genkit

Features

Google GenAI

• Added the Gemini Deep Research agent (deep-research-preview-04-2026) to the vertexAI plugin as a background model. (#6872)

  const { operation } = await ai.generate({
    model: vertexAI.model('deep-research-preview-04-2026'),
    prompt: 'Compare TCP and UDP in a short markdown table.',
  });
  

• googleAI now sends store on generateContent requests when it's set, so you can opt out of the Gemini API's new default request logging for paid projects. (#6919)

• Interactions API models now accept Word, Excel and other document types instead of only PDF and CSV. (#6904)

• gemini-nano-banana-* models on googleAI are now treated as image models and accept imageConfig. (#6910)

• generateContent models now report tool-use tokens as usage.custom.toolUseTokens. (#6963)

Bug Fixes

Google GenAI

• On Interactions API models, a request-level store now overrides the plugin-level store option. (#6963)
• Server-side tool steps (URL context, file search, Google Maps, MCP) are now kept in multi-turn history on Interactions API models. (#6909)
• vertexAI stateless multi-turn conversations that used Google Search no longer fail with a 501. (#6909)
• Output settings are now sent as response_format, which also fixes JSON output losing its schema when combined with audioFormat or videoFormat. (#6910)
• responseMimeType: 'application/json' and JSON output without a schema now work on Interactions API models. (#6963)
• Removed the shut-down Veo models from googleAI. veo-* names now fail with a clear error pointing to gemini-omni-1.1-flash. Veo is still available on vertexAI. (#6918)

Breaking Changes

• Google GenAI, Interactions API usage: per-modality token counts moved from inputCharacters, inputImages, outputAudioFiles, etc. to usage.custom.*Tokens (e.g. inputTextTokens). Those old fields held token counts by mistake. See #6963 for the full mapping.

Full Changelog: https://github.com/genkit-ai/genkit/compare/[email protected]@1.44.0

CLI

This release improves the Genkit MCP server and documentation commands with automatic project language detection and more reliable runtime restarts, adds trace indexing and filtering by session ID, and introduces experimental flags for Reflection API authentication and OTLP telemetry routing.

New Features

• Automatic Project Language Detection for Docs & MCP Tools: genkit docs:list and genkit docs:search, as well as the MCP list_genkit_docs, search_genkit_docs, and get_usage_guide tools, now automatically detect your project's language (js, go, python, or dart) from workspace files instead of defaulting to JavaScript.
• Session ID Trace Indexing & Filtering: Local trace storage now indexes agent session IDs, and genkit trace:list adds a --session-id <sessionId> option to filter traces by session.
• Experimental Reflection API Authentication & Host Binding: Added --experimental-auth (and GENKIT_REFLECTION_SECRET_TOKEN support) across CLI workflow and runtime commands (start, start:flutter, mcp, flow:run, flow:batch-run, eval:run, eval:flow, and eval:extract-data) to require shared-secret authentication from connecting runtimes. The Reflection v2 WebSocket server now binds to loopback (127.0.0.1) by default and adds --reflection-v2-host <host> on start and start:flutter for remote or containerized runtimes.
• Experimental OTLP Telemetry Routing: Added genkit start --experimental-use-otel to point an application's own OpenTelemetry SDK at the local Dev UI telemetry server via standard OTEL_EXPORTER_OTLP_* environment variables while suppressing native direct telemetry export.

Fixes

• Reliable MCP Runtime Restarts & Tool Metadata: Fixed restart_runtime in the MCP server so process start options (such as working directory and non-interactive mode) are preserved across restarts and the tool waits for the restarted runtime to finish registering before returning. Added server-level MCP instructions, clearer tool descriptions, and improved error formatting.
• Compact list_genkit_docs MCP Output: Omitted section headers from list_genkit_docs tool output to reduce token overhead when AI assistants browse documentation.
• Smaller CLI & Telemetry Package Tarballs: Build dependency improvements and clean up.

Dev UI

This release introduces live interactive rendering, validation, pop-out inspection, and end-to-end action simulation for A2UI surfaces across the Agent, Model, and Prompt runners, alongside layout and preference improvements in the Trace Viewer and navigation sidebar.

New Features

• Live A2UI Surface Previews & Validation: A2UI Surface cards now render interactive live previews for Basic Catalog surfaces directly inline. Surfaces are validated against catalog schemas, surfacing error or warning callouts and per-node issue badges in the component tree when validation issues or custom catalogs are detected.

• Pop-Out A2UI Surface Inspector: Added an Inspect button on previewable A2UI surface cards that opens a resizable split-pane workbench pairing the live preview with the Component Tree and Active Data Model. Hovering or focusing nodes in the component tree highlights the corresponding rendered element in the live preview, and edits made in interactive preview controls sync live with the inspected data model and bound tree properties.

• Interactive A2UI Action Simulation & Action Cards: You can now trigger client-to-server A2UI actions directly by in

Highlights

Agents that recover from failures, plus background sub-agents

Agent runs can now pick up where they stopped after a failure or an abort, and orchestrators can hand work to sub-agents that run in the background.

• throwOnError: false on generate() returns the failure and the conversation up to the last completed tool round, so you can retry from there. Code that doesn't set it still gets the same errors as before. (#6359)

  const opts = { model: googleAI.model('gemini-flash-latest'), tools: [getWeather], throwOnError: false };
  let res = await ai.generate({ ...opts, prompt: 'What should I wear in Tokyo today?' });
  if (res.finishReason === 'failed' || res.finishReason === 'aborted') {
    res = await ai.generate({ ...opts, messages: res.messages });
  }
  

• Failed and aborted agent turns are now resume points. A failed turn is saved as a failed snapshot and a stopped run is saved as an aborted snapshot. Both can be resumed. (#6361, #6362, #6360)

• waitForSnapshot is the blocking version of getSnapshot. It waits on the server until a detached turn finishes, so you don't need a polling loop. (#6272)

  const task = await reportAgent.chat().detach('Write the report');
  const snap = await task.wait();
  

• Background delegation in the agents middleware. With async: true, delegation tools accept background: true and return a task ID right away. New check_background_tasks, wait_for_background_tasks and abort_background_tasks tools let the orchestrator manage those tasks. (#6273)

  use: [agents({ agents: ['researcher'], async: true })]
  

• continue_task continues a failed, aborted or finished sub-agent task from its handle, with optional new instructions. Snapshot reads also accept metadataOnly to skip loading the state. (#6693)

• Transport-specific options in AgentAPI. In-process agents accept a per-chat or per-call context, so auth carries through detached tasks and background jobs. The building blocks for custom AgentTransport implementations are now exported. (#6448)

  const chat = weatherAgent.chat({}, { context: { auth: { uid: 'alice' } } });
  

New contextCompression middleware

@genkit-ai/middleware now includes contextCompression, which keeps long agent conversations within a token budget. It truncates tool responses, removes duplicate tool responses, caps message counts, and can summarize old turns with an LLM. It tracks compaction history so later rounds build on earlier ones. (#6266, #6267, #6269, #6270)

use: [
  contextCompression({
    maxInputTokens: 80000,
    deduplicateToolResponses: { matchBy: 'name-and-input' },
    toolResponses: { maxChars: 2000, preserveRecent: 2 },
    summarize: { model: googleAI.model('gemini-flash-lite-latest'), preserveRecent: 6 },
  }),
]

Gemini: Interactions API, Gemini 4 Argon, Omni and more

• The googleAI plugin now sends requests for -latest aliases, newer Gemini models and any model it doesn't know to the Interactions API. Older built-in models stay on generateContent. (#6344)
• The vertexAI plugin routes the same way, with one difference: image models and the -latest aliases stay on generateContent. (#6851)
• New models: gemini-4-argon, gemini-3.8-flash, gemini-3.8-flash-tts, gemini-3.8-flash-lite-tts, gemini-nano-banana-2.1, gemini-omni-1.1-flash (replaces Veo 3.1), lyria-3.5 and antigravity-preview-09-2026. Gemini 2.5 and the retiring preview models are now hidden: they still work but no longer show up in autocomplete. (#6807)
• vertexAI gets the same model updates, plus text-to-speech, Omni video generation and the gemini-embedding-2 embedder. (#6859)

Breaking for models on the Interactions API: response.raw/custom now has a different shape. safetySettings, toolConfig, thinkingBudget and candidates > 1 are dro

Tools and interrupts are stable and typed end to end. A tool that pauses is resumed with the type it declared, a tool error can go back to the model to fix, and an approval can be decided by a judge model or refused with a reason. Every run reports what it spent, a budget can stop or pause it, and context compression keeps long tool loops inside the window. Agents can delegate to agents that other Genkit apps serve, sessions persist in PostgreSQL, and telemetry plugs in with or without OpenTelemetry. A preview plugin adds decision models.

go get github.com/firebase/genkit/[email protected]

Tools that pause are resumed by type

genkit.DefineResumableTool declares a third type: what the tool is resumed with. The tool gets it as a pointer, nil on the first call. The caller claims each paused call with the tool's Interrupted method, which matches the part to the tool and decodes its input, so nothing after the claim can fail.

Before: the caller matches tools by name, reads the input as an untyped map, and the tool reads its answer back one key at a time.

transferMoney := genkit.DefineTool(g, "transferMoney", "...",
	func(tc *ai.ToolContext, in TransferInput) (*TransferOutput, error) {
		if !tc.IsResumed() && in.Amount > approvalLimit {
			return nil, ai.InterruptWith(tc, map[string]any{"reason": "over_limit"})
		}
		if approved, ok := ai.ResumedValue[bool](tc, "approved"); ok && !approved {
			return &TransferOutput{Status: "declined"}, nil
		}
		// ...
	})

var restarts []*ai.Part
for _, part := range resp.Interrupts() {
	switch part.ToolRequest.Name {
	case transferMoney.Name():
		in := part.ToolRequest.Input.(map[string]any) // untyped on the part
		approved := askHuman(in["amount"].(float64), in["toAccount"].(string))
		restart, err := transferMoney.RestartWith(part,
			ai.WithResumedMetadata[TransferInput](map[string]any{"approved": approved}))
		if err != nil {
			return nil, err
		}
		restarts = append(restarts, restart)
	}
}
resp, err = genkit.Generate(ctx, g,
	ai.WithMessages(resp.History()...), ai.WithTools(transferMoney), ai.WithToolRestarts(restarts...))

After: the claim does the match and the type check, and the answer has one type at both ends.

transferMoney := genkit.DefineResumableTool(g, "transferMoney", "Transfers money to another account.",
	func(ctx context.Context, in TransferInput, resume *Confirmation) (*TransferOutput, error) {
		if resume == nil && in.Amount > approvalLimit {
			return nil, tool.Interrupt(ctx, map[string]any{"reason": "over_limit"})
		}
		if resume != nil && !resume.Approved {
			return &TransferOutput{Status: "declined"}, nil
		}
		// ...
	})

var parts []*ai.Part
for _, part := range resp.Interrupts() {
	if call, ok := transferMoney.Interrupted(part); ok {
		approved := askHuman(call.Input.Amount, call.Input.ToAccount) // TransferInput, decoded
		parts = append(parts, call.Restart(Confirmation{Approved: approved}))
	}
}
resp, err = genkit.Generate(ctx, g,
	ai.WithMessages(resp.History()...), ai.WithTools(transferMoney), ai.WithResume(parts...))

call.Respond(out) answers without running the tool again, and call.RestartWithInput(in, res) runs it with revised arguments, which history then records as the request. The tool reads the input it was first called with through tool.OriginalInput[In](ctx). Code that holds only the part, such as a generic handler, builds the same parts untyped with part.ToToolRestart, part.ToToolRestartWithInput, and part.ToToolResponse: the payload is checked against the resumeSchema the tool advertises and decoded into the resume type, so a map with the right fields behaves like the typed value, in process or after a wire hop.

The resume is checked as a whole before any tool runs. A restart that would charge a card no longer runs while a sibling request fails the resume, and WithResume rejects a part that is still paused, such as `r