Qencode MCP

Qencode cho phép các trợ lý AI chuyển mã, phân tích, chỉnh sửa, tối ưu hóa và phân phối video bằng ngôn ngữ tự nhiên, được hỗ trợ bởi nền tảng xử lý video đám mây.

GitHub

Tài liệu

qencode-mcp

Model Context Protocol (MCP) server for the Qencode Transcoding API.

Connect any MCP-compatible AI client — Claude, Cursor, ChatGPT, Grok, Gemini — to your Qencode account and let it submit, monitor, and reason about transcoding jobs on your behalf.

Quick example

Once your client is connected (see Connect a client), ask your agent in plain English:

Transcode https://example.com/input.mp4 as an HLS ladder at 1080/720/540/360. Put it in my R2 bucket videos/demo/.

The agent picks the hls_abr recipe, fills in the per-rendition encoding params, submits via start_encode2_raw, and polls until the job is done.

Prerequisites

A Qencode portal account with at least one project — sign in at the portal for your environment: https://portal.qencode.com (production) or https://portal-qa.qencode.com (QA). You select the project during the OAuth consent step.
An MCP-compatible client (Claude, Cursor, ChatGPT, Grok, Gemini, or any custom client).

There are no API keys to copy into client config — authentication is browser-based OAuth.

How it works

The connector uses standard OAuth 2.1 — no API keys in client config. On first use, your client opens a browser, you sign in to your Qencode portal account, pick a project, and approve the requested scopes. The client stores the token; subsequent calls are silent until the token expires.

Scopes the client should request at authorize time (published via Protected Resource Metadata):

Scope	Purpose
`openid`	OIDC identity
`profile`	Display name
`email`	Account email
`offline_access`	Refresh token
`transcoding:read`	`get_job_status`, `wait_for_job`, docs tools
`transcoding:write`	`transcode_video`, `start_encode2_raw`

The RS enforces transcoding:read and transcoding:write on access tokens at the transport layer.

Your Qencode API keys never leave the portal. The MCP server derives a short-lived session token per request via an internal portal endpoint.

Environments

The same connector is deployed in two environments. Each has its own domains, accounts, projects, and credentials — sign in to the portal that matches the endpoint you connect to.

Role	Production	QA (testing)
MCP endpoint (connect here)	`https://mcp.qencode.com/mcp`	`https://mcp-qa.qencode.com/mcp`
Portal (sign in / projects)	`https://portal.qencode.com`	`https://portal-qa.qencode.com`
Authorization server	`https://auth.qencode.com`	`https://auth-qa.qencode.com`
Qencode API	`https://api.qencode.com`	`https://api-qa.qencode.com`

The instructions below use the production endpoint. To test against QA, swap in the QA URL and sign in at the QA portal.

Connect a client

Endpoint: https://mcp.qencode.com/mcp — same for every client below. Sign in to your Qencode account when the browser opens and approve access.

QA (internal testing): use https://mcp-qa.qencode.com/mcp and sign in at the QA portal instead.

Client	Where to add it	MCP URL / config
Claude (chat)	Message box → + → Connectors → Add connector	`https://mcp.qencode.com/mcp`
Claude Code	Terminal	`claude mcp add --transport http qencode https://mcp.qencode.com/mcp`
ChatGPT	Apps → search Qencode → Connect; or Developer Mode → Build app	Connector URL: `https://mcp.qencode.com/mcp`
Gemini	`~/.gemini/settings.json` → `mcpServers`	`"httpUrl": "https://mcp.qencode.com/mcp"` — then `/mcp auth qencode` in the CLI
Cursor	Settings → Tools & MCP → New MCP Server (or `~/.cursor/mcp.json`)	`"url": "https://mcp.qencode.com/mcp"` — restart Cursor after saving

Cursor (mcp.json):

{
  "mcpServers": {
    "qencode": { "url": "https://mcp.qencode.com/mcp" }
  }
}

Gemini (settings.json):

{
  "mcpServers": {
    "qencode": {
      "httpUrl": "https://mcp.qencode.com/mcp",
      "timeout": 30000,
      "trust": false
    }
  }
}

Tip: sign in to portal.qencode.com in your browser before connecting — OAuth goes smoother.

What the connector exposes

Tools

Transcoding & jobs

Tool	Description
`transcode_video`	Submit a job from a source URL to one or more outputs. Convenience wrapper — auto-injects `encoder_version: 2` (or `1` for VMAF) when omitted.
`start_encode2_raw`	Escape hatch — submit a job with the full `query` JSON exactly as the Qencode API expects.
`get_job_status`	One-shot status snapshot by `task_token`.
`get_job_status_detailed`	Full, authoritative job status, including per-rendition progress and output details.
`wait_for_job`	Poll until terminal state, timeout, or internal poll cap. Do not call in parallel with other tools in the same client batch.
`search_qencode_docs`	Search the built-in knowledge base of recipes and reference docs.
`fetch_qencode_doc`	Fetch the full content of a knowledge-base resource by `qencode://` URI (tool-based counterpart to `resources/read`).

Media Storage

Bucket management and ingest for Qencode Media Storage. These ride the same OAuth grant as the transcoding tools — no extra scope and no re-consent.

Tool	Description
`list_buckets`	List the Media Storage buckets available to the account.
`create_bucket`	Create a new bucket. Called only on an explicit request — not to satisfy a missing `destination`.
`list_objects`	Browse the contents of a bucket.
`get_download_url`	Return a time-limited download URL for an existing object.
`download_url_to_bucket`	Server-side copy of a public URL into a bucket (ingest, no transcoding).

Resources

The server ships a knowledge base of recipes and reference docs, exposed as MCP resources so the agent can fetch only what it needs. Notable URIs:

qencode://docs/best-practices — composition defaults the agent applies automatically
qencode://docs/storage — destination compatibility matrix (Qencode S3, R2, AWS S3, Azure, B2, FTP/SFTP)
qencode://docs/error-codes — error code → cause → fix
qencode://docs/gotchas — non-obvious API quirks
qencode://schema/digest — full attribute reference for start_encode2
qencode://recipe/<slug> — one per feature flow: hls_abr, mp4_ladder, audio_outputs, thumbnails, speech_to_text, subtitles, stitching, drm_widevine_ezdrm, drm_fairplay_ezdrm, drm_playready_ezdrm, drm_aes128, drm_buydrm, drm_expressplay, codec_av1, per_title_encoding, incremental_abr, refresh_abr_playlist, callbacks, reliability, video_metadata

Use search_qencode_docs to discover the right recipe URI for a goal.

Prompts (slash commands)

In clients that surface MCP prompts, 21 one-shot templates are available. Each tells the agent to read the matching qencode://recipe/... resource and submit via start_encode2_raw.

ABR / packaging: encode_hls_abr, encode_dash_abr, encode_mp4_ladder, encode_incremental_rung, encode_refreshing_playlist

Codecs / quality: encode_av1, tune_per_title

Audio / images / text: extract_audio, generate_thumbnails, transcribe, add_subtitles

Probe / stitch: get_video_metadata, stitch_videos

Production hooks: enable_callbacks, enable_reliability

DRM: encode_aes128_hls, encode_widevine_ezdrm, encode_playready_ezdrm, encode_fairplay_ezdrm, encode_drm_buydrm, encode_drm_expressplay

Source URL rules

transcode_video and start_encode2_raw accept source values with schemes https://, http://, s3://, or tus:. FTP/SFTP and private/metadata URLs are rejected at the tool boundary (SSRF defence). See docs/security/THREAT_MODEL.md for limitations.

Security

Authentication is OAuth 2.1 only — there is no static-API-key mode. Your Qencode API keys never leave the portal; the server derives a fresh, short-lived session token per request via an internal portal endpoint. Source URLs are validated at the tool boundary (SSRF defence — see Source URL rules).

Full threat model and adversarial test coverage: docs/security/THREAT_MODEL.md.

Development

uv venv && uv pip install -e ".[dev]"
NO_NETWORK=1 pytest -q              # offline L1 + L2 + L5 (~540 tests)
pytest -m protocol                # MCP wire conformance only
pytest -m unit                    # per-tool logic (FakeQencode)
pytest -m security                # OWASP MCP Top 10 adversarial suite

Protocol tests run fully offline (mocked Authorization Server, portal, and Qencode API). CI is the Jenkins job mcp_automated_tests (Jenkinsfile.manual): manual checkboxes for any layer, nightly L3 cron, weekly L4 cron.

Supported MCP protocol versions

Clients negotiate a version at initialize. This server targets MCP 2025-11-25 as the primary version. CI also runs conformance tests against 2025-06-18 because JSON-RPC batching behavior differs between earlier revisions. We do not claim support for 2025-03-26 or older wire semantics beyond what the underlying SDK negotiates.

Version	Support	Notes
2025-11-25	Primary	Streamable HTTP, resumable SSE where used
2025-06-18	CI matrix	Regression guard for mid-2025 clients
2025-03-26	Not targeted	Batching semantics differ from 2025-06-18

More docs

Local server / env vars: docs/local-development.md
Test layers (L1–L5): docs/testing.md and tests/README.md
L3 against live QA / PROD: tests/integration/README.md
L4 agent evals: evals/README.md
Pre-release gate: docs/release-checklist.md

Versioning policy

The connector follows SemVer applied to the MCP surface — tools, prompts, resources, OAuth scopes, and supported protocol versions. Qencode HTTP API changes are out of scope (they are the API's own concern, not the connector's).

MAJOR — a breaking surface change: a tool/prompt/resource is removed or renamed, a previously optional argument becomes required, an OAuth scope is added or tightened in a way that forces re-consent, or a supported MCP protocol version is dropped.
MINOR — a backward-compatible addition: a new tool/prompt/resource, a new optional argument, or a newly supported protocol version.
PATCH — no change to the surface shape: tool/prompt description rewordings, knowledge-base/doc updates, and bug fixes.

Surface changes are guarded by snapshot tests under tests/protocol/. When you change the surface, regenerate the snapshots (python scripts/regen_tools_snapshot.py) and bump the version in the same PR: pyproject.toml, src/qencode_mcp/__init__.py, server.json, and a new CHANGELOG.md entry must all agree.