omni_ai 0.6.1

One Flutter API for every AI provider: Gemini, Vercel AI Gateway, OpenRouter and more. Stream text, call tools, build agents, and generate images, speech and video.

0.6.1

• Fixed: calling ModelCatalog.load() when the catalog was already fresh left it marked as loading for good, so isLoading stayed true and every later load and refresh() did nothing, including the refresh that follows setApiKey. A newly added key's models then never appeared until the app restarted.
• Fixed: refresh() during a load now does fetch again when that load ends. 0.6.0 listed this but still returned the same load.
• Raised the minimum speech_to_text to 7.4.0 and flutter_svg to 2.0.17, the first versions this package compiles against, and shortened the package description.

0.6.0

• Addis AI: batch generation. ai.addis.batches sends up to 10,000 text to speech requests at once (off-peak, within 24 hours, half price): createFromRequests, create, retrieve, list and requests (paged streams), cancel, results, audio and waitFor (with onProgress; rides out dropped connections and busy answers, and resumes by id after a restart). ai.addis.files has upload (one request up to 25 MB, signed upload up to 100 MB), retrieve, delete, contentUrl and content. Requests are checked before upload; creating a batch sends an Idempotency-Key, reused across retries, so a retry can't start a second batch; downloads never send the API key to the storage host. Adds a direct dependency on http_parser, already used through http.
• Jev: pictures and JSON state in every tool. jev_decide, jev_batch (per item), jev_noul, jev_choice, jev_score and jev_box take images ({placeholder: file or link}), turned into bytes by the app's JevToolkit.resolveImage; JevImage.detect reads PNG/JPEG from the bytes. Each picture is sent under a unique id (JevPictures): Jev swaps a placeholder wherever it appears, so the call's own names are replaced as whole words only, and a picture state doesn't mention is added under images. Requests with pictures use visionModel (jev-preview). state written as JSON text is sent as JSON (JevToolkit.stateOf). jev_noul / jev_choice / jev_score now pass the app's images too (they dropped them before), and decideEach takes per-state images, so batches can include box questions.
• Fix, Jev: JevQuestion.fromJson accepts options the way models send them: under options, choices or levels as well as criteria; as a map, a list (of names or {name, description}), a JSON string or a comma-separated string. A choice or score with no options now says what shape it expects instead of only "A choice takes 1 to 255 options".
• Jev: batches. JevClient.decideEach asks the same questions about many states (one request each, since the API has no batch endpoint; four at a time, order kept, a failed state holds its error). The jev_batch tool takes up to 100 items (strings or {id, state}) and returns compact answers per item plus totals (choice counts, yes/no counts, mean score), so an agent judges a list in one call instead of one per item.
• Fix, Skills: previewHit / installHit use the catalog's SKILL.md address (Skyll refs.raw) to read only that skill's folder, instead of scanning every skill in the repository when names differ. SKILL.md and bundled files are fetched six at a time, and each GitHub request gives up after 20 s, so a preview no longer appears to hang.
• Fix, MCP: servers that reject the 2026-07-28 discovery probe as an unsupported protocol version, while listing older ones ("supported versions: 2025-11-25, …"), now fall back to initialize instead of failing to connect. A rejection that lists no older version still does not downgrade.
• Security, MCP: a registry server's header templates (Bearer {api_key}) are filled from the values given (McpCatalogHit.inputs, McpInput); a required one still holding a placeholder throws instead of being saved as if it were a key, and install refuses unfilled {placeholders}. Plain http:// endpoints are only accepted on localhost or the local network (McpRegistry.isLocalHost). The agent's mcp_install / mcp_install_catalog no longer save local command (stdio) servers unless allowAgentStdio is set: connecting runs that command.
• Security, Jev: the key is only sent over https:// (plain http for localhost). Retried decisions carry one generated Idempotency-Key, so a retry is billed once. Every jev_* tool passes the run's cancel token.
• Security, skills: remote skills install over https:// only; a SKILL.md or bundled file is at most 512 KB and a skill bundles at most 40 files / 2 MB. Descriptions put in the agent prompt are capped at 300 characters (the full text is in load_skill).
• Skills: SkillRegistry is Listenable (installs and removals notify). search asks Skyll first, so hits carry a description and the SKILL.md rawUrl, and falls back to skills.sh (searchUrl: '' for skills.sh only). previewHit reviews a hit without installing. load_skill, read_skill_file and uninstall_skill declare their required inputs.
• Fixed: a provider whose key was set while the model catalog was loading (e.g. keys restored at app start) never got its models listed, and refresh() during a load returned that same load. A forced refresh mid-load now fetches again when it ends, setApiKey refreshes a catalog that is loading, and ModelCatalog.isLoading says whether one is under way.

0.5.0

• Fixed misleading MCP response-id errors for redirected endpoints: show the final URL and HTTP status. Manual remote setup adds endpoint/authorization fields, keeps advanced JSON, and fixes controller disposal while the dialog closes.

• Added on-demand MCP install/configure, installation checks, removal, tool-schema inspection, resource and prompt tools; managementTools() avoids eager connections. Skills add remote previews and installation checks. Host apps can opt out of automatic installed-skill exposure when enforcing tool permissions.

• Added reusable ExtensionsView with catalog search, skill review/installation/removal, MCP configuration editing, and tool inspection; the example app adds Extensions and a fully editable Jev playground.

• Jev preserves structured JSON criteria, validates score/type/image requirements, supports boolean naming and normalized API base URLs, and cancels during retry backoff.

• MCP negotiates stateless 2026-07-28 or legacy initialization, reads HTTP SSE results without waiting for connection closure, supports immediate mcp_call_tool after installation, preserves catalog configuration, and migrates credentials to secure storage.

• Fixed skill selectors, GitHub folder filtering, multiline YAML descriptions, root/deep reference files, and concurrent registry loading. Added protocol, persistence, subprocess, cancellation, and UI regression coverage.

• Jev (TypeSafe System One): JevClient / JevToolkit / ai.jev / ai.decide(). Full control of the System One API — noul (yes/no), choice, score and box questions, images, long context, retries, idempotency keys, custom base URL/headers, and one-shot helpers (noul, choice, score, box). Agent tools: jev_decide, jev_noul, jev_choice, jev_score, jev_box, jev_models. Same decision contract as the Vercel AI SDK experimental_decide.

• MCP: createMcpClient (HTTP, SSE, stdio on desktop) lists tools/resources/prompts and turns tools into omni_ai Tools, following @ai-sdk/mcp. McpRegistry / ai.mcp installs, lists and uses servers (install, list, connect, tools, toolkits). Search the official MCP Registry (ai.mcp.search, installHit, installFromCatalog) — no API key.

• Skills: SkillRegistry / ai.skills installs from markdown, a SKILL.md URL, or GitHub (owner/repo, owner/repo@name), lists and uses them (load_skill). Search skills.sh (ai.skills.search, installHit) — the same public catalog as npx skills find. Installed skills are available to agents on that OmniAI.

0.4.1

• Fixed: Addis AI speech failed with a FormatException when the audio came as a data: URL. GeneratedFile.fromBase64 (and fromDataUrl) now take data URLs, line breaks, URL-safe base64 and missing padding; decodeBase64 does the same for any string.
• Fixed: restricted ElevenLabs keys (without user_read) failed verify() though they work; ElevenLabsProvider.isMissingPermission tells the two apart.

0.4.0

Found while moving the teleai app onto omni_ai's agents.

• Approvals can edit: onToolApproval may return ToolApprovalDecision.approve(input: …) (run with an edited input, e.g. a message draft the user changed) or ToolApprovalDecision.deny('why'); bools still work.
• ToolContext.current inside a tool's execute: the call (id, name, input) and the cancel token. Agent hooks now get the real call id (it was empty).
• parallelToolCalls: false on streamText, generateText, Agent and ai.agent() runs a step's tools one at a time, in order.
• ToolOutput(modelFiles: …): pictures a tool fetched for the model to look at, sent after the results as a user message.
• Tool results stream as each tool finishes, and a tool starts only after listeners have seen its call (UIs can show it running).
• Toolkit.where(test) keeps only some of a toolkit's tools (its prompt stays).
• AgentMemory is a Listenable with save, update, remove and clear for your own screens; remember and update_memory return the content; memories saved with a created date load.
• ai.verifyKey(id, key: …) checks a typed key without setting it, for settings fields.
• ai.agent() takes temperature, providerOptions and parallelToolCalls.
• Dictation: onTranscribed reports each cloud transcription with the recording's length (usage, cost). Fixed: errors and status from the device recognizer were lost when deviceAvailable() ran first.
• Voices: VoicePicker speaks Addis AI samples in the voice's own language (VoicePicker.samples), tries Addis's preview endpoint, and can stop a generated sample; AiAudio.play(id:). ElevenLabsProvider.speechModels lists the models every account has.
• Composio: a new key or user id gets a new session (the old one was kept); disconnect uses the package's HTTP client.
• E2B: E2BExecution.count and .text (the cell's value, results marked main), E2BSandbox.startedAt, typed E2BClient.sandboxes() / E2BSandboxInfo, listFiles with paths and dates; kill treats a gone sandbox as stopped; a blank template uses the default.
• Cloudflare: listObjects(delimiter: false) lists everything under a prefix; objects carry their ETag.
• Browser: BrowserSession.host and stop().
• AiHttp.delete.

0.3.0

• On-device sandbox: DeviceSandboxToolkit / ai.deviceSandbox runs Python (Pyodide) or JavaScript in a hidden web view, with input files and output files/charts. Free, no key.
• Browser: BrowserToolkit / ai.browser, a real browser the AI drives (open, read, type, click, back, run JS, close), with BrowserView / showBrowser to watch it live or take over. AiChatView shows a "Watch the browser" button.
• Both are in ai.allToolkits on Android, iOS and macOS. New dependency: webview_flutter.
• Toolkits are only offered where isConfigured (keyless toolkits included).

0.2.0

• Agents, modelled on eve: Agent / ai.agent() with instructions, tools, toolkits, skills (SKILL.md), memory, helper agents, limits, hooks, ask_user, and sessions you can save and restore. agent.chat() plugs into AiChatView.
• Toolkits: ExaToolkit (web search, page reading, similar pages, cited answers), E2BToolkit (cloud code sandbox with charts), ComposioToolkit (users' apps: connect, find and run actions), CloudflareToolkit (browser rendering, DNS, cache, Workers, KV, R2), BasicToolkit (time, calculator, fetch). Keys are saved like provider keys and shown in ApiKeySettings.
• Toolkits on any call: toolkits: on generateText, streamText, ai.ask, ai.chat and AiChatController; ai.allToolkits gives the basic tools, memory and every toolkit with a key.
• Memory: AgentMemory / ai.memory with teleai's tools (remember with tags, update_memory, forget_memory, list_memories), memories in the prompt as id: content; and MemoryFilesToolkit / ai.memoryFiles, a memory tool with the commands of Claude's native memory tool over /memories files. Both stored on the device by default, with pluggable stores.
• CloudflareWorkersAiProvider (workersai:) for Cloudflare's models.
• Tools: ApprovalPolicy (never, always, once, when), ToolOutput to hand files to the app, label. Calls to unknown tools now go back to the model as errors instead of stopping.
• Widgets: showToolApprovalDialog, showAskUserDialog; chat shows tools waiting for approval and helper progress.

0.1.0

First release.

• OmniAI: one object with keys and providers; model ids like google:gemini-2.5-flash.
• Providers: Vercel AI Gateway, Google Gemini (native), OpenRouter, LM Studio, ElevenLabs, Addis AI, any OpenAI-compatible API, and on-device speech.
• Text: generateText, streamText (text, reasoning, tool and file streams), multi-step tool loops with approval, generateObject, streamObject, classify.
• Media: generateImage (with references), generateSpeech, speak, transcribe, live and cloud dictation, generateVideo, embed, embedMany, ElevenLabs sound effects and music.
• Catalog: every provider's models merged with models.dev (prices, context, modalities, logos), cached on device.
• Widgets: AiChatView, AiChatController, ModelPicker, VoicePicker with previews, ApiKeySettings, DictationButton, AiAudio player.
• UiMessageStreamClient for Vercel AI SDK backends.