v1.3.7

Added

• Projects can now be listed and managed through the Management API and MCP server, matching what was already available elsewhere.
• Proxy responses now include an X-MVGC-Upstream-Path header showing the exact path a request was forwarded to, making provider-side errors easier to diagnose.
• Gemini requests now support both Google's native API and its OpenAI-compatible endpoint, with the correct authentication automatically selected for each.
• Policy conditions now support not_regex, the negated form of regex, for excluding matches by pattern.
• Qwen's US (Virginia) endpoint is now supported for routing and pricing.

Security

• Alert channel webhook URLs are now masked in API and console responses after creation, matching how connector credentials are already protected.

Fixed

• Self-hosted gateways connected to the Axemere control plane now clearly reject attempts to edit base provider pricing directly, instead of silently accepting the change and having it overwritten by the next pricing sync.
• Management API keys and MCP keys created through the console now actually work for authentication — previously a newly created key always failed to authenticate against the Management API or MCP server.
• Policy bundle load failures no longer expose internal error details in denial reasons; you'll see a clear, actionable message instead.
• Qwen credentials using a region-specific workspace endpoint now resolve pricing correctly instead of being left unpriced.
• Credentials with a region or custom base URL override now actually apply when the gateway routes a request, instead of being silently ignored after being saved.
• AWS Bedrock requests can now use either SigV4 credentials (access key/secret) or a Bedrock API key, instead of only SigV4.
• Zhipu (GLM) requests routed through its api.z.ai endpoint are now recognized and priced correctly, instead of being rejected as an unrecognized host.
• Together AI's routing policy now points at its correct API host (api.together.ai) instead of a retired domain (api.together.xyz).

Removed

• The ByteDance/Doubao (Volcengine Ark) provider has been removed; it's no longer available to configure.

v1.1.16

Added

• Automation Keys can now be renamed and given a description after creation, matching Gateway Keys.

Fixed

• Budget caps on a managed gateway now enforce consistently across every scope and time period (org/project/workload × daily/weekly/monthly/etc.) instead of only some combinations.
• Report-ready notifications (Chat, email) no longer include a direct download link that works without logging in; they now link to the console, which requires sign-in and verifies you belong to the report's organization.
• Report downloads are now blocked once a report's retention period has passed, instead of remaining downloadable until cleanup ran.

Changed

• Report-ready notifications now show a "View Report" button/link instead of a raw URL.

v1.1.12

Added

• The Executive Summary report now includes Spend by Workload and Spend by Model breakdowns, alongside the existing Spend by Project and Spend by Provider sections.

Fixed

• On-demand Executive Summary reports can now use "Previous billing period" as their date range; this option previously only worked for recurring billing-period-cadence schedules.

v1.1.11

Fixed

• Attribution summaries no longer silently exclude spend that has no assigned project, account, customer, or other attribution dimension — it now appears in an explicit "(none)" bucket, matching Spend Breakdown, so both pages' totals agree.

v1.1.10

Fixed

• Streaming completions now report accurate token usage and cost across all providers (OpenAI, Azure OpenAI, Anthropic, Bedrock, Vertex AI, Gemini, Cohere) instead of showing zero even on success.
• The Budget page no longer shows $0 spend for a project or org that has real usage when the analytics pipeline is still catching up.
• A gateway node whose control-plane certificate was due for renewal (but not yet expired) would fail to start until an operator manually re-enrolled it. It now keeps using the still-valid certificate while renewing in the background, as intended.
• Self-hosted operators authenticating with an admin token could no longer view or update policies; token-based access to the policies endpoint is restored.
• A request denied for exceeding your quota now says so specifically, instead of showing the same generic "denied by policy" message as an unrelated policy block.
• The gateway now finishes in-flight requests before shutting down on a restart or redeploy, instead of dropping active connections immediately.
• Self-hosted orgs running more than one gateway node no longer see undercounted spend breakdowns, attribution summaries, budget usage, billing, model efficiency, or discovered-domain reports — these now reflect traffic from every node instead of only the node handling the request.

v1.1.5

Fixed

• Anthropic connector no longer reports negative when the prompt cache is warm.
• Gateway now accepts requests with a missing , filling it from workload context or node defaults.
• OpenAI streaming now returns content correctly when is set via the gateway API.

v1.1.0

Added

• AWS Bedrock connector: The gateway can now route requests to Amazon Bedrock using AWS SigV4 request signing. Configure a Bedrock credential in the console with an AWS access key and secret, choose your target region, and the gateway handles authentication and routes to Bedrock's Converse API. Supports all models available via the Converse endpoint (Claude, Titan, Llama, Mistral, and more).

• GCP Vertex AI connector: The gateway can now route requests to Google Cloud Vertex AI using OAuth2 service account authentication. Configure a Vertex credential with a GCP service account JSON key, and the gateway handles token exchange and routes to the Vertex AI inference endpoint. Supports Gemini models and any other model available through your Vertex AI project.

• Multi-tier pricing rules engine: Provider pricing now supports tiered rates based on token count and request mode — for example, Gemini 1.5 Pro's context-size tiers (≤128K vs >128K input tokens) or OpenAI reasoning models that bill thinking tokens separately. All built-in providers (Anthropic, OpenAI, Gemini) ship with pre-configured rules covering cache tokens, reasoning tokens, and context-size tiers.

• ** field in action response:** Every response now includes a object with , , and . Streaming responses include metering as an SSE event immediately before the terminal signal. Client applications can display cost and token usage directly from the response without a separate records query.

• ** in pricing manifest:** now includes a field on each model entry when a multi-rule pricing config is available. Tooling can use this to pre-populate pricing editors with provider defaults.

• Per-connector request timeout: Custom connector definitions now support a field under to override the gateway's global connector timeout for that specific connector — useful for slow self-hosted models or high-latency regional endpoints.

• Attribution label filter on records endpoint: now accepts (repeatable, ANDed) to filter execution records by attribution labels. For example returns only records whose contain both entries.

• SDK: field in : The Python SDK's now accepts an parameter, allowing workload token configuration without environment variables.

Fixed

• Anthropic streaming reliability: Streaming requests submitted via the action params interface now correctly activate SSE on the Anthropic connector. Non-streaming and raw-body requests are unaffected.

• SIEM export delivery: Execution records are now reliably delivered to configured SIEM webhooks and syslog destinations. Previously, high-throughput conditions could cause export to be interrupted before delivery completed.

• Accurate HTTP status codes on upstream errors: The gateway now returns when a connector request times out and when the client disconnects — with a structured error body in both cases — making it straightforward to distinguish these conditions in logs and load balancers.

• Workload credential scoping after cache refresh: Workload-scoped credentials now maintain their scope restrictions correctly across cache refresh cycles, ensuring consistent enforcement as credentials are updated.

• Pricing markup saved on partial updates: Updating only the markup multiplier on an org pricing override now saves and appears correctly in pricing history. No pricing data is required alongside the markup to persist the change.

v1.0.7

Added

• MiniMax, Moonshot, and Zhipu provider integrations: These three providers can now be enabled in the console and used via the gateway — BYOK, credential injection, and path-prefix routing all work. Previously they appeared in the provider catalog but were missing from the routing registry, so every request failed silently.

• Custom provider connectors: Self-hosted gateways can now be configured to route to your own OpenAI-compatible endpoint — useful for privately-deployed models, regional AI infrastructure, or any API endpoint that speaks the OpenAI chat format. Add a custom connector in the console with a base URL and your choice of authentication (bearer token or client passthrough), and the gateway routes to it like any built-in provider.

Fixed

• Zhipu AI international endpoint: Zhipu AI requests now route to for reliable connectivity. The previous configuration targeted the China-mainland endpoint, which is unreachable from most regions.

• Zhipu AI documentation link: The Zhipu AI provider card in the console now links to the correct documentation page.

v1.0.6

Added

• Groq BYOK support: Client-supplied Groq API keys are now passed through correctly. Previously, only injected credentials were supported for Groq; the gateway would silently ignore any header the client provided.

Fixed

• DeepSeek credential injection: DeepSeek inline credentials configured in the console now inject correctly into outbound requests. A missing provider registry entry caused the auto-rule to be skipped at startup, leaving requests unauthenticated.
• Together AI provider integration: The Together AI provider integration now routes to the correct endpoint (). The previous configuration referenced the deprecated domain.
• BYOK keys silently ignored: When a provider integration was enabled alongside a BYOK credential, client-supplied API keys were being silently ignored and the injected credential was used instead. Passthrough rules now fire at the correct priority so client-supplied keys are honoured.

v1.0.5

Fixed

• Workload-scoped credential injection: Inline API keys scoped to specific workloads
  are now reliably selected by the gateway when injecting credentials on behalf of a request.

• Policy rules refresh after credential changes: The gateway now picks up updated
  auto-generated policy rules immediately when credentials are added or removed, without
  requiring a restart.