In Development · Register Interest

Memory the agent actually consults.

CxMS Agent is OpenCxMS's own AI coding agent. It runs its own loop, so “consult memory, verify, then respond” is enforced at the gate rather than suggested from the side.

Same Cortex memory engine as CxMS Pro AI. Different layer of enforcement.

Register InterestSee the three tiers

Why a separate agent

Every AI coding assistant on the market today runs the same default behavior loop: receive prompt, generate answer, act. Not: receive prompt, search memory, verify context, answer. That loop belongs to the model vendor and the host CLI, not to the memory layer bolted on top.

CxMS Pro AI pushes back on this as far as you can from outside the loop. Hooks fire synchronously through your host CLI's hook system. Sensitive writes get gated until source-of-truth has been verified. Context gets injected at session start. It's the strongest scaffolding possible without controlling the agent loop itself.

But the host's loop is still the host's loop. The agent can still ignore retrieved context in its reasoning. The gate blocks specific actions, not the underlying agent step.

CxMS Agent runs its own loop. When the agent is ours, the gate becomes enforcement, not suggestion.

How CxMS Agent works

A standalone AI coding agent built on the Cortex memory substrate. Python-primary. Three inference tiers. Same memory engine as CxMS Pro AI, in a different shape.

1

Memory consultation enforced

Before the agent acts on anything stakes-bearing, the loop checks the Cortex memory engine for relevant context, contradictions, and source verification. Not a hook on the side. A required step in the agent's own execution path.

2

Three inference tiers

Local Ollama for offline / cost-free. Hybrid for smart-routed cascading. Cloud (BYOK) for best raw capability on the hardest tasks. You choose, or let the Hybrid router choose for you.

3

Same Cortex engine

The Cortex memory database is shared with CxMS Pro AI. Same plain-markdown files, same KMAP refinement loop, same confidence tiers. Your memory follows you from your CLI sessions into Agent sessions.

Three inference tiers

Pick the tier that fits the work. The memory engine, hooks, and governance are the same across all three. Only the model layer changes.

Local

Runs on your machine

Backend
Ollama with Qwen3-Coder-30B (or other supported local models)
Cost
$0 per token
Best for
Privacy-sensitive code, offline work, cost-floor control
Tradeoff
Slower than cloud on the hardest tasks. Best for the 70-80% of work where local is plenty.

Hybrid

Smart-routed cascade

Backend
Routes between Local and Cloud based on task complexity
Cost
Mostly $0 (Local), Cloud calls billed via your BYOK key
Best for
Power users who want cost control without giving up cloud capability when it matters
Tradeoff
Routing decisions are conservative by default. Override per-session or per-task.

Cloud

BYOK across providers

Backend
Claude (Anthropic), OpenAI, Google, xAI, DeepSeek, Mistral, OpenRouter
Cost
Your BYOK provider rates only. We don't intermediate.
Best for
Hardest tasks, frontier reasoning, complex codebases
Tradeoff
Cost scales with usage. Internet required.

CxMS Pro AI vs CxMS Agent

Same memory engine. Different layer of intervention. Sibling products on the same substrate.

CapabilityCxMS Pro AICxMS Agent
Cortex memory engine + KMAPYesYes (shared DB)
Runs on your machineYesYes
Works with your existing AI coding CLI9 CLIs supportedStandalone (alternative to a CLI)
Layer of interventionHooks via host CLIOwns the loop
Memory consultationSuggested via context injectionEnforced at the gate
Write gatingYes (PreToolUse blocks)Yes (loop-native)
Local inferenceThrough host CLINative Ollama support
Smart-routed Hybrid inferenceNoYes
BYOK across multiple providersThrough host CLINative across 7+ providers
StatusShipping today ($49)In development

Most users will run CxMS Pro AI today and adopt CxMS Agent for projects where loop-level memory enforcement matters more than the host CLI's other features. The same Cortex database is shared between them.

Engineering choices that matter

Python-primary developer surface

`pip install cxms-agent`. The Pydantic AI framework adapter is the Tier 1 integration. Most agent developers already live in the Python ecosystem; Python adoption beats engine convenience. The compiled memory engine sits underneath.

One-time purchase + BYOK

No SaaS as the primary mechanic. No inference reselling. You buy the agent once. You bring your own API keys (Cloud tier). You own your memory, your code, and your spend.

Cortex memory shared with CxMS Pro AI

Plain-markdown memory files. Same KMAP continuous classifier refinement. Same confidence tiers. Same contradiction detection. Your memory follows you from Pro sessions into Agent sessions, and back.

Audit trail at the loop level

Every reasoning step, every memory consultation, every gate check is logged. Useful for governance, compliance (EU AI Act takes full effect August 2, 2026), and for managed-operations use cases where certified operators need full accountability.

Use cases

CxMS Agent is built coding-first. The same agent supports several adjacent applications because the memory + audit + governance substrate is the same.

Coding with enforced memory

The primary use case. Agent reasoning passes through memory consultation by design. Corrections stick because they have to. Decisions persist because the gate looks for them.

Privacy / offline work

Local tier with Ollama runs entirely on your machine. Air-gapped environments. Code that can't leave the network. Cost-sensitive work where token bills accumulate.

Cost-controlled AI development

Hybrid tier defaults to Local for the 70-80% of tasks where local is plenty and routes Cloud only when complexity warrants. BYOK pricing means your spend is what your provider charges, not what we mark up.

Multi-agent orchestration

Multiple CxMS Agent instances coordinating on a project. Cross-instance memory sharing through the shared Cortex substrate. For enterprises and complex engagements.

Remote AI-assisted operations

Certified OpenCxMS operators delivering managed services through the local agent. Context audits, governance deployment, forensic collection, incident response. Full audit trail because the agent runs locally with logged reasoning.

SASM micro-agent architecture

The same agent codebase ports to the SASM hardware safety module's micro-agent layer. One memory engine, two deployment targets (developer machine + safety-critical hardware).

Pricing

One-time purchase across all three tiers. BYOK for Cloud and Hybrid. No SaaS. Specific pricing locked when CxMS Agent reaches early access. Optional memberships (cloud sync, fleet coordination, priority support) coming as a fast-follow.

In Development
Register Interest

Early access opens when Cortex v4.0 + KMAP foundation work completes and the agent is ready for first beta users.

Pricing target: one-time purchase under $200 for the base agent across all three tiers. BYOK additional for Cloud and Hybrid usage. Final pricing locked at early access.

FAQ

How is CxMS Agent different from CxMS Pro AI?
Same Cortex memory engine. Different layer of intervention. CxMS Pro AI is scaffolding for your existing AI coding CLI (Claude Code, Cursor, etc.). It influences the agent through the host CLI's hook system. The host owns the loop. CxMS Agent is OpenCxMS's own coding agent. It owns the loop, so consult-memory, verify, then respond is enforced at the gate rather than suggested from the side. If you want memory the agent can't bypass, you eventually need an agent we control. That's CxMS Agent.
What are the three inference tiers?
Local (Ollama with Qwen3-Coder-30B) runs entirely on your machine. Zero per-token cost. Offline-capable. Best for privacy-sensitive work and cost control. Hybrid is a smart-routed cascade. It routes between local and cloud based on task complexity, defaults conservative, override-able per session. Cloud uses BYOK API keys (Claude, OpenAI, Google, xAI, DeepSeek, Mistral, OpenRouter) for best raw capability on the hardest tasks.
Do I need a separate API key?
Bring Your Own Key for the Cloud tier. The Local tier needs no key. Ollama runs models on your machine. The Hybrid tier routes intelligently and uses your BYOK key only when cloud inference is selected.
Why one-time purchase instead of subscription?
Same posture as CxMS Pro AI. You buy it, you own it, forever. We don't intermediate AI inference (BYOK across all tiers), so there's no SaaS billing motive baked in. Optional memberships are coming as a fast-follow for ongoing services like cloud sync, fleet coordination, and priority support. The agent itself is one-time.
What language is CxMS Agent built in?
Python-primary developer surface. `pip install cxms-agent`. The Pydantic AI framework adapter is the Tier 1 integration. The memory engine and IP-bearing components sit underneath as a compiled layer. Python adoption beats engine convenience.
When does CxMS Agent ship?
In development. Cortex v4.0 + KMAP, the memory substrate the Agent is built on, is the foundation work and partially shipping today in CxMS Pro AI. Register interest below to hear when early access opens.
Does CxMS Agent replace my existing AI coding CLI?
It's an alternative path, not a forced replacement. Many users will run CxMS Pro AI with their existing CLI (Claude Code, Cursor, etc.) AND eventually adopt CxMS Agent for projects where loop-level memory enforcement matters more than the host CLI's other features. The same Cortex memory database is shared between them.
What about remote operations and certified operators?
Higher-tier application of having your own agent. Managed operations where certified OpenCxMS operators handle context audits, governance deployment, forensic collection, and incident response remotely. The remote piece exists because CxMS Agent runs locally with full audit trails, so certified operators can deliver services through it without losing accountability. One capability among many, not the headline.

Register Interest

Be first when CxMS Agent opens early access

CxMS Agent is in development. The Cortex v4.0 memory substrate it's built on is partially shipping today in CxMS Pro AI. Register your interest and we'll notify you when early access opens.