Cursor ships hosting and agent security
Your AI dev stack just got more hands-on. Today’s batch brings a new home for Cursor projects, security testing that fixes itself, and sharper ways to manage agent cost and voice operations.
In today’s edition
- Cursor launched Origin, a code hosting platform integrated with Cursor.
- ElevenLabs brought voice and chat agent management into Claude via MCP.
- Replit added black-box penetration tests with one-click Agent fixes.
- OpenRouter cut GPT-5.6 Sol pricing through September 18.
- OpenRouter added a Cost per Session data layer for model comparisons.
Cursor launches Origin code hosting
Cursor’s Origin code hosting platform is live, with direct integration into Cursor and a starting path for syncing repositories from GitHub.
- Origin is positioned as a code hosting platform built around the Cursor workflow.
- You can get started by syncing existing repos from GitHub.
- The rollout begins in beta today.
- Vercel, Buildkite, and Depot integrations are already available, with more planned.
What it means Try Origin on a noncritical repository if your team wants fewer handoffs between AI-assisted coding, source control, and the services that build and deploy your app.
Source: cursor_ai
ElevenLabs puts agent operations in Claude
ElevenLabs MCP is now available in Claude, letting teams manage voice and chat agents from the place they already use for AI work.
- Review recent agent performance from Claude.
- Create agents and update their configurations.
- Estimate LLM costs before changes go live.
- Use OAuth sign-in without running a server or managing API keys.
What it means Give the people who own customer conversations a controlled Claude-based workflow for reviewing, changing, and cost-checking voice agents before production updates go live.
Source: ElevenLabs
Replit adds attacker-style app testing
Replit apps can now run black-box penetration tests that probe them the way an external attacker would, then hand fixes to Replit Agent.
- The scans test deployed Replit apps from the outside.
- Replit says Agent can fix findings in a single click.
- The company recommends pairing black-box scans with white-box tests that inspect code from the inside.
What it means Add attacker-style scans to your pre-launch checklist, then use Agent to speed remediation while keeping white-box tests in the loop for code-level coverage.
Source: Replit
GPT-5.6 Sol gets a temporary price cut
OpenRouter cut GPT-5.6 Sol pricing in half, including batch, flex, and priority tiers, with flex pricing as low as $1.25 input and $7.50 output.
- Use the model ID `openai/gpt-5.6-sol` to receive the discount automatically.
- The offer applies to the OpenAI provider and non-BYOK requests.
- The promotion runs through September 18.
- Batch, flex, and priority tiers are included.
What it means Run your expensive agent and evaluation workloads through GPT-5.6 Sol before September 18, then compare quality and total cost against your current production route.
Source: OpenRouter
OpenRouter tracks cost per session
OpenRouter added Cost per Session to its API and data layer, aiming to measure what models cost across real tasks and agent harnesses rather than token price alone.
- The metric covers real-world model cost across tasks and harnesses.
- OpenRouter says it uses medians across 30 days of tasks seen by each major harness.
- Its Dataset API can be queried to create your own charts.
- The data is licensed under CC BY 4.0.
What it means Stop choosing models on token rates alone and benchmark session-level cost for your own agent flow, where tool calls and token efficiency decide the actual bill.
Source: OpenRouter