Changelog
What's new and fixed in Multi-Router.
| Date | Type | Change |
|---|---|---|
| Fixed | Tools sent to Gemini models keep their full JSON Schema, so coding tools' tool calls work on Gemini. | |
| New | Codex Proxy and Gemini Proxy connect Codex CLI and Gemini CLI to Multi-Router. The Pi Proxy is removed: pi connects through an Adaptive Router. | |
| Improved | Model group members name a variant and region, and the same model can appear more than once. Custom models belong to an organization and are managed by its admins. | |
| New | Adaptive Router: one router serves many features and customers from a model group. Rules on the feature, customer or requested model pick the model, with fallback and weighted splits, and Traffic records the rule that chose it. Create and edit it in the dashboard or the API. | |
| Fixed | Guardrail redaction is more reliable on streamed answers, including emails split across chunks and nested filters. | |
| Fixed | Traffic records and stored payloads are written out before the server shuts down. | |
| New | Integrations: connect guardrail, response cache and telemetry services and attach them to routers. | |
| Fixed | Claude Proxy refuses member keys that are missing or belong to another organization. | |
| Fixed | Deleting a provider key in the dashboard is refused while a router still uses it. | |
| Fixed | Connect GitHub opens your configured GitHub App. | |
| Fixed | Budget reads in the API include the budget's live status again. | |
| New | OpenAI provider keys can carry an organization ID, sent as OpenAI-Organization on every request. | |
| Fixed | A provider key can't be deleted while an AI agent policy uses it. | |
| New | API: every resource carries a version. Send If-Match to avoid overwriting someone else's change. | |
| New | API: list AI agents across an account or organization, and retire them. |
| Date | Type | Change |
|---|---|---|
| New | API: copy or move a provider key between organizations you manage. | |
| New | API: create, list, cancel and resend invitations. | |
| New | API: create, update, rotate and delete provider keys. | |
| New | API: read the model catalog, with model summaries and details. | |
| New | API: create, list, update and delete organizations. | |
| New | API: manage SSO connections, verify domains and run a test sign-in. | |
| New | API: manage machine identity issuers and organization bindings. | |
| New | API: manage prompt repositories and their GitHub connection. | |
| New | Traffic: filter by Claude Proxy and Pi Proxy routers, in the dashboard and the API. | |
| Fixed | Traffic records the router type for Claude Proxy and Pi Proxy requests. | |
| New | API: list and read API tokens. | |
| New | API: update budgets and manage their rules and overrides. | |
| New | API: read Traffic with the same filters and figures as the dashboard. | |
| New | API: create, update and delete every router kind, with version checks on writes. | |
| New | API: manage model groups and their members. |
| Date | Type | Change |
|---|---|---|
| Improved | Dashboard pages have stable URLs you can bookmark and share. | |
| Fixed | OpenAI reasoning models accept max_tokens on the passthrough route. | |
| New | PDF documents in requests, and passthrough when the request and router use the same API format. | |
| New | OpenAI file search calls are priced at the catalog fee. | |
| Fixed | Pi Proxy requests to OpenAI are priced, including their web searches. | |
| New | OpenAI service tier and cache write token counts are passed back in responses. | |
| New | OpenAI service tiers and 30-minute cache writes are priced, and custom models can set the 30-minute rate. | |
| New | Custom models: set tool fees in the price editor. | |
| Fixed | Streaming requests are billed for the input tokens reported in the final message. | |
| New | Anthropic fast mode and web searches are priced, and cancelled streams are labelled in Traffic. | |
| New | Long-context rates and 5-minute cache writes are priced, and every custom model price is editable. | |
| Fixed | Long streaming responses no longer drop after a stretch with no output. | |
| Improved | Simpler SCIM provisioning settings and token management. | |
| New | SCIM 2.0 provisioning: sync users and groups from your identity provider. | |
| Improved | Simpler budget panels, and spend charts with a dollar axis. |
| Date | Type | Change |
|---|---|---|
| New | Budget overrides: raise a limit for the rest of a period, request more budget from the Portal, and approve requests. | |
| Improved | Claude Proxy follows the Claude gateway compatibility specification. | |
| Fixed | One-hour cache writes are priced at the one-hour rate. | |
| Improved | Machine identity settings show every issuer, and the audience and host each workload calls. | |
| New | People management for each organization. | |
| New | Machine identity: CI jobs and Kubernetes workloads sign in with their own OIDC token instead of an API key. | |
| New | Account roster: every person in the account with their AI agents, and an activity log. | |
| Fixed | Router setup snippets show the API host. | |
| New | Portal: reveal an agent's key, copy its environment variables and open setup help. | |
| New | Person page: each person's access, AI agents, policies and budget. | |
| Fixed | Invited members can sign in to the dashboard. | |
| New | Traffic shows the full model coordinate: vendor, variant and region. | |
| New | Invitation emails and an invitation landing page. | |
| New | Invite people as operators or as managed users. Managed users get a tools policy when they accept. | |
| Improved | Budget detail: a details card, the creator, and hourly spend for daily budgets. |
| Date | Type | Change |
|---|---|---|
| Fixed | Policy changes apply to every agent already set up, including Claude Desktop. | |
| New | Redesigned New policy and policy pages, with inline editors for each AI agent. | |
| Improved | Budget form redesign: scope, period, a searchable model picker and an Enabled switch. | |
| New | Portal: budgets for each person's agents, per-person model caps, key rotation and Claude Desktop setup. | |
| New | API: /v1/budgets with idempotent create, list, read with status, and delete. | |
| New | Budget enforcement: Block rules stop requests at the cap. | |
| New | Custom models: clone a catalog offering with your negotiated prices, or add a model on your own endpoint. | |
| New | Tools policies set which AI agents each person can use, with which keys and budgets. | |
| New | Admin: configure email transport, branding and templates, and send a test. | |
| Fixed | Claude Proxy requests are priced by the resolved catalog model. | |
| New | People and the Portal: invite people, assign roles, and give employees a Portal for their AI agents. | |
| New | Budget rules: notify when spend reaches a percentage of a cap, or is projected to. | |
| New | Choose how a Static Router or Claude Proxy routes when you create it. | |
| New | Claude Proxy routes through catalog routers. | |
| Fixed | Claude Proxy uses the router's region, with the key's region as the default. |
| Date | Type | Change |
|---|---|---|
| New | Claude Proxy serves any Anthropic model name in the catalog. | |
| Improved | API: update routers with PATCH instead of replacing the whole router. | |
| New | Custom model aliases, and router aliases work on Claude Proxy. | |
| Improved | Catalog 5.0: models are addressed by catalog slug, and each offering has its own lifecycle. | |
| New | Claude Proxy on Vertex and Bedrock: model discovery and token counting. | |
| New | Offerings are priced and routed per variant: standard, global, geo and in-region. | |
| New | Claude Proxy model discovery. | |
| Improved | The model picker hides retired models by default. | |
| New | One shared model picker with catalog display names. | |
| New | Model groups: organization-owned groups of models, a Models page, and model governance for Claude Proxy. | |
| New | Pi Proxy: run the Pi coding agent through a router, including its Codex endpoint. | |
| Fixed | Routers refuse models their provider doesn't offer. | |
| Improved | The payload viewer decompresses compressed request and response bodies. | |
| New | Custom OpenAI-compatible providers on your own base URL. | |
| New | Region-aware Amazon Bedrock routing, with region checks and the x-mr-region header. |
| Date | Type | Change |
|---|---|---|
| Fixed | Fallback Routers move to the next member when retries run out or a member is misconfigured. | |
| Improved | Self-hosted servers can fetch new model catalog versions and stage them for an admin to approve. | |
| New | Budgets: set a spend cap for an organization, routers or tags, with a detail page and a daily spend chart. | |
| Improved | Your role in each organization shows on the Organizations page. | |
| Improved | One email-first page to sign in or sign up. | |
| New | Search Traffic by request, session or correlation ID, and send a correlation ID with each request. | |
| New | Enterprise single sign-on. | |
| Improved | Errors show whether they came from the caller, the provider or Multi-Router. | |
| Improved | Router search also matches models, providers, tags and aliases. | |
| New | Amazon Bedrock native endpoints and structured outputs. | |
| Improved | Request timings show the time to upload the request body apart from router overhead. | |
| Fixed | Switching organizations keeps you signed in. | |
| Improved | Traffic records the exact prompt revision and GitHub repository behind each request. | |
| New | The x-mr-router header picks a router on any request, and each router chooses whether the request's model or its own is used. | |
| New | Move a provider key to another organization. |
| Date | Type | Change |
|---|---|---|
| New | Storage modes choose which fields Traffic records for each request. | |
| New | Request storage: keep request and response bodies in your object storage, choose what each router captures, and open them from Traffic. | |
| New | API: list provider keys as metadata only; the key itself is never returned. | |
| Improved | A redesigned Models page with filters and columns, and a catalog that separates publishers from vendors. | |
| New | Claude Proxy: run Claude Code through a router, with usage recorded for each person. | |
| Improved | One model, vendor and key picker for creating and editing routers. | |
| New | Open any request in Traffic to see its route, timings, cost breakdown and error class. | |
| Improved | Request cost is split into input, output, cache read and cache write. | |
| Improved | Consistent pagination on the Routers, Prompts and Traffic tables. | |
| New | Customize any Traffic view's stats and charts, and save it as your own view. | |
| New | API: list, read, create, edit, enable, disable and delete routers, with idempotency keys on create. | |
| New | API: manage inline prompts and their revisions, and connect GitHub repositories as prompt sources. | |
| New | Anthropic Messages API at /v1/messages, with streaming, extended thinking and bound prompts. | |
| New | Traffic views for Overview, Latency, Errors, By model and Customers, each with its own stats and charts. | |
| New | Traffic: every request recorded with its router, model, tokens, cost and latency, in a live table with filters, saved views, stats and CSV export. |
| Date | Type | Change |
|---|---|---|
| Improved | A new Prompts page showing where each prompt is used, with tags, filters and a columns picker. | |
| New | Connect your own ClickHouse for Traffic, with a connection test and automatic schema setup. | |
| New | Email verification at sign-up. | |
| New | Fallback and Weighted Routers combine other routers: a fallback tries members in order on errors you choose, a weighted router splits traffic by weight. | |
| New | Tags for routers and prompts. | |
| New | A new Routers page with search, filters, saved views, sorting and 24-hour metrics for each router. | |
| New | Delete an empty organization, with a type-to-confirm step. | |
| Fixed | Password managers no longer fill the provider key form. | |
| New | Model presets on each router, with a Models tab. | |
| New | Admin pages for the database, the secret store and feature settings. | |
| New | Self-hosting: a first-boot install wizard checks the secret store, connects the database and creates the first admin. | |
| New | Server admins can deactivate, activate, delete, promote and demote users. | |
| New | Sign in with Google or GitHub. | |
| New | Aliases: give a router more names, and move an alias from one router to another. | |
| New | Bind a prompt to a router; it is applied to every chat request the router serves. |
| Date | Type | Change |
|---|---|---|
| New | Inline prompts stored in Multi-Router, with a structured editor for chat prompts. | |
| New | Prompts from GitHub: connect an organization and a repository, and prompt files sync on every push with their history. | |
| New | Invite people to an account, manage their access, and accept invitations from a landing page. | |
| New | In-app notifications, with a bell in the top navigation. | |
| New | Organizations: create, rename and switch between organizations inside an account. |
Start for free
Create a router, set your SDK base URL to Multi-Router and send a request.