without waiting for a slow tool to finish. A run is also contained automatically after a set number of refusals, or when a decoy record is touched.
What you receive
One action that stops an agent at the gateway, with a recorded reason.
Constraints
A call already passed to an external service may still finish, and its effects cannot be recalled. LIMEN13 does not terminate the agent's own process.
See it today
The kill-switch scenario in the demo.
  • AVAILABLEDisable and enable agents, contain runs, automatic containment.
07

Model gateway and usage controls

The problem
Agents that hold provider keys can call any model, send anything and spend without limit.
How it works
Agents call models through LIMEN13 with the OpenAI-compatible or Anthropic API. Provider keys stay in LIMEN13. Each agent has allowed models, a rate limit and an estimated daily token budget, and prompts are scanned for secrets and decoy markers.
What you receive
A record of each model call (prompts are not stored, only their SHA-256), and a refusal when a rule or a budget is reached.
Constraints
The budget is estimated accounting, not a billing cap: set a hard limit at your provider too. It only controls model traffic that passes through it, which is why outbound network rules matter.
See it today
The decoy scenario in the demo.
  • AVAILABLEOpenAI-compatible and Anthropic gateway, allowed models, rate and estimated token budgets, prompt scanning.
  • PILOTValidation with your provider accounts and network rules.
08

Event timelines and investigation

The problem
After an incident, reconstructing what an agent did from scattered logs takes days.
How it works
Every run has a storyline: each step, decision and reason, in order. Alerts are mapped to the OWASP Top 10 for LLM Applications and to MITRE ATLAS, and can be acknowledged, resolved and reopened.
What you receive
A step-by-step account of any run, event logs for each agent, and an alert queue with its status.
Constraints
It shows what passed through LIMEN13. Activity outside it is not in the timeline.
See it today
The Overview and Activity views of the demo.
  • AVAILABLERun storylines, event logs, alert triage, framework mapping.
09

Signed and verifiable audit evidence

The problem
A log that can be edited quietly proves nothing to an auditor, or after a dispute.
How it works
Each record is redacted, signed with HMAC-SHA-256 and chained to the one before it. The chain is verified at every start and on demand. Copies of the chain head can be sent off the machine at intervals, so that a rewrite is detectable even by someone who obtained the key. Reading the evidence is itself recorded.
What you receive
A verification you can run yourself, and a report that keeps “unchanged” and “covers the whole period” as two separate verdicts.
Constraints
Without off-site anchors, local signing cannot prove integrity against someone who controls both the host and the key. Renewing the evidence-signing key on an existing chain is not supported in this release.
See it today
The Evidence view of the demo: verify the chain, alter a record, verify again.
  • AVAILABLESigned hash chain, verification at start and on demand, recorded evidence reads.
  • PILOTOff-site anchoring to your own storage.
10

Security reporting and integrations

The problem
Leadership, auditors and an MSSP's customers need a regular, consistent account, and the SOC needs alerts where it already works.
How it works
A monthly report is computed only from verified evidence: reliability of the record, coverage of each agent, inventory, activity, refused actions and why, approvals, alerts and incidents, and rule-based recommendations. Alerts go to a SIEM over syslog (CEF or JSON), to HMAC-signed webhooks, to Slack and to Teams.
What you receive
A draft report in Markdown and HTML for an analyst to review, and alert delivery that resumes after a restart.
Constraints
The report is a draft for a person, in English, with no PDF export in this release. Webhook delivery is at least once: receivers should de-duplicate event IDs.
See it today
The sample report, with fictional data.
  • AVAILABLEMonthly report; delivery over syslog, webhooks, Slack and Teams.
  • PILOTValidation with your SIEM and chat workspace.

See these controls decide, step by step.

The demo is a simulation that runs in your browser. A pilot runs the real engine on your own agent.