Coding Agent Capture, and Gateway Billing

What You Can Do Now

See what your coding agents are doing

• Capture Copilot across three surfaces: the CLI wrapper, the app login agent, and VS Code Chat.

• Capture the conversation from a plain codex run, not just the wrapper.

• Install a first-class LangWatch plugin for Claude Code, built on Agent Plugins 1.0, and it keeps itself up to date from every wrapped run.

• Review session economics, with the repositories and pull requests behind each session.

Bill and reconcile gateway spend

• A billing-grade spend pipeline with a webhook endpoints platform and a reconciliation surface, plus anchored cycles and idempotency keys.

• A budget now applies to the traffic it was scoped to, and each route states which key it requires.

• One filter vocabulary across both spend reads, and grouping that no longer double-counts.

• Routing-policy and cache-rule edits propagate through the change feed.

Self-host with your own identity and email

• AWS Cognito and OneLogin join the supported self-hosted SSO providers.

• One switch for email, with SMTP and Resend gateways, outbound proxy support, and a self-hosting page walking through setup.

• Sign-up recovers when the account already exists.

Turn traces into datasets

• A trace edit overlay, and an annotation-queue-to-dataset workflow, so review work becomes training or evaluation data.

• Oversized spans spool to S3 by default, so configured storage means no truncation.

• The Events column in Traces V2 reads real events, and a category with no result no longer shows as a passing zero.

Build against the API

• The REST API the SDKs actually call is now published as reference docs, with route coverage gated in CI.

Model providers

• Google Agent Platform is now part of the Gemini provider, reachable through either endpoint.

• Test a credential you have already saved, and an extra header no longer wipes stored Azure credentials.

Housekeeping you will feel

• Seven production defects found in a five-day error review, fixed.

• We stopped tracing every Redis command, which had been 93 percent of all spans.

• The legacy Ragas evaluators are gone.

• Seat changes no longer get blocked by a personal workspace, and refusals say what was refused.

Screenshots

Coding-agent sessions, with the repositories and pull requests behind them.

Repository capture in the sessions lens.

The billing events ledger.

Self-hosted SSO through Cognito.

The trace edit overlay with its annotations column.

REST API reference, published from what the SDKs call.

Notes

• Three app releases this week: 3.9.0, 3.10.0 and 3.11.0, plus python-sdk 1.2, typescript-sdk 1.5, mcp-server 2.0 and langevals 3.0.

• mcp-server 2.0 is a major: every request to the standalone HTTP server is now authorized.

• Performance: resident memory cut on the server, content-hashed assets can be served from a CDN, and prod entrypoints are bundled to run on node rather than tsx.