---
title: "Opencontroller Capability Reference"
description: "Every Opencontroller capability, block by block, plus Opencontroller Edge and the Q4 2026 roadmap."
canonical: https://opencontroller.lyzr.com/capability-reference
last-updated: 2026-10-27
---

# Opencontroller Capability Reference

[Opencontroller](/)

Search capabilities

[Compare](/compare/controlplane)[Benchmark](/llm-gateway-benchmark)[Book a demo](https://www.lyzr.ai/book-demo/)

The route

## Ten stops, four stages.

Find it, ship it, run it, improve it. Each stop is a control block that works with agents on the platforms where they already run. The number on each card counts its capabilities.

Find

[0133 capabilities

Discover
Know every agent you run, and what makes it risky.](#discover)

Ship

[0219 capabilities

Register
Give every agent an owner, a purpose and an approval.](#register)[0323 capabilities

Identity
Know who is acting, and on whose behalf.](#identity)[0420 capabilities

Evals
Test agents before you trust them.](#evals)[0515 capabilities

Deploy
Govern every release through the pipeline you already run.](#deploy)

Run

[0623 capabilities

Agent Gateway
Decide who can reach each agent, and stop it when you need to.](#agent-gateway)[0731 capabilities

LLM Gateway
One governed door to every model.](#llm-gateway)[0824 capabilities

Guardrails
Check every prompt, answer and tool call.](#guardrails)[0924 capabilities

Observability
See every run, what it did and what it cost.](#observability)

Improve

[108 capabilities

Refine
Turn production evidence into better agents.](#refine)

[AlongsideOpencontroller Edge (formerly Basecode)The same control for coding agents on developer laptops.19 capabilities](#edge)
[UnderneathFoundationsRuns in your cloud, keeps secrets in your vault, audits every change.16 capabilities](#foundations)
[NextRoadmapTen major releases, all by the end of 2026.10 releases](#roadmap)

![](data:image/webp;base64...)

Follow one agent

At every stop we follow one illustrative agent: an invoice assistant that a finance team built in Microsoft Copilot Studio. Its story shows what each block does in practice.

Where agents are built

* AWS
* Microsoft
* Google Cloud
* Databricks
* Snowflake
* Anthropic
* Salesforce
* SAP
* HubSpot
* Claude Enterprise
* OpenAI Enterprise
* + any agent platform

Read-only scan
every 5 min

One inventory and asset graph

* Agents
* Models
* Tools
* MCP servers
* Knowledge
* Identities
* API keys
* Licences

RegisteredOwned, approved, governed

Shadow AIFound, not yet on record

Posture findings mapped toOWASP LLM and Agentic · MITRE ATLAS · NIST AI RMF · ISO/IEC 42001 · EU AI Act · AWS Security Hub

### One inventory across rival clouds

AWS, Microsoft, Google Cloud, Databricks, Snowflake and Anthropic in one asset graph, with the models, tools, MCP servers and identities each agent uses.

### Shadow AI, surfaced

Anything nobody has registered shows up as shadow, together with the keys, licences and spend behind it.

### Posture built for agents

31 detectors flag exposure, risky identities, leaked secrets and missing controls, mapped to the frameworks your auditors use.

All 33 capabilities in Discover

#### Inventory 8

* Agents and everything they use: models, tools, MCP servers, knowledge, gateways and identities.
* See how each agent connects to its models, tools, guardrails and identities, with a map view.
* Every discovered agent, tool or model of a governed kind that nobody has registered.
* Discovery never writes to your cloud. Only a separate, opt-in control credential can change anything.
* Rescans every 5 minutes by default, with per-connection intervals and a live scan indicator.
* Each scan hashes every asset and records only what changed.
* Every connection is verified per surface, naming any permission that is missing.
* Setup, permission and troubleshooting guides for every platform.

#### Platforms 10

* Bedrock AgentCore runtimes, gateways and Agent Registry, classic Bedrock Agents, knowledge bases and guardrails, across regions.
* Copilot Studio, Agent Builder, the Agent 365 catalog, Azure AI Foundry agents and Entra agent identities.
* Agent Runtime, Managed Agents, Agent Registry and CX Agent Studio, at project, folder or organization scope.
* Serving endpoints, Agent Bricks, Genie, agent apps, external models, MCP services and Unity Catalog skills.
* Cortex Agents, managed and external MCP servers, Cortex Search, semantic views and Model Registry.
* Claude models in use, API keys and the models each one calls, workspaces and tokens.
* Agents built on the major business application platforms.
* Agents and usage inside enterprise AI plans.
* A connector framework for any platform that lets teams create agents.
* A cross-account read role with an external ID, created from a template. No stored keys.

#### Keys, licences and access 3

* Keys found per provider and linked to the models they call.
* Who holds which AI licence, when they last used it, and which seats sit idle.
* Resource policies read into access links, exposure level and blocked state.

#### Posture and risk 12

* Risk and permission detection over the whole discovered estate, run after every scan or policy change.
* Findings mapped to OWASP LLM and Agentic Top 10, MITRE ATLAS, NIST AI RMF, ISO/IEC 42001, the EU AI Act and AWS Security Hub.
* Agents with no sign-in, unauthenticated gateways, internet-reachable endpoints and org-wide sharing.
* Agents using their maker's credentials, shared identities, non-expiring or unused keys, keys made by leavers.
* Credentials in instructions or environment variables, and agents that mix outside input, private data and a way to act.
* Destructive tools, code interpreters, weak instructions and agent-to-agent calls.
* Unregistered MCP servers, missing TLS and models from providers you have not allowed.
* Agents with no guardrail or gateway, model calls that skip the gateway, agents nobody traces or evaluates.
* Shadow agents, missing or departed owners, dormant published agents.
* Severity rises when an agent is reachable, can act, or touches data.
* Category-by-cloud heatmap, framework coverage, most-at-risk assets and CSV export.
* A guide for each framework with live coverage of every item.

Found by a scanShadowDiscovered, no owner on record

Owner + purpose

RequestedPendingCan be rejected or withdrawn

A different approver

ApprovedRegisteredGoverned until revoked

Every transition records who, when and why. Rejecting or revoking needs a reason.

### Approvals with separation of duties

Requesters name an owner and a purpose. A different person approves, matched on their IdP identity.

### Register in the same click you block

Blocking a shadow agent opens its registration in the same dialog, so nothing is contained without an owner on record.

### Agents, tools, MCP servers and skills

One registry for every governed asset, including agents outside the clouds and A2A agents read from their card.

All 19 capabilities in Register

#### Agent registry 9

* Every registered asset with its owner, purpose, platform and account.
* Requesters state owner and purpose. Where the platform names an owner, it is locked in.
* Admins approve, reject or revoke from one page.
* Nobody approves their own request, matched by IdP subject rather than display name.
* A note is required to reject or revoke, and optional to approve.
* Pending, registered, rejected, revoked or withdrawn, with who, when and why.
* The requester or an admin can withdraw a pending request.
* Choose which asset kinds are governed. Each registration records the policy version it met.
* Control actions run only on registered, approved assets.

#### Beyond the clouds 3

* Register agents outside every cloud, over OpenAI-compatible or HTTPS endpoints, with no credentials in the URL.
* Reads skills, auth schemes, version and provider from the A2A agent card.
* Batch and external registration from pipelines and service accounts.

#### Ownership 2

* Teams own assets by provider, account, environment or region. The narrowest rule wins.
* Summary cards, an estate report and overview shortcuts for leadership.

#### Tools, MCP servers and skills 5

* A central list of approved tools and MCP servers.
* Skills found across the estate start as shadow until they are registered.
* Skills are scanned for vulnerabilities and leaked credentials, and register only once findings are fixed.
* Requesters state a business purpose. Approvers can reject with a reason.
* Skills are managed and versioned centrally.

PersonSigns in with SSOOkta, Entra ID, Google and more

AgentOwn workload identityActs on the person's behalf

Gateway keyMapped from IdP claimsScoped models, tools and budget

Model or toolCall allowed or deniedRecorded against the caller

Every hop recorded The audit trail names the person, the agent and the key for every call.

### Your IdP, end to end

OIDC or SAML sign-in, SCIM provisioning and role mapping from IdP groups, for Okta, Entra ID, Google and more.

### Agents as principals

Each agent carries its own workload identity, and its actions are attributed to the person it acts for.

### Identity becomes the key

The gateway validates the caller's IdP token and maps its claims to a scoped key, so every call has a name.

All 23 capabilities in Identity

#### People 6

* Authorization Code with PKCE. Tokens are checked for signature, issuer, expiry and audience.
* Signed assertions with an HttpOnly session and CSRF protection.
* Okta, Entra ID, Google Workspace, Keycloak, Ping, OneLogin, ADFS and Auth0, each with a guide.
* Users and groups pushed by your IdP. Deprovisioning cuts access on the next request.
* Sign-ins from outside your domains are refused.
* argon2id passwords, lockout after failed attempts, rate limits and session lifetimes.

#### Roles and access 6

* Viewer, operator, admin and owner, with a permission on every route.
* Your own roles with wildcards, mapped from IdP groups, local accounts or keys.
* Groups, app roles or SAML attributes map to roles, with a default role.
* Nobody grants more than their own role, and at least one owner always remains.
* Teams see only their own assets, registry, traces and costs.
* Every API route and the permission it needs, published and pinned.

#### Agents as principals 5

* Agents as first-class principals with federated workload identity.
* Agent actions attributed to the person they act for.
* The gateway checks the caller's IdP token and maps its claims to a virtual key.
* Keys that require sign-in record each caller on every call.
* Google Cloud and Snowflake connections trust your deployment's own identity. No stored keys.

#### Sharing and audit 6

* Share and revoke on Copilot Studio, Agent 365, Foundry, AgentCore, Google Agent Runtime and Snowflake.
* Removes everyone and public grants while keeping named people.
* Control actions carry the operator's identity into AWS CloudTrail.
* OAuth client-credential machines that are never owners, audited as services.
* Role-scoped keys for automation and break-glass access.
* Every sign-in, failure, role, rule, key and SSO change, with actor and outcome.

Answer correctness, judged weekly on live trafficIllustrative

### Judges you can check

LLM judges and zero-cost code evaluators, with a grader-agreement check against human labels.

### Regressions caught with statistics

Monitors compare runs with Welch tests and alert only when a drop is beyond chance.

### Golden sets from real traffic

Turn production traces into test items and compare a new version item by item against the baseline.

All 20 capabilities in Evals

#### Evaluators 8

* A template catalog and builder, with Anthropic and OpenAI-compatible judges.
* Sandboxed JavaScript checks with no model cost: exact match, keywords, regex, JSON and length.
* Each judge returns a score from 0 to 1 or a label.
* Judges answer through a tool that accepts only valid labels.
* Answer relevance, correctness against retrieved context, and more.
* Whole conversations evaluated, beyond single turns.
* Adversarial probes for injection, bias and privacy leaks.
* Judges checked against human labels with precision, recall and Cohen's kappa.

#### Data and experiments 5

* Capture from traces, bulk-add expected answers, import CSV, TSV, JSON or JSONL, edit as a grid.
* Item-by-item comparison of a new version with the current one.
* Test a new version on the agent's latest real traffic.
* Versioned prompts with notes, tested on rows or typed text, with rollback.
* Every judgment across tasks and monitors, with filters you can share.

#### In production 7

* Production traces sampled on a schedule and judged.
* Welch tests with Bonferroni correction. Alerts fire only past chance.
* Errors, latency, tokens, cost, loops and retries, grouped into one incident.
* Webhook alerts to Slack or Teams, with a test alert.
* Token budgets that judges never overspend, counted per run and per day.
* Creates an evaluator over any agent's gateway traces.
* Ask proposes evaluators, rows and tasks. Nothing changes until a person accepts.

DevVersion 7Model, prompt, tools, parameters

Eval gate

StagingVersion 7Golden set and regression checks

Named approver

ProductionCanary: 10% on v790% stays on v6 until promoted

Rollback One step back to version 6. Separation of duties applies at every gate.

### Promotion with gates

Dev to staging to production, with eval floors, named approvers and separation of duties.

### Safe rollout

Canary and blue-green releases where the runtime supports them, and one-step rollback.

### Your CI, governed

Your pipeline builds, scans and signs. Opencontroller governs the promotion and keeps the evidence.

All 15 capabilities in Deploy

#### Release path 6

* Dev, staging and production for every agent.
* Every change mints a version: model, prompt, tools and parameters.
* Promotion waits on named approvers.
* Authors cannot approve their own promotion.
* Approvers approve, reject with a reason or request more information.
* Approve releases from coding agents or any MCP client.

#### Gates and rollout 5

* Releases stop on eval floors and regressions.
* Shift traffic to a new version gradually.
* Switch between two live versions.
* Return to the previous version.
* Pull a version or an agent out of service.

#### Your pipeline 4

* Auto-registration, eval gate and gateway access inside the workflow.
* The same hooks in the pipeline you already run.
* Skills move through the same governed pipeline.
* AWS Bedrock AgentCore, Google Vertex AI and Azure AI Foundry.

Native control actions, in each platform's own terms

| Platform | Block and unblock | Share and revoke | Delete with cascade |
| --- | --- | --- | --- |
| Microsoft Copilot Studio | ● | ● | ● |
| Microsoft Agent Builder and Agent 365 | ● | ● | ● |
| Azure AI Foundry | ● | ● | ● |
| AWS Bedrock AgentCore | ● | ● | – |
| Google Cloud Agent Runtime | – | ● | – |
| Snowflake Cortex Agents | ● | ● | – |

Every action needs a reason, is audited, and shows as pending until a rescan confirms it. Unblock restores the platform's policy exactly.

### Native control, platform by platform

Block, share, revoke and delete in each platform's own terms. Every action is reasoned, audited and confirmed by rescan.

### Behind the gateway in one step

The agent gets its own key, its model calls are repointed in its own cloud, and a test proves the call arrived.

### Tools under policy

MCP servers behind the gateway, with Cedar and OPA rules checked before a tool runs.

All 23 capabilities in Agent Gateway

#### Control actions 8

* Quarantine Copilot Studio, Agent Builder, Agent 365, Foundry, AgentCore and Snowflake agents. Reversible.
* Unblock puts the platform's own policy back exactly as it was.
* Removes Microsoft agents with their dependent flows and bot services, after name confirmation.
* Share and end public access for Agent Runtime, and control CX Agent Studio publishing.
* Control actions for Anthropic Claude assets from the same console.
* Every action records who and why, then shows pending until a rescan confirms it.
* Shows which actions apply now, and why the others do not.
* Microsoft admin tokens live in memory for their hour and are never stored.

#### Rogue agents 2

* Behaviour deviations trigger automatic quarantine or kill.
* Rogue-agent policies applied across the whole agent estate.

#### Agents behind the gateway 4

* A per-agent key, model calls repointed and a test call, with no code change.
* Restores the agent's previous model and deletes its key.
* Method, path, header, body and CEL policies in front of any service.
* Deny, access control, require auth, rate limit, transform and redact, with hot reload.

#### Tools and MCP 9

* MCP servers behind the gateway with per-server auth and health checks.
* Bundle tools into groups that keys and teams can use.
* Global, client, tool and request filters must all allow a tool.
* Allow or deny tools by key and identity claims, or strip a tool from the call.
* Rules checked before each tool runs, validated at save, with ready templates.
* Rego runs before a tool runs and on what it returns.
* Per-session and per-call MCP records.
* Multi-turn tool-calling loops, with a code mode.
* The console cannot add stdio MCP servers or code-running plugins.

Callers

* Apps
* Agents
* Coding agents

Identity

* Virtual key
* IdP claims
* Access profile

Policy

* Budget and rate limits
* Allow or deny
* Guardrails

Router

* Rules and semantic routes
* Classifier
* Complexity

Providers

* OpenAI
* Anthropic
* AWS Bedrock
* Azure
* Google
* + more

FallbackIf a provider fails, the next in the chain answers.

CacheSimilar prompts are answered from the semantic cache.

TraceEach allow or deny decision is written to the span.

### Every model, one endpoint

An OpenAI-compatible API across 12+ providers, with drop-in support for the SDKs your teams already use.

### Routing that pays for itself

Rules, semantic routes, classifiers and a complexity router send each request to the right model, with fallbacks and caching.

### Spend under control

Virtual keys, budgets, rate limits and cascading policies by key, team, customer and business unit.

All 31 capabilities in LLM Gateway

#### Access 8

* One endpoint in front of 12+ model providers.
* Works with the OpenAI, Anthropic, Bedrock, Google GenAI, LangChain, LiteLLM and PydanticAI SDKs.
* Chat, Responses, embeddings, images, speech, transcription, batch, files and token counts.
* Queue requests and collect results later.
* Scoped keys with model and provider access and rotation. The value is shown once.
* Reusable bundles of model and provider access.
* Users request access and admins approve.
* Keys, budgets, teams, providers, calls, usage and cost in one console, every change recorded.

#### Routing and resilience 8

* Provider fallback chains with configurable backoff.
* Weighted and adaptive balancing across keys and providers.
* Reuse answers to similar prompts, on Weaviate, Qdrant, Redis or Pinecone.
* Route by model, provider or request attributes.
* Route by what the prompt is about.
* A classifier picks the route, with a playground to test it.
* Simple requests go to cheaper models.
* Analytics on every routing decision.

#### Spend and policy 8

* Spend caps per key or globally, reset on your schedule.
* Token and request limits per model, per key or globally.
* A hierarchy for spend and policy.
* Allow or deny by key, team or globally, applied in a fixed order.
* An OPA/Rego sidecar decides on each request.
* Trust levels per model, and quarantine for a model you no longer trust.
* Override prices and keep model metadata in one catalog.
* Versioned prompts deployed with traffic splits.

#### Operations 7

* Each request's allow or deny decision is written into its span.
* Filterable request logs with usage and cost export.
* Slack, webhook and PagerDuty alerts.
* Reports whether encryption is on, and can require it.
* Multi-node gateway clusters.
* Deploy in your VPC with Helm and Terraform.
* Available as an independent module, with or without the rest of Opencontroller.

Prompt inBefore the model sees it

* Prompt injection
* PII
* Secrets
* Banned topics

Answer outBefore the user sees it

* PII
* Toxicity
* Groundedness
* Valid SQL

Before a tool runsOn the tool call

* Cedar rules
* OPA policies
* Allowed tools

After a tool returnsOn the tool result

* OPA policies
* Secrets
* PII

The same guardrail runs through the LLM Gateway, the Guardrails API or Claude Code hooks.

### Four hooks

Checks run on the prompt, on the answer, before a tool runs and after it returns.

### Use the clouds' own guardrails

AWS Bedrock Guardrails, Azure AI Content Safety and Google Model Armor, called side by side from one guardrail.

### Policy as code

OPA/Rego and Cedar rules sit next to built-in checks for injection, PII, secrets and content safety.

All 24 capabilities in Guardrails

#### Built-in checks 9

* Model-based scoring of injection attempts.
* Detects personal data and redacts or blocks it.
* Masks keys and credentials in prompts and answers.
* Toxicity, sexual content and gibberish.
* Zero-shot topic classification.
* Block or redact by keyword.
* Structural checks on what the model returns.
* Answers checked against their sources at runtime.
* Checks that the agent stays on its assigned task.

#### Policy as code 2

* Custom policies, validated by a dry run before they go live.
* Allow only listed tools or block a named tool, from templates.

#### Providers you already use 4

* AWS Bedrock Guardrails, Azure AI Content Safety with Prompt Shields, and Google Model Armor.
* Lakera, OpenAI moderation, Zscaler AI Guard, Azure Content Safety, Bedrock or your own webhook.
* Redaction on requests and responses at the gateway.
* Provider credentials linked to Integrations, with write-only keys.

#### Run them anywhere 9

* Built and managed in Opencontroller and run inside your own deployment.
* Every guardrail syncs to the gateway and is repaired every 5 minutes.
* Answers are checked with answer checks, prompts with prompt checks.
* One endpoint for all four hooks, in OpenAI or Anthropic format.
* curl, Python and TypeScript examples with the guardrail ID filled in.
* Managed settings check prompts and tool calls for the whole organization.
* Templates for data protection, threat protection, moderation and custom rules.
* Try settings on a prompt or answer and see each check's result.
* Blocks and redactions per user, type and day. No text is stored.

One run, traced end to endIllustrative

Model callTool callGuardrail

### Traces with context

OpenTelemetry traces kept in your own deployment and joined to the agent, key, user and cost.

### Spend across every bill

Cloud invoices, Copilot licences and idle seats, Claude usage and gateway calls in one view.

### Anomalies flagged early

Spend spikes and behaviour deviations raise alerts before the month closes.

All 24 capabilities in Observability

#### Traces 14

* OTLP traces signed per service account and stamped so cost cannot be misattributed.
* Reads both span vocabularies.
* Traces stay in a store inside your own cell, read through a read-only user.
* A query language, saved views and a view of where the time went.
* A session shown as a chat beside the trace of each turn.
* Failures grouped by project, step and error.
* Average, P90 and P95, with heatmaps and split views.
* By provider, user, customer, key, tool and MCP server.
* Prompts and answers visible only to roles you allow.
* 18 integrations, a key and a test trace.
* Gateway spans filed under the agent whose key made them.
* OTLP export to Datadog, New Relic, Grafana, Arize Phoenix and ClickHouse.
* Claude Code telemetry, read in its own terms.
* Live traces monitored for deviation, with alerts.

#### Cost 10

* One cost page across every project.
* Cost per agent and per version.
* Azure Cost Management, AWS Cost Explorer, BigQuery billing, Databricks and Anthropic in one view.
* A monthly forecast compared with the previous period.
* Microsoft 365 AI licence spend and the seats nobody uses.
* Joined by email across licences, Claude Code and traces.
* Billed AI assets matched to their registry standing.
* Model mix, with estimated prompt-caching and batching savings.
* Flags any service running at three times its previous week.
* Usage attributed to each person.

1. Production traffic
2. Judges grade it
3. Failures clustered
4. New test items
5. Version vs baseline
6. Fine-tune or down-route

### Judges on live traffic

Scheduled monitors grade sampled production traces and keep regressions open until they recover.

### Failures become tests

Real questions from production become golden-set items for the next version.

### Smaller, cheaper, proven

Fine-tune smaller models and down-route simple requests, then prove the result against the baseline.

All 8 capabilities in Refine

#### The improvement loop 5

* Scheduled monitors grade sampled live traces.
* Regressions stay open until they recover, and recovery is reported.
* One item per real question, with expected answers seeded for review.
* Score a new version against the baseline on the same set.
* Prompt Hub versions tried on dataset rows before they ship.

#### Optimization 3

* A fine-tuning studio managed from Opencontroller, its console shown in place.
* Complexity and classifier routing in the LLM Gateway.
* Ask reads failed runs and names the common cause.

Alongside the ten blocks

## Opencontroller Edge

Formerly Basecode

The same control for the coding agents on every developer laptop.

One laptop, three layers

Coding agentsClaude Code, Codex, GitHub Copilot, GeminiOpencontroller Edge

* PromptsPolicy-checked
* Provider keysVirtual only
* MCP serversApproved list
* PackagesApproved stack
* Model callsRouted
* SpendHard ceilings

AppsIDEs, terminals, developer toolsYour MDM

DeviceEncryption, patches, complianceYour MDM

### Rolled out by IT

Packaged and pushed through the MDM you already run, scoped to the groups you already manage.

### One standard for every coding agent

Rules, skills and approved MCP servers live in a git repo and sync to every laptop.

### Keys and spend stay central

Virtual keys only, policy in the request path, and hard ceilings enforced before a request is sent.

All 19 capabilities in Opencontroller Edge

#### Rollout 5

* Packaged, scoped to groups and pushed by IT. Engineers do not have to install anything.
* Engineers can enroll themselves for a pilot.
* Claude Code, Codex, GitHub Copilot and Gemini, in any IDE.
* VS Code, Cursor and Windsurf.
* Opencontroller Edge adds the team's rules and skills, and only steps in when a policy applies.

#### Standards 2

* Rules, persona, knowledge and skills in the open OpenGAP format, synced to every laptop on change.
* Frameworks, MCP servers, skills and subagents per group. Agents decline anything off the list.

#### Policy and spend 5

* OPA or Cedar allows, denies or redacts before a model sees the request.
* Policy runs in the request path, on your Opencontroller Edge gateway.
* Coding agents never hold raw provider keys.
* Enforced before a request is sent. Models you have not allowed are refused.
* Simple edits go to cheaper models.

#### Context and visibility 5

* Sessions, tokens and spend by team, project, person and device.
* Decisions and project knowledge kept on your gateway.
* Move a session from Claude Code to Codex mid-task.
* Send a message to every terminal at once.
* Prompts, model calls and tool calls recorded per session.

#### Together with Opencontroller 2

* One managed-settings file brings traces and guardrails to every Claude Code user.
* Claude Code sessions in Opencontroller observability, with spend by person.

Underneath every block

## Foundations.

The platform runs inside your own cloud account, with your secrets in your vault and an audit trail that cannot be switched off.

![](data:image/webp;base64...)

### Runs in your cloud account

One isolated cell per organization on Kubernetes, with its own database, trace store, vault, gateway and guardrail engine.

![](data:image/webp;base64...)

### Secrets in your own vault

HashiCorp Vault, AWS Secrets Manager, Azure Key Vault or Google Secret Manager. Write-only in the console.

![](data:image/webp;base64...)

### Modular by profile

Start with discovery, add governance, or run the full platform. Modules switch on without a fork.

![](data:image/webp;base64...)

### Ask, the built-in assistant

Answers questions across inventory, findings, spend and traces, steers the console and works inside your editor.

More in the foundations

#### Platform 12

* Every change and every attempt is recorded. It cannot be switched off.
* Every portal action and sign-in, with who, what, which organization and from where.
* Operate Opencontroller from Claude Code and other MCP clients, with a Claude Code plugin.
* Every change Ask proposes waits for a person to accept it.
* OpenTofu modules and GitOps with Argo CD.
* Own namespace, database and host per organization, with network policies.
* Every route documented, with the route contract pinned in CI.
* RFC 9457 problem details with request IDs.
* OpenTelemetry traces and logs, Prometheus metrics, health and readiness checks.
* Gateway, evals and fine-tuning consoles inside Opencontroller, role-checked.
* Module calls never reach link-local or cloud metadata addresses.
* A ready sandbox with every module connected from the start.

Roadmap

## Ten major releases, by the end of 2026.

Every item below is planned for delivery in Q4 2026, before 31 December. Each one extends a block you have just read about.

OctNovDecToday, 27 October**All ten delivered by 31 December 2026**

01DiscoverQ4 2026

### Deeper shadow AI discovery

Find AI wherever it hides, beyond the managed agent platforms.

* Agents running on Kubernetes, EC2, Lambda and Cloud Run
* Shadow AI in code repositories and personal accounts
* A data access map of what each agent can reach
* Onboarding for hundreds of accounts at once, and more platforms

02IdentityQ4 2026

### Zero-trust identity for every agent

A verifiable, least-privilege identity for every agent, including short-lived ones.

* Microsoft Entra Agent ID support
* Just-in-time secrets injected from your vault
* Identities for sub-agents and ephemeral agents
* Attribute-based access and delegated admin by business unit

03Evals and RefineQ4 2026

### Self-improving agents

Agents that get measurably better in production, with every change proven first.

* An improvement engine that recommends prompt, model and config fixes
* Closed-loop fine-tuning behind an eval gate
* Trajectory and tool-call evaluators for multi-agent workflows
* A/B tests and shadow canaries that weigh cost against accuracy

04Agent GatewayQ4 2026

### One governed front door for every agent

Every agent and tool call through one authenticated, policy-enforced entry point.

* A per-agent entry point that routes to approved versions
* Governed agent-to-agent calls, including third-party agents
* Existing MCP servers and REST APIs behind central auth and trust tiers
* Graduated containment, from isolating one tool to an estate-wide stop

05LLM GatewayQ4 2026

### Sovereign, policy-aware model routing

Sensitive data stays with the models you trust, with a receipt for every routing decision.

* Routing by sensitivity, PII or jurisdiction to sovereign or local models
* Cost- and latency-aware routing across vendors
* Your self-hosted model endpoints as providers
* Bring-your-own gateway, managed from Opencontroller

06GuardrailsQ4 2026

### Next-generation threat protection

Defence against attacks hidden in documents, tool output and other agents' messages.

* Indirect prompt-injection defence
* A purpose-built guard model
* Multilingual PII detection, and guardrails for images and files
* More bring-your-own providers, including Presidio

07ObservabilityQ4 2026

### CFO-grade spend and chargeback

AI spend turned into accountable, forecastable cost for every business unit.

* Chargeback and showback by business unit, team, cost centre and agent
* Budgets at agent, user, team and model level, with 60, 80 and 100% alerts
* A finance view that answers why cost went up, with ROI
* Spend from more sources, including Copilot Studio

08PlatformQ4 2026

### Deploy anywhere, even air-gapped

Run Opencontroller in your own cloud, across many accounts, or fully disconnected.

* Customer-hosted installs on AWS, Azure and Google Cloud
* Air-gapped mode with local models
* Hub-and-spoke control across many accounts and VPCs
* Listings on the cloud marketplaces

09Posture and policyQ4 2026

### Regulatory policy packs and evidence

Maintained policy packs and tamper-evident evidence for the rules you report against.

* Packs that produce evidence for GDPR, the EU AI Act, HIPAA and DORA
* A write-once, tamper-evident evidence store with full export
* A replayable record of every decision across sub-agents
* Policy inheritance from organization to business unit to agent

10Opencontroller EdgeQ4 2026

### Workforce and developer AI governance

Govern the AI tools people use every day, from Copilot to coding agents, with Opencontroller Edge (formerly Basecode).

* Opencontroller Edge fleet, devices and spend inside the main Opencontroller console
* GitHub Copilot, Microsoft 365 Copilot and Claude Cowork traffic through the gateway with per-user limits
* A browser extension for PII and unsanctioned AI chat
* Self-hosted Opencontroller Edge, entirely inside your network

## See it on your own agents.

Book a demo and we will walk through every block on the platforms you run today.

[Book a demo](https://www.lyzr.ai/book-demo/)

Or write toopencontroller@lyzr.com

## Your privacy choices

With your permission, we and our service providers use cookies and similar technologies to recognize returning visitors, measure how the site is used, and support marketing. Essential website functions remain available if you decline. [Read our privacy policy.](/privacy)
