Cortex
Cortex connects coding agents to the work around the code: persistent project workspaces, Docker Compose deployments, runtime diagnostics, and shared memory. Work from a terminal or a connected chat channel, then inspect jobs and services in Cortex Control. Policies, approvals, and audit are configurable parts of that workflow, not a claim that every session runs with the same protections.
Part 01 · 6×4 · seats on the baseplate · Engram, Gauntlet, OMP, Cortex Control seat on it
From a coding session to an operated service.
ShippingShipping. Local downloads, a live Cloud product site, and a public compliance library. Available surfaces depend on the deployment and plan.
Python · TypeScript · Docker Compose · Terraform · WorkOS · Stripe
Where it is useful
Continue work beyond a chat
Start a project from a template or an existing repository and return to the same workspace. Cortex Code wraps a native coding agent in your terminal; connected Telegram topics, Slack threads, and Discord channels provide remote entry points. Transcripts and artifacts make a long-running task inspectable without following every message.
Deploy the app you already have
Describe the repository, Compose services, routes, and health checks in .cortex/app.yml. Preview the deployment request, submit it to the control plane, and follow its job. Cortex manages existing software as well as agent-built projects; it does not require a new application framework.
Keep operating after the build
Inspect app routes, containers, health, and diagnostic events from the CLI or Cortex Control. Refresh a route or restart a service through the same operating surface. These are useful tools for maintaining an internal app or an experiment service, not just generating its first version.
What the operator does
For local coding, choose a model provider and run cortex code from the project directory. Cortex supplies the surrounding environment; OMP, Codex, Claude Code, or OpenCode supplies the coding-agent interface. OpenClaw handles channel routing and session lifecycle for connected chat workflows.
For a managed deployment, log in to the target instance, inspect its status, and review the project manifest. Run cortex deploy --dry-run before cortex deploy, then follow cortex jobs watch with the returned job ID. Check the resulting app and its health rather than treating job submission as proof that the service works.
An administrator also owns the configuration: account groups, runtime policies, credentials, enabled channels, and backups. Connecting a channel or installing a tool does not establish the right access policy on its own.
Governance is a posture you configure
Human account groups and agent runtime policies are separate controls. Operators can set capability grants, service scope, file roots, package rules, and approval requirements. Governed actions can be allowed, denied, or sent for approval, with decisions available for inspection.
The execution path matters. Cortex Code supports governed sessions, but an authorized local operator can explicitly select an unrestricted session without Cortex tool gates or per-tool audit. OpenClaw harness-level governance is separately opt-in and defaults off. Review the active policy and integration settings before relying on a boundary; neither the product name nor a benchmark result is a safety guarantee.
The parts and their boundaries
- MemoryEngramShared memory built from linked Markdown notes, source references, and a derived search index. Browse a graph or retrieve prior project context rather than reloading every transcript. Scheduled maintenance is configurable; model-driven distillation needs a configured gardener. The Control memory surface is currently Enterprise-only.
- BenchmarkGauntletThe evaluation system. It compares base and Cortex-wrapped coding harnesses on attack resistance, code quality, generative tasks, and whole-repository builds. Its public reports expose task counts, scoring methods, and individual results.
- Coding agentOMPThe coding agent pre-configured in the local download. Cortex Code integrates through its tool hooks and preserves its native interface. OMP is one supported harness, not another name for the Cortex control plane; its active protections depend on the launch path and policy.
- Control planeCortex ControlThe web operations surface for app and runtime inventory, deployment jobs, diagnostics, accounts, and policy configuration. What a person can see and operate depends on their account group and the stack’s entitlements.
- Runtime and compilerSemaThe lab’s neurosymbolic language and compiler research, also included in the local tool bundle. It is a separate runtime, not the implementation language every Cortex app must use.
Evidence, not a blanket guarantee
The Gauntlet record covers 96 runs. These are the lab’s own evaluations, not an independent certification. The public index includes dated reports, different task counts, rescored runs, and cases where the raw harness leads. Read the report for the task and harness you care about; headline improvements are not a promise about every workload.
The compliance library publishes twelve documents in three bands: management system, information security, and AI management. Its own distinction is the useful one: conformity-ready, not certified. The documents identify implemented controls, procedures, partial coverage, and open gaps; they do not replace evidence from operating a particular deployment.
Inspect the product surfaces
- Inside Cortex ControlScreenshots of diagnostics, deployments, containers, and support
- Engram memoryThe product’s memory workflow and graph interface
- Local, Cloud, and EnterpriseCurrent packaging and pricing on Cortex’s own site
Links & access
- Cortex CloudProduct site, pricing, and self-serve signup
- The Gauntlet benchmark96 runs; inspect the individual reports and methods
- Compliance libraryTwelve documents, three bands, publicly readable
- Field notes on agentic engineering
- Download Cortex CodeLocal coding tools; a Cloud account is optional
- Page
- 03 of 09
- Part
- 01
- Grade
- shipping