Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

Mastra changelog: what's new each month

Every stable Mastra release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 10 months covered; the current month updates daily.

November 2026 · so far

4 releases: 1.81.0 → 1.84.0

November brought significant API refinements across the platform, with a focus on cleaning up deprecated components and stabilizing core features. The month included breaking changes to observability, memory, and factory APIs alongside new capabilities for agent resumability and MCP server configuration.

Action needed
  • Metrics chart card API cleanup: removed ChartCard description/summary and several MetricsCard components.
  • Removed fluid-hover utilities (useFluidHover, FluidHoverHighlight) and CommandList.highlightClassName from playground-ui.
  • Removed runtime-importable subpaths @mastra/core/network/vNext and @mastra/core/telemetry/otel-vendor.
  • defineBoard() and MastraFactory boards/includeDefaultBoards are deprecated in favor of stored boards APIs.

Highlights

  • Agents now support resumable streams, allowing clients to reconnect after disconnection without losing message history.
  • New MCP server configuration options let you control toolsets, features, OAuth settings, and read-only tool serving.
  • Agents can now authenticate with websites via setCookies() on MastraBrowser, enabling access to login-protected content.
  • Factory storage now supports collection existence checks before table registration.
  • Channel agents can open binary file attachments that require authenticated downloads, such as Slack files.

New

  • Resumable streams for all agents, not just durable ones, with replay caches for client reconnection.
  • Knowledge alpha scope matching with exact single-segment matching and canonical encoding requirements.
  • Workflow recovery loop that restarts crashed default-engine workflow runs without waiting for another instance.
  • Span query sorting by model and provider fields.
  • OAuth discovery and token audience configuration for the built-in MCP server.
  • onDecision callback in ClassifierProcessor for writing guardrail decisions to audit logs.
  • collectionExists() method for Factory storage to check table existence before registration.
  • Agent scopes function can now narrow or replace a run's incoming scopes.

Improved

  • Chat SDK logs from channels now use your Mastra logger instead of console output with [chat-sdk] prefix.
  • Evented durable agents no longer stop when the API process shuts down; workers complete cleanup and free the thread.
  • Instruction file loading and tool-result instruction discovery no longer block the event loop.
  • Temporal steps now validate inputSchema and requestContextSchema before running, with validateInputs option to opt out.
  • Advanced trace queries now support requestId, serviceName, and experimentId filters with expanded query operators.
  • Agent durable option now controls only resumable streams, not checkpoint-backed execution.

Fixed

  • Evented durable agents stopping unexpectedly when the API process shuts down.
  • Missing behavior for instruction files from untrusted Git refs during loading.

October 2026

8 releases: 1.73.0 → 1.80.0

October brought major architectural refinements to Mastra, including session and memory management improvements, breaking changes to controller APIs, and enhanced observability features. The month focused on simplifying core concepts and improving reliability across agent execution, storage, and integrations.

Action needed
  • Durable streams: use finishReason === 'tripwire' and output.tripwire instead of separate tripwire chunk.
  • @mastra/connect: API renamed from integrations to providers; removed legacy connect() alias, environment() sandbox credential surface, and several resolver helpers.
  • Slack channel integration id changed from 'slack' to 'slack-channels' in channels() method.
  • AgentController model switching: removed modeId/scope; session.model.switch(modelId, options?) signature changed.

Highlights

  • Agent controller sessions are now keyed by sessionId instead of scope, simplifying multi-session per resource patterns.
  • Durable agents automatically recover from transient provider failures, assistant-prefill rejections, and history incompatibilities with no configuration needed.
  • Observational memory now supports group filtering and generation ordering for more flexible history retrieval.
  • MCP tools no longer require approval by default, streamlining tool discovery workflows.
  • Concurrent schema-based working memory updates on PostgreSQL no longer lose fields when updating simultaneously.

New

  • tool-call-resumed stream chunk emitted when suspended or approval-gated tool calls resume on regular and durable agents.
  • agent.getMessages() available in tool execution context for standard and durable agent loops.
  • StorageMemoryRef type allows stored agents to reference registered memory instances or hold inline configs.
  • Explicit reconcile step for channel installations whose connect flow finishes outside the app.
  • File attachments now supported in AgentController steer and followUp methods, matching sendMessage behavior.
  • Authenticated message authors in AgentController sessions with author tracking across sendMessage, queueMessage, and steer.
  • queueSignal() to Agent and AgentController Session for queuing typed signals behind active runs.
  • Memory can contribute error processors to agents via MastraMemory.getErrorProcessors().

Improved

  • Invalid tool-call IDs on Anthropic models are rewritten in outbound requests to prevent failures.
  • Agent run abort operations now complete immediately when abortSignal fires instead of waiting up to 10 seconds.
  • Chat channel agents now correctly respond to mentions of their display name after being renamed.
  • Slack-triggered Factory runs no longer fail with credential errors after server restart.
  • Repository-backed Slack sessions recover from failed GitHub authentication without requiring restart.
  • Factory sessions now have a single global thinking default instead of per-mode defaults.
  • Trace lists now include outputPreview and errorPreview alongside inputPreview for better observability.
  • processLLMRequest now receives the workspace of the step for better processor context.

Fixed

  • Durable agent crash recovery was re-running still-executing runs; recovery now skips runs with live claims.
  • Aborting an agent run was taking up to 10 seconds when provider reported low rate-limit tokens.
  • Slack-triggered Factory runs failing with credential errors after server restart.
  • Repository-backed Slack sessions could not recover from failed GitHub authentication attempts.
  • Concurrent schema-based working memory updates on PostgreSQL were losing fields.
  • Chat channel agents ignoring mentions of their current display name after being renamed.

September 2026

9 releases: 1.64.0 → 1.72.0

September brought major updates to observability, workflow capabilities, and UI components, with significant API changes across factory configuration, trace querying, and playground components. The month included a major MCP rewrite and new features for advanced trace aggregation, workflow building, and model selection.

Action needed
  • @mastra/factory: sandbox config must be a callback, not options object; workdir/maxSandboxes removed.
  • @mastra/factory: defineBoard() phases require kind, working phases require role, initialPhase must be resting.
  • @mastra/factory: global rules object removed; use MastraFactory({ rules }) instead.
  • subscribeQueuedMessages replaced with subscribeThreadEvents (listener receives typed events).

Highlights

  • Studio Workflow Builder backend enables visual editing and persistence of workflow definitions.
  • Advanced trace queries now support aggregation, thread grouping, filtering by span names and outcomes, with deterministic cursor pagination.
  • ModelSelectionProcessor lets agents pick different models per request based on a classifier, routing simple requests to cheaper models while keeping a capable default.
  • MCP 2.0 major rewrite with 2026-07-28 protocol support and new suspend/resume APIs replaces legacy handshake.
  • Observability storage gains trace deletion, feedback review status tracking, and per-endpoint discovery features.

New

  • Studio Workflow Builder backend for persisted, visually-authored workflow definitions.
  • Classifier registration and classifier steps with typed answers and token usage in workflows.
  • ModelSelectionProcessor to dynamically select models per request with fallback handling.
  • Advanced trace aggregation via planTraceAggregate() and aggregateTraces() with grouping and measures.
  • Thread query contract (queryTraceThreads/queryThreads) for grouping traces by thread.
  • Cross-agent communication tools and peer thread signal mechanisms (experimental, off by default).
  • Trace filters for span names, model providers, timing, outcomes, and version lineage.
  • Per-endpoint discovery features for observability storage (entity types, services, environments).

Improved

  • Sandbox configuration now uses callbacks (ctx => new E2BSandbox) for dynamic setup.
  • Workflow control-flow entries (.parallel, .branch, .foreach, etc.) now support optional id, description, and metadata.
  • Trace deletion now cascades to spans, scores, feedback, and logs; dataset.deleteExperiment() deletes associated traces.
  • Agent Controller stream messages changed to messagestart plus ID-addressed messageupdate deltas for efficient delivery.
  • Background execution now supports per-call dispositions, awaited calls, and bounded delegated tool execution.
  • Observability feedback now carries reviewStatus (needs-review/reviewed) with filtering and update endpoints.
  • Model timeout control via modelSettings.timeout.firstChunkMs to bound first chunk latency.
  • Consistent, resumable prune() execution for storage adapters with bounded work and cancellation.

Fixed

  • Server-defined toModelOutput now supported for client-side tools without execute functions.
  • Guardrail processors now have configurable error handling instead of always warning and failing open.
  • Observer detach() now works correctly; early exit from fullStream also stops observation.
  • Background task execution fenced with persisted, expiring ownership leases to prevent duplicate execution.
  • Workspace grep now includes .sas, .log, and .jsonl files by default with extensible text-extension whitelist.
  • Model failure in Observer/Reflector no longer ends the run when failurePolicy is 'continue'.

August 2026

8 releases: 1.56.0 → 1.63.0

August brought declarative workflow persistence, improved experiment controls, and foundational changes to how agents handle signals and traces. The month also refined storage efficiency and fixed several memory and session management issues.

Action needed
  • MASTRAPLATFORMSECRETKEY auth removed for workspace providers; use MASTRAPLATFORMACCESSTOKEN or pass accessToken explicitly.
  • Stored workflows renamed to dynamic workflows: addStoredWorkflow → addDynamicWorkflow, StoredWorkflow types → DynamicWorkflow.
  • CostGuardProcessor renamed to TokenCostControl; deprecated aliases will be removed in future major version.
  • Channel agent replies now post as markdown by default; set textFormat: 'plain' to preserve plain-text posting.

Highlights

  • Stored workflows renamed to dynamic workflows with new persistence and loading semantics.
  • Agents now support transient signals that remind the model without retaining history, and can filter tool results through a new processToolResult lifecycle hook.
  • TokenCostControl processor replaces CostGuardProcessor with better reliability, diagnostics, and budgeting options.
  • Caller-driven experiments let external orchestrators own the experiment loop while Mastra tracks results.
  • Channel agents now post replies as markdown by default, with plain-text posting available via configuration.

New

  • Item-level scorer selection for dataset experiments.
  • Batch trace ID filtering in observability metric queries.
  • Per-run controls to keep experiment and score results in memory without writing to storage.
  • Tool definitions now captured in MODELGENERATION span attributes.
  • Transient agent signals that deliver reminders without retaining conversation history.
  • Typed experiment provenance and grouping fields that persist across runs and support filtering.
  • instructions.ts as a code alternative to instructions.md for file-based agents.
  • ProcessHandle.closeStdin() to signal end-of-file to background processes.

Improved

  • Declarative, persistable workflow graphs with strict save-path validation and lenient load-path handling.
  • RegexFilterProcessor now reports what the redact strategy rewrote via the onViolation callback.
  • Browser UI can now import toAISdkMessages without requiring Node.js APIs.
  • AgentController supports live-session deletion with process-local listeners.
  • Cost checks now issue fewer queries against observability storage.
  • Sessions resume existing conversations instead of opening empty threads on startup.
  • MCP tools served over HTTP see the authenticated caller when running behind a Mastra server.
  • Graceful shutdown options for generated servers with configurable drain time.

Fixed

  • Memory leak where suspended agent runs were never released from memory.
  • Sessions opening new empty threads on every start instead of resuming conversation.
  • Traces starting under external parent spans not appearing in Mastra Studio.
  • Resumed agent and workflow runs now properly maintain link to suspended run's trace.
  • Fixed worker processes to discover schedules created after startup without restart.
  • Output stream processor traces remaining open when streams close or cancel before finish.

July 2026

9 releases: 1.47.0 → 1.55.0

July brought major architectural updates to Mastra, including a complete migration from the legacy Harness system to AgentController, expanded AI SDK support, and new file-system-routed configuration patterns. The month also saw significant improvements to sessions, workflows, storage, and observability.

Action needed
  • Harness system removed: migrate /harness/... routes to /agent-controller/... and harness: permissions to agent-controller: permissions.
  • @mastra/client-js removed deprecated getHarness, listHarnesses, Harness, and HarnessSession classes.
  • Harness config no longer accepts resolveModel, customModelCatalogProvider, modelAuthChecker, or WorkspaceConfig shorthand for workspace.
  • Session identity now mandatory: id and ownerId required on harness.createSession() and Session constructors.

Highlights

  • Harness system fully replaced by AgentController with new API routes and permission model.
  • Added support for AI SDK v7 models alongside existing v4, v5, and v6 support.
  • File-based agents now support nested subagents up to three levels deep.
  • Chat channel support added to AgentController for Slack, Discord, Telegram and other messaging platforms.
  • New file-system-routed configuration for observability, storage, and workflows with auto-discovery.

New

  • Support for AI SDK v7 models (LanguageModelV4) passed directly to agents.
  • Storage retention policies with configurable maxAge and storage.prune() deletion.
  • Nested subagents in file-based agents up to three levels deep.
  • Chat channel support for AgentController with messaging platform adapters.
  • Built-in web fetch and web search tools for agents.
  • Workspace-level provider registry for MastraEditor.
  • Optional provider-driven authorization for system actors with requireActor() method.
  • Optional networking capability and writeFiles() method on WorkspaceSandbox interface.

Improved

  • DurableAgent surface now mirrors in-process Agent's stream, resume, and generate APIs.
  • Session identity now required with mandatory id and ownerId; workspace moved to per-session.
  • Internal judge agent's agentic-loop iteration limit now configurable via optional maxSteps.
  • Tenancy scoping added to DatasetsManager.get and DatasetsManager.delete across all storage operations.
  • Workflow run event topics automatically cleaned up after reaching terminal state.
  • Build deployment now happens in one step for push-style deployers with deployOnBuild flag.
  • Custom workers option now merges with defaults instead of replacing them.
  • Stream error retry policies now support custom processors and maxRetryAfterMs limits.

Fixed

  • Cross-tenant data-access issue in datasets by applying tenancy filtering.
  • Sub-agent tool approvals now resume the correct parent agent run after restart.
  • Thread title generation no longer receives duplicate conversation messages.
  • Channel tool approval actions prevented from creating conversations under button clicker's identity.
  • Zod 4 installations no longer report peer dependency warnings.
  • Missing thread mappings in channel reactions cannot create unintended conversations.

June 2026

7 releases: 1.38.0 → 1.46.0

June brought major API simplifications to the testing harness and streaming infrastructure, along with improvements to tool approvals, notifications, and state management. The month focused on making the framework more flexible for production use cases like background jobs and long-running threads.

Action needed
  • Harness no longer exposes singleton harness.session; callers must use await harness.createSession().
  • Run control moved from Harness to Session (sendMessage, sendSignal, abort, respondToToolSuspension now on session).
  • Event subscription moved from Harness.subscribe() to session.subscribe().
  • Thread lifecycle APIs (createThread, switchThread, cloneThread, renameThread) moved to session.thread.

Highlights

  • Harness API refactored to session-based model, moving all run control and thread lifecycle methods away from the singleton.
  • Tool approval system now supports conditional, function-based approvals evaluated per tool call with full context.
  • Streaming can now stay open across background-task continuations with the new untilIdle option.
  • Trusted actor support added for background jobs and scheduled tasks to run without JWT authentication.
  • Experimental notification signals with priority-aware delivery and thread-scoped inbox storage.

New

  • agent.sendMessage() and agent.queueMessage() APIs for injecting user input into agent threads.
  • channels.threadContext.addSystemMessage to opt out of built-in channel system messages.
  • Workspace sandbox now accepts a resolver function for per-request custom sandboxes.
  • MessageFactory component in @mastra/react for type-safe rendering of MastraDBMessage with custom per-part components.
  • Processor state signals for memory-backed thread context.
  • Gateway embedding routing support with mastra/provider/model IDs.
  • Actor option for agent.generate() and stream() invocations to identify trusted system processes.
  • UntilIdle field on /agents/:agentId/stream and /agents/:agentId/resume-stream endpoints to keep streams open across continuations.

Improved

  • MCP tool approvals now honored end-to-end with per-tool requireApproval functions preserved during conversion.
  • Global requireToolApproval option now accepts functions in addition to booleans, evaluated per tool call.
  • Tripwire reason now persisted as tripwire.reason field in addition to message text.
  • CostGuardProcessor now falls back to MastraMemory context when running without auth middleware.
  • Removed Hono from core and auth runtime dependencies; auth providers now use framework-agnostic request types.
  • State signal restoration improved so long threads can resume without scanning every message.
  • Agent signal drains now record pending signals through canonical signal transcript path.
  • ACP tools now keep their default session alive across executions.

Fixed

  • Fixed dropdown, combobox, select, and popover popups silently failing when opened outside side panels.
  • Fixed combobox popups so model picker dropdowns open correctly outside side dialogs.
  • Fixed CostGuardProcessor thread and resource scope resolution in Studio dev mode.
  • Fixed agent signal drains to prevent follow-up signal turns from losing context.

May 2026

8 releases: 1.30.0 → 1.36.0

May brought major architectural improvements to Mastra agents, including durable and resumable stream support, fine-grained authorization controls, and significant enhancements to channel adapters and observability. The month also featured numerous breaking changes to APIs and storage schemas that require migration.

Action needed
  • @mastra/inngest now requires inngest@^4 and Inngest Dev Server v1.18.0+; remove @inngest/realtime and import from inngest/realtime.
  • @mastra/client-js Vector return types changed: getIndexes() returns string[], upsert() returns {ids: string[]}, query() returns QueryResult[].
  • @mastra/client-js paginated list methods now require orderBy: {field, direction}; AgentBuilder.stream() now requires runId.
  • Schedules storage schema changed: mastraschedules and mastrascheduletriggers tables add ownerType/ownerId and rename ScheduleTrigger.status to outcome; scheduled workflows require migration.

Highlights

  • Durable agents with resumable streams now allow agent execution to persist across interruptions, with EventedAgent and InngestAgent variants for different execution models.
  • Fine-grained authorization (FGA) enables relationship-based, resource-level access control across agents and resources.
  • Platform channels framework now supports multiple connection flows (OAuth, deep link, immediate) with richer observability and batch-capable embedders.
  • Slack channel adapter gains streaming text, adaptive typing indicators, and flexible tool display modes (cards, text, or timeline).
  • Agent metadata support and favorites storage let you organize, filter, and mark agents and skills as favorites.

New

  • Durable agents extend Agent with resumable streams; EventedAgent and InngestAgent provide fire-and-forget and Inngest-powered execution variants.
  • Azure OpenAI Responses API and v1 routing controls for Azure deployments.
  • ACP-compatible coding agents can now run as Mastra tools or lightweight subagents with incremental response streaming.
  • Platform channels framework with ChannelProvider interface, ChannelsStorage domain, and discriminated union for OAuth, deep link, and immediate connection flows.
  • Fine-grained authorization (FGA) system for relationship-based, resource-level access control with route policies and resolver hooks.
  • Agent metadata and favorites storage with visibility controls (private/public) and favorite counts on stored agents and skills.
  • ProcessLLMRequest processor hook runs after message conversion and before model requests for temporary prompt modifications.
  • Target-aware tool payload transforms for display streams and transcript messages across input, output, errors, approval, and suspension payloads.

Improved

  • loadtool now accepts arrays of tool names for bulk loading with clearer response shape (loaded, notFound, alreadyLoaded arrays).
  • Microsoft Entra ID authentication support for Azure OpenAI gateways eliminates need for API keys when using Azure SDK credentials.
  • Workspace search now supports batch-capable embedders with optional maxBatchSize for single-call chunk embedding.
  • Trajectory scorers now receive Trajectory objects instead of raw agent messages, fixing crashes when accessing run.output.steps.
  • Slack channel adapter now supports streaming text with adaptive typing indicators reflecting agent activities (working, thinking, typing, calling tools).
  • Slack tool display modes offer flexibility: cards (default rich blocks), text (plain text), or timeline (streaming widget with status icons).
  • Agent signal ordering stays correct when conversations reload from memory.
  • Background process output retention and background-task observability improved.

Fixed

  • Fixed channel adapter startup crash in 1.29.0 caused by 'Custom API route must not start with /api' validation.
  • Fixed MCP client support in the agent editor.
  • Fixed workspace PATCH bug where omitted config fields overwrote previously-persisted values with undefined.
  • Fixed crash when importing @mastra/core/workflows/workflow from tests or apps due to undefined class extension.
  • Fixed scheduled workflows created from public @mastra/core/workflows entry point so declared schedules apply correctly.

April 2026

6 releases: 1.24.0 → 1.25.0

April brought major observability enhancements with RAG tracing support, improvements to agent execution and memory handling, and security fixes for credential leakage in telemetry data. Multiple versions focused on reliability and adding flexibility to core features.

Action needed
  • CloudExporter endpoint format changed: configure a base endpoint URL with automatic publisher path derivation.
  • Sarvam voice provider (@mastra/voice-sarvam@1.0.0) drops deprecated models and changes TTS payload from inputs[] to text.

Highlights

  • RAG observability now tracks ingestion pipelines, embeddings, vector operations, and chunk actions alongside agent and workflow traces.
  • Fixed critical credential leakage in observability spans where API keys and authentication headers could be sent to telemetry backends.
  • Browser automation now supports CLI-driven screencast functionality with automatic CDP injection.
  • Forked subagents can now run on a clone of the parent thread, enabling parallel branches within workflows.
  • Observational memory improved to prevent message duplication and correctly handle streamed tool results.

New

  • RAG observability with span types for ingestion, embeddings, vector operations, and chunk actions.
  • CLI-driven browser automation with screencast support and automatic CDP injection.
  • Opt-in temporal-gap markers for observational memory to signal gaps in agent awareness.
  • S3 prefix (subdirectory) mount support for filesystem configuration.
  • Custom language server registration via LSPConfig servers field.
  • Forked subagents that execute on a clone of the parent thread.
  • Scorer utilities for input preparation with prepareRun and filterRun functions.
  • Stream error retry processor with OpenAI Responses stream error matching.

Improved

  • CloudExporter now derives publisher paths automatically from a configured base endpoint URL.
  • Span filtering with excludeSpanTypes and spanFilter options in ObservabilityInstanceConfig.
  • Workspace filesystem configuration now accepts a resolver function in addition to static instances.
  • Per-entry model settings, provider options, and headers in agent model fallback arrays.
  • Sub-agent version overrides with cascading propagation from Mastra instance defaults.
  • Progress throttling for background tasks via progressThrottleMs to coalesce high-frequency output.
  • Client-disconnect aborts no longer logged as errors in LLM execution.
  • Provider registry and model documentation updated with latest models and providers.

Fixed

  • Foreach parallel iterations no longer lose suspendPayload when sibling iterations are resumed.
  • Message duplication in observational memory caused by savePerStep interaction.
  • Output processors returning undefined causing undefined chunks in consumer streams.
  • Streamed tool results being replayed during mid-stream observational memory execution.
  • False positive provider change detection in observational memory metadata.
  • Mastra workspace list files silently returning no results with empty patterns.

March 2026

8 releases: 1.8.0 → 1.16.0

March brought significant improvements to workflow execution tracking, multi-agent coordination, and observability in Mastra. The month focused on enhancing debugging capabilities, adding supervisor patterns for agent coordination, and refining how tools and models are handled across the platform.

Action needed
  • LocalFilesystem now treats absolute paths as real filesystem locations; pass relative paths to target the workspace.
  • ProcessHandle.pid changed from number to string.
  • MCP tool calls now use MCPTOOLCALL span type instead of TOOLCALL in traces.
  • MetricType (counter/gauge/histogram) deprecated in favor of raw events with aggregation at query time.

Highlights

  • Added supervisor pattern for multi-agent coordination with delegation hooks and memory isolation.
  • Workflow execution path tracking lets you see which steps ran and access the path during or after execution.
  • Tool input examples now guide AI models on valid inputs for better tool usage.
  • Dynamic model selection with fallback arrays enables context-based routing and flexible model selection.
  • Experiments now support dataset targeting, status tracking, and agent version pinning for comprehensive evaluation.

New

  • Supervisor pattern for multi-agent coordination using stream() and generate() with delegation hooks and iteration monitoring.
  • Workflow execution path tracking in results and execution context to see which steps ran.
  • Input examples support on tool definitions to show models what valid inputs look like.
  • Dynamic functions returning fallback model arrays for context-based model selection.
  • OnStepFinish and onError callbacks to NetworkOptions for per-LLM-step progress monitoring.
  • Dataset-agent association and experiment status tracking in the Evaluate workflow.
  • LLM-powered dataset item generation and failure analysis endpoints for experiments.
  • Agent version support for experiments to pin which agent version is used.

Improved

  • QueryVectorParams now makes queryVector optional to support metadata-only queries.
  • Harness.sendMessage() now uses files instead of images, supporting any file type and preserving filenames.
  • Execution logs are now smaller and more readable with automatic duplicate payload removal.
  • Dynamic model selection now supports nested dynamic functions within returned arrays.
  • Tool timeout parameter now accepts seconds instead of milliseconds.
  • Provider tools are now properly handled when using custom gateways that return AI SDK v6 models.
  • Observability metrics refactored to use raw events with aggregation at query time.
  • MCP tool calls now use MCPTOOLCALL span type in traces with server name and version information.

Fixed

  • ModelRouterEmbeddingModel.doEmbed() crash with AI SDK v6's embedMany when result.warnings was not iterable.
  • Provider-executed tool calls being dropped or saved out of order during memory replay.
  • Provider-executed tool calls from deferred tools being lost or incorrectly persisted.
  • ReplaceString utility incorrectly interpreting $ characters as regex backreferences.
  • AllowedPaths resolving against working directory instead of basePath, causing unexpected permission errors.
  • Tool invocation updates not preserving providerExecuted and providerMetadata from original tool calls.

February 2026

5 releases: 1.3.0 → 1.7.0

February brought major infrastructure improvements to Mastra, including mount support for multi-storage workspaces, new Datasets and Experiments features for testing AI outputs, and a powerful generic Harness class for orchestrating agent workflows. The month also focused on better file handling, streaming tool previews, and improved type safety across the platform.

Action needed
  • @mastra/memory: observe() now takes a single object parameter instead of positional arguments
  • Workspace.getPathContext() is deprecated in favour of getInstructions()

Highlights

  • Mount support lets agents access files from multiple storage providers (S3, GCS, local disk) under a single directory tree.
  • New Datasets and Experiments features enable versioning test inputs and running AI outputs against them with results tracking.
  • Generic Harness class provides unified orchestration for agents with built-in tools, state management, and memory integration.
  • Streaming tool argument previews show code diffs, file content, and command arguments in real-time as the model generates them.
  • AST edit tool performs intelligent code transformations including renaming, import management, and pattern-based find-and-replace.

New

  • CompositeFilesystem for combining multiple filesystems under one directory tree.
  • Datasets and Experiments core features with versioning, schema validation, and experiment execution tracking.
  • Glob pattern support in workspace configuration and listfiles tool for flexible file filtering.
  • Direct skill path discovery allowing workspace skills to reference specific directories or SKILL.md files.
  • Vercel AI Gateway support in the model router for routing through official AI SDK gateway providers.
  • AST edit tool for code transformations including identifier renaming and import management.
  • Thread lock callbacks in HarnessConfig for preventing concurrent thread access across processes.
  • HarnessDisplayState for reading unified agent state snapshot instead of handling individual events.

Improved

  • Native Groq support in model router using official AI SDK package instead of OpenAI-compatible fallback.
  • MastraFilesystem and MastraSandbox base classes with safer concurrent lifecycle handling.
  • Flattened scorer type system with top-level type, instructions, scoreRange, and presetConfig fields.
  • LocalFilesystem.resolvePath fixed to properly handle absolute paths and filesystem info.
  • Typed workspace providers now return concrete types improving autocomplete and eliminating type casts.
  • Workspace instruction improvements with accurate context distinguishing sandbox-accessible paths from workspace-only paths.
  • Tool definitions can now include toModelOutput function to transform results before sending to model.
  • Added background process management to workspace sandboxes.

Fixed

  • Fixed absolute path resolution in LocalFilesystem that was incorrectly stripping leading slashes and causing PermissionError.
  • Fixed memory observe() method signature inconsistency in observational memory module.
  • Fixed workspace context injection to distinguish between sandbox-accessible and workspace-only paths in agent instructions.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Mastra: pricing, features and alternatives · All changelogs