One managed route
Use the same authorized model choices and team policy across supported assistants, rather than distributing upstream credentials to each tool.
Resources / Implementation guide
How Cortega manages coding agents
A practical guide to assistant configuration, model authorization, telemetry, project costs and the boundaries between gateway and endpoint controls.
Connection and control paths
Cortega connects supported coding assistants to approved models and provides separate controls for local AI applications. The AI Booster Pack configures clients; AI Border Gateway enforces policy on routed requests; Endpoint Guard covers configured applications on enrolled Windows and macOS devices.
Internal callersEmployees · service identities
Attacker entry pointsExternal messages · pages · attachments · tool output
↓ Task and context
Coding assistant on a developer’s deviceAI Booster Pack configures supported connections. Endpoint Guard applies configured device policies.
↓ Separate model and tool requests
Cortega AI Border GatewayIdentity · authorized models and MCP tools · guardrails · routed budgets · records
↓ Allowed requests only
Approved LLM routesPrivate inference · hosted providers
Governed toolsMCP tools · destination permissions
Independent usage collection
Claude Code and Codex can report usage while developers retain their existing subscriptions. Collection by itself does not route inference or enforce a gateway budget.
Project attribution
Associate governed usage with explicit project selections or reviewed session assignments. Keep unmatched usage and unpriced models visible.
Choose what to manage
| Control | Configuration | What it changes |
|---|---|---|
| Model routing | Route AI assistant traffic through AIBG | Supported requests use a Cortega identity and the team’s model authorization. Requests sent directly to a provider bypass this route. |
| Usage collection | Collect AI assistant usage | Claude Code and Codex send sessions, token counts and cost information. This is separate from model routing. |
| Message content | Collect prompts, responses and tool use | Enables additional text collection supported by the assistant. Codex does not supply response text through this telemetry path. Apply retention and access policy. |
| Repository inventory | Collect repository inventory | Collects repository metadata such as remotes, branch and working-tree state. It does not upload source files. |
| Device policy | Endpoint Guard application rules | Observe, police or block configured application traffic. This is a separate coverage path from a configured model connection. |
In the product, these are separate switches. Decide which combination is appropriate for each team; do not treat a token report as proof that a request was governed inline.
Assistant coverage
| Assistant | Managed connection | Operational detail |
|---|---|---|
| Claude Code | Shared CLI and editor gateway settings | Restart or open a new session after synchronization. Keep existing local permissions and hooks under review. |
| Codex | Custom provider shared by supported clients | Verify the selected endpoint’s Responses API, streaming and tool-call compatibility. |
| Continue | Managed chat, edit and apply profile | Select the managed profile; project or cloud profiles may use another route. |
| Cursor | Saved BYOK endpoint, key and model selection | Quit before Sync and reopen afterward. BYOK requests still pass through Cursor’s servers; private inference does not make the whole client local. |
| OpenCode / Pi | Provider registration and model defaults | Separate profiles, environment settings or explicit client choices can override defaults. |
| Goose / Aider | Custom endpoint and supported model settings | Test streaming, auxiliary model choices and tool behavior for the client version being deployed. |
AI Booster Pack configures installed tools; it does not install every standalone assistant or govern all editor features. Configuration support and runtime API compatibility are separate checks.
Roll out a governed configuration
Prepare the team policy
Register approved models, assign the development team’s authorization, set budgets and choose relevant guardrails. Test the endpoint contract needed by each assistant.
Sign in and discover tools
Use AI Booster Pack in desktop VS Code on Windows or macOS. Set the organization’s Cortega application URL and sign in. Installed editor extensions and supported standalone installations can be discovered.
Synchronize and restart
Close Cursor before synchronization. Choose router or direct-model behavior as appropriate, run Sync, then start fresh client sessions. Check the per-tool status rather than assuming every adapter completed.
Verify on the gateway
Submit a harmless coding request and confirm identity, team, requested model, actual serving model and policy decisions. Test an unauthorized model and the configured budget boundary.
Review overrides and restoration
Check project profiles, command-line flags and environment overrides. Test Undo in a pilot so the team knows how settings are restored and how later user edits are handled.
Projects, sensitive work and costs
| Situation | Cortega behavior | Review point |
|---|---|---|
| Developer selects a project | The server-held selection is associated with subsequent governed requests. | Explicit selections take precedence over inferred conversation assignment. |
| Project is marked sensitive | The selected project puts requests into the team’s Sensitive model choice path. | Configure suitable Sensitive choices; test the refusal when no eligible route is available. |
| No explicit selection | The optional Project agent can classify captured sessions from project descriptions and conversation context. | Review uncertain suggestions. A session assignment does not change unrelated work. |
| Routing is off; telemetry is on | Usage estimates can still be visible for supported assistants. | Codex costs are list-price estimates, not subscription invoices. Delayed telemetry attribution has timing limits. |
| No price is available | The report exposes unpriced usage. | Do not treat missing prices as zero cost or use them unreviewed for billing. |
Projects can be managed through the shared Projects MCP integration in Claude Code and Codex, subject to user permissions. Project writes are audit-logged. The classifier is optional; explicit project reporting does not depend on enabling it.
Operating checks and failure handling
The AI Assistants routing switch also controls access for this tenant’s device assistant keys: turning routing off revokes those keys, and turning it back on reactivates them. Verify a new gateway request after changing it. This stops the governed connection; it does not terminate the local coding assistant or disable an independent provider subscription.
A tool still reaches its old provider
Inspect its active profile, environment overrides and resumed session state. Re-sync, restart and verify a new gateway record.
Tokens appear but no gateway record exists
Check whether this is collected usage from a subscription-backed tool. Telemetry is an observation path; routing is a separate setting.
A sensitive task has no route
Review the team’s Sensitive choices and the serving model’s capabilities. Do not silently fall back to a standard model.
Costs do not match an invoice
Separate governed provider consumption from telemetry estimates, subscriptions and tool charges. Check missing prices and delayed exports.
Use a small pilot covering each client and operating system. A successful model response does not establish that every editor feature or local command is covered. Local execution permissions remain part of the assistant and device configuration.
Review the deployment with Cortega.
Bring the tools, models, access requirements and operating constraints for your team.