Skip to content
CortexFor AI coding agents

Load the Rules and Reasons Before You Act

Flash one Module. Get its declared dependencies, foundations first.Use Cortex when team knowledge can change your next action.

CLI

Flash one Module.

socra cortex flash <module-name>
MCP

Call context_flash.

context_flash { "target": "<module-name>" }
Returns

Target and declared dependencies, foundations first.

Then

Verify current state before acting.

Flash the Module, then verify current state

Cortex is pull-based. Call it only when Account knowledge can change the task.

  1. Read the Issue

    Use its Module name. For code, confirm the target repository. Issue access does not grant Module access.

  2. Flash the Module

    Call Cortex explicitly. Assignment and installation do not preload a session.

  3. Verify current state

    Inspect the repository, runtime, tests, and external sources. Cortex cannot inspect them.

  4. Act or expose the gap

    Follow the active instructions. Report missing or conflicting knowledge before acting.

Connect through the CLI or hosted MCP

Choose the path your agent supports. Both return the same declared knowledge.

CLI

Requires a terminal, the Socra CLI, an Account login, and access to Cortex in that Account.

socra install cortex
socra account login
socra cortex map
socra cortex flash <module-name>

Hosted MCP

Requires a remote HTTP MCP client with OAuth and Account access to Cortex. First-party plugins use Socra’s registered clients. Custom clients require a Cortex-enabled Cloud Project; follow the MCP reference.

https://cortex.socra.cloud/mcp
context_flash { "target": "<module-name>" }
Read the MCP reference
Identity
A normal MCP connection acts as the signed-in Account member. It uses a separate Agent principal only when one is created and connected explicitly.
Write surface
Hosted MCP can write and delete. Account permissions apply. Review its tools and the client's confirmation policy.
Empty Account
An empty Module map is valid. Continue real work and preserve useful specific knowledge as it emerges, following the human’s permissions and review preference. Do not create a test Module just to connect.
socra cortex module create --name <module-name> --content-file <path>

Know what Cortex will do before you call it

These are the current limits that can change whether Cortex is right for the task.

Good to know

Scope

A successful flash returns only the target and its declared dependency closure. It does not find missing knowledge or semantic conflicts. A stored cycle makes a later flash fail.

Access

The caller needs access to the target. Included dependencies need no separate access. Inclusion grants no later access.

Output

The CLI prints the assembled bundle, Flash ID, estimated token total, and Module count. MCP returns the bundle as tool text.

Record and replay

Each success records a metered Flash and its Module revisions. Reading the record preserves that manifest; reading its assembly rebuilds current knowledge without another Flash. Save the original output for exact replay.

Meter

Estimated tokens sum ceil(content.length / 4) once per Module in the closure. JavaScript length counts UTF-16 code units. Wrapper text is excluded.

Repository boundary

Cortex sees only data sent to its API. It does not inspect a repository or runtime, and it does not infer the target repository.

Authority

Cortex output is tool context. The host controls instruction precedence and tool confirmation. The team controls approval, and discovered instruction files still apply.

Cost

$20 per person monthly, with matching pooled usage credit that resets each cycle. Managed AI Teammates cost $3 monthly plus $1 per running hour, drawn from that credit. Bring your own supported AI subscription.

Access and saved copies

Access depends on Account entitlement and resource permissions. Save needed records and files before access ends using the documented per-record steps. Those copies do not form a complete Account backup or restore package.

Evidence

This contract supplies no controlled performance benchmark. Treat performance impact as unproven and run a controlled evaluation.

Compare repeated tasks with the model, settings, tools, task, and repository revision held constant. Check final state, evidence quality, tool calls, tokens, latency, cost, and failures.

Read full pricing

Technical contract reviewed 17 October 2026

Test Cortex on one decision that changes the work

Use one useful Module maintained under the owner’s review preference. Run the same task once with its flash and once without it. Keep the model, settings, tools, and repository revision fixed.