v3.5.0 — Streamable HTTP transport

@felores/kie-ai-mcp-server@3.5.0

Adds a remote Streamable HTTP transport (MCP spec 2025-11-25) alongside the default stdio path. Reimplements the capability from closed PR #3 against the current monorepo.

Added

• Streamable HTTP transport — opt in with MCP_TRANSPORT=http or --http; stdio stays the default. Single /mcp endpoint (POST + GET/SSE + DELETE), stateful sessions keyed by Mcp-Session-Id (each session gets its own MCP Server over a shared Kie.ai client + task DB).
• Health endpoint GET /health → {status, transport, sessions, version}, unauthenticated and exempt from Origin/DNS-rebind checks for container probes.
• Bearer-token auth (KIE_MCP_HTTP_TOKEN) and DNS-rebinding protection (MCP_ALLOWED_HOSTS, required when binding beyond loopback; the server fails fast otherwise). New env: MCP_HTTP_HOST (default 127.0.0.1), MCP_HTTP_PORT (default 3000).
• Deployment — packages/mcp/Dockerfile (multi-stage, bundled, non-root mcp user, HEALTHCHECK), root .dockerignore, docker-compose.coolify.yml, and docs/DEPLOY_HTTP.md.
• README documents the feature and its benefits in English and Spanish.

Notes

• stdio remains the default; the @felores/kie-cli package is unaffected.
• Out of scope for now: OAuth 2.0 (RFC 9728) and eventStore stream resumability — the transport supports adding both later.

See CHANGELOG.md and docs/DEPLOY_HTTP.md.

npm: npx -y @felores/kie-ai-mcp-server

v3.4.0 — single-call wait_for_task + MCP progress

Highlights

wait_for_task now waits for a generation to finish in a single call, so the agent no longer loops over get_task_status. By default it polls the Kie API directly (no callback infrastructure); the callback rendezvous is kept as an optional fallback.

While waiting, the MCP server streams notifications/progress on the still-open tools/call request (when the client sends a progressToken), which resets the client's request timeout so the call stays open past the default 60s. For long jobs (video), enable resetTimeoutOnProgress with a generous maxTotalTimeout in your client.

Terminal results reuse get_task_status's per-API parsing, so completed Suno/multi-output tasks return every URL.

Packages

• @felores/kie-ai-mcp-server@3.4.0
• @felores/kie-cli@0.2.0 (single-call wait too, without the MCP-only progress stream)

See CHANGELOG.md for details.

v3.3.2 — README + Spanish + em-dash cleanup

Docs and polish release (no functional changes).

• Rewrote the README around the token-efficiency story (load only the tools you need; the CLI costs zero context tokens until called), fact-checked and trimmed.
• Added a Spanish README (README.es.md) with a 🇬🇧/🇪🇸 language switcher.
• Removed em dashes from docs and a few tool description strings (clean kie-cli --help / MCP listTools).

Packages: @felores/kie-ai-mcp-server@3.3.2, @felores/kie-cli@0.1.2.

v3.3.1 — code===200 task-creation guard

Fixed

Six generation tools (kling_video, veo3_generate_video, nano_banana_image, wan_animate, hailuo_video, flux2_image) created a task and returned success: true whenever a taskId was present, even on an application-level failure (HTTP 200 with code !== 200). They now gate task creation and the success return on response.code === 200 && taskId, surfacing the API error otherwise — consistent with the other 19 tools.

Addresses the CodeRabbit review on #2 (thanks @evertbal).

Packages: @felores/kie-ai-mcp-server@3.3.1, @felores/kie-cli@0.1.1.

v3.3.0 — Monorepo + standalone kie-cli + MCP SDK 1.x

What's new

Two independently installable packages from one shared tool registry:

• @felores/kie-ai-mcp-server@3.3.0 — MCP server (for Claude Desktop and other MCP clients).
• @felores/kie-cli@0.1.0 — new standalone CLI (npm i -g @felores/kie-cli, then kie-cli --help). Generate images/video/music/speech from the terminal, no MCP client needed.

Highlights

• Monorepo with a private, bundled core feeding both surfaces; adding a model is one file (npm run add-tool).
• MCP listTools/dispatch and CLI commands both derived from a single registry; inputSchema derived from Zod (no more duplication).
• Upgraded @modelcontextprotocol/sdk 0.4.0 → 1.29.0.
• Resources & prompts are now generated from the registry (fixes pre-existing ENOENT; 33 resources + 2 prompts work).
• Recovered 222 parameter descriptions into the schemas (MCP inputSchema + CLI help + docs).
• Includes a Claude Code agent skill in skills/kie-ai/.

See CHANGELOG.md for details.

Release v3.2.1

What Changed

Removed

• Sora 2 (sora_video): Removed following OpenAI deprecation. All Sora 2 endpoints are no longer available.

Full Changelog: v3.2.0...v3.2.1

v3.2.0

What's New

HappyHorse 1.0 Video (happyhorse_video)

New multi-mode video generation tool by Alibaba ATH:

• Text-to-Video: prompt-only generation
• Image-to-Video: 1 input image with prompt
• Reference-to-Video: up to 9 reference images
• Video Edit: edit existing video with reference images and audio
• Supports 720p/1080p, 5 aspect ratios, 3-15s duration

GPT Image 2 (gpt_image_2) - Replaces openai_4o_image

• Text-to-image and image-to-image (up to 16 reference images)
• Resolution options: 1K, 2K, 4K
• Aspect ratios: auto, 1:1, 9:16, 16:9, 4:3, 3:4
• Uses Market API (/jobs/createTask)

Wan 2.7 (wan_video) - Upgraded from 2.5

• 4 modes: text-to-video, image-to-video, reference-to-video, video-edit
• New parameters: first_frame_url, last_frame_url, reference_video, reference_voice, audio_setting
• Duration range expanded to 2-15 seconds

Full Changelog: v3.1.0...v3.2.0

Release v3.1.0

Seedance 2.0 Upgrade

Breaking Changes

• Removed: quality, image_url, end_image_url, camera_fixed, seed, enable_safety_checker
• Renamed: image_url → first_frame_url, end_image_url → last_frame_url
• duration changed from string (2-12s) to integer (4-15s)
• resolution now 480p/720p only (removed 1080p)
• Model IDs changed to bytedance/seedance-2 / bytedance/seedance-2-fast

New Features

• mode: standard (quality) or fast (iterative workflows)
• reference_image_urls: up to 9 reference images for style/subject guidance
• reference_video_urls: up to 3 reference videos for motion guidance
• reference_audio_urls: up to 3 audio clips for sound-guided generation
• generate_audio: native audio generation (default: true)
• web_search: prompt enhancement via web search
• nsfw_checker: content filtering
• adaptive aspect ratio: matches input frame dimensions
• Prompt length expanded to 3-20000 characters

Release v3.0.1

Added

• Topaz Image Upscale: New topaz_upscale_image tool for AI-powered image enhancement and upscaling
  • Supports 1x, 2x, 4x, and 8x upscale factors
  • High-fidelity detail restoration and natural texture reconstruction
  • Max output dimension: 20,000 pixels per side
  • Pricing: 10 credits (≤2K), 20 credits (4K), 40 credits (8K)
  • Model ID: topaz/image-upscale

See CHANGELOG.md for details.

Release v3.0.0

Breaking Changes

• Nano Banana 2: Upgraded from Gemini 3.0 Pro to Gemini 3.1 Flash. Removed upscale mode. Renamed image_urls → image_input (max 14). Added google_search param. Model: nano-banana-2
• Seedream 5.0 Lite: Default version changed from V4 to V5 Lite. Version enum: ["4", "5-lite"]. Model: seedream/5-lite-*
• Kling 3.0: Complete upgrade. Flexible 3-15s duration, multi-shot storytelling, character elements, std/pro modes. Single model: kling-3.0/video

See CHANGELOG.md for full details.