
Skavio PDF, OCR & Document Processing MCP
io.github.skavio-eu由 GitHub 账号 skavio-eu 发布v0.2.9更新于 Nov 7, 2026
OAuth PDF, OCR, invoice and audio processing with prepaid company credits; card or USDC funding.
概览
远程 MCP 服务器,提供 OCR、PDF 与图片处理、发票提取、媒体转换和转写,按公司预付额度计费。
- 可以做什么
- 将发票和文档字段提取为 XLSX、CSV 或 JSON,并附带证据与复核标记
- 编辑 PDF:合并、拆分、旋转、压缩、加水印、编页码、加密或解锁、OCR
- 处理图片与音视频:缩放、裁剪、格式转换、压缩、裁剪时长、静音
- 转写语音并执行会议工作流
- 提交单任务、最多 100 项的批处理或批量清单,轮询状态并下载结果
- 适用场景
- 适合需要把扫描发票或 PDF 转成结构化数据、做常规 PDF 或媒体转换、或转写录音而不想自建处理流水线的场景。适合已有或将充值 Skavio 公司钱包的团队。
- 运行要求
- 使用提供方的远程 Streamable HTTP 处理端点;通过 Authorization 头使用 OAuth(PKCE)或签发的公司 MCP Bearer 凭证。需要 Skavio 公司账户和预付额度,按页、按起始分钟或按体积计费。需要访问提供方的网络。
安装
在 SourceWeft 中
- 打开 控制台中的 Skavio PDF, OCR & Document Processing MCP,将其添加到工作区。
- 连接 所需的 凭据。它们会被私密存储,绝不会在公开页面上展示。
- 为需要使用其工具的对话启用该服务。
Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。
交给智能体安装
把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体,或 SourceWeft 对话中。智能体会读取该服务器的安装指南,先向你展示发布者、端点和登录方式,经你同意后再把它添加到你的客户端。密钥不会写进任何命令。
阅读 https://sourceweft.com/mcp/io.github.skavio-eu%2Fprocessing/install.md,并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前,先告诉我它的发布者、端点和登录方式,等我确认。不要把 API 密钥写进命令或我会提交的文件;问我密钥应从哪里获取。其他 MCP 客户端
把它添加到你客户端的 mcpServers 配置中。 客户端会在首次连接时完成 OAuth 登录。
{
"mcpServers": {
"skavio-pdf-ocr-document-processing": {
"type": "http",
"url": "https://getskavio.com/processing/mcp"
}
}
}README
Skavio Processing API
New no-account integration example: Fictional Flow invoice to reviewable CSV — no key, no upload or charges. Public EN sample gallery and DE examples. For actual authenticated processing use the maintained quickstarts below and obtain a quote before submitting.
[Skavio Processing MCP — MCPRush publisher listing]
Smithery links: Skavio Processing catalog profile identifies this server. The work-6dqn namespace gateway is a Smithery-managed MCP interface requiring Smithery authorization (connections:execute); its available connections depend on the authorized namespace. It is not the canonical Skavio Processing endpoint. Use https://getskavio.com/processing/mcp with company OAuth for direct connections. Anonymous checks on 2026-10-27 confirmed the gateway's HTTP 401 and Smithery OAuth metadata, but did not establish access to Skavio tools through it. Never send a Skavio company credential to the namespace gateway.
Discover Skavio MCP tools
Cursor / Claude Desktop / Claude Code setup: step-by-step remote MCP connections · ready-to-use Cursor JSON · optional Claude Desktop JSON. Also discover Skavio USDC micro-APIs and their MCP client guide.
Skavio remote MCP server — invoice OCR and extraction, PDF text extraction, document automation, image and media processing, transcription and meeting workflows for AI agents. Discover it in the official MCP Registry, or connect a compatible MCP client to https://getskavio.com/processing/mcp and complete OAuth, or use an explicitly issued company MCP bearer credential. Processing requires authentication.
- MCP setup and all available tools
- MCP pricing and prepaid company credits
- Official MCP server manifest · Smithery link roles and verification
- x402 paid USDC verification micro-APIs · machine-readable x402 discovery
The Processing MCP uses company-scoped prepaid credits; the separate x402 micro-APIs advertise payment-required responses. Neither requires exposing private keys to Skavio.
Card-funded prepaid processing: Processing MCP uses the authenticated company's Skavio wallet (1 USD = 1,000 credits). Only a company administrator funds the correct company wallet through the website checkout; an MCP agent cannot charge a card. The separate Agent Tools Dashboard offers Stripe-hosted $5–$500 card top-ups or USDC for its bound Agent Tools company wallet and eligible processing jobs paid with payment_mode=prepaid. Company credentials are NOT interchangeable. The four public blockchain-proof/DNS tools on https://getskavio.com/mcp and standalone x402 HTTP API calls remain native-USDC x402 only. Card funding and credit debit/refund logic passed controlled tests, but a real paid customer card-to-MCP completed-job E2E has not yet been independently verified.
One company-scoped API for document extraction, PDF and image tools, media processing, transcription and meetings.
- Product and pricing
- Live developer documentation
- Company workspace: register and create an API key
- OpenAPI v1 · Developer guide
- Python quickstart · Node.js quickstart · Bulk guide
- Postman public documentation · Collection
- DEV.to integration article
- Webhook verification
API route version: /v1. Contract snapshot: 1.9.0 + Pricing V2, captured 2026-10-04.
What is new in API 1.7.0
- Monthly charged/reserved usage and separate payment history, bounded CSV and key attribution.
- Atomic company/key budgets for API, batch, bulk and paid web admission.
- Expiring API keys with granular permissions and administrator-only policy writes.
- Queue visibility, persisted alerts, cancellation of unstarted standalone jobs and governance audit.
- Idempotent replay of failed job/batch/bulk webhook deliveries using the frozen original event.
- Structured errors and X-Request-Id correlation; Czech/German/English dashboard additions.
Read the Usage, budgets and recovery guide for authorization, UTC month rules, limits and compatibility.
Storage Fabric introduced in API 1.6.0
- Private S3 sources/results and compatible legacy local files.
- Verified direct uploads and resumable imports from approved customer cloud URLs.
- Company-scoped retention, archive, deletion hold and opt-in separate-bucket backups.
- Manual verified backup/restore, signed result downloads and encrypted import credentials.
- Existing Pricing V2, jobs, batches and bulk contracts remain supported.
Read the Storage Fabric guide for endpoints, production limits and browser upload handling.
Bulk features introduced in API 1.5.0
- Bulk manifests up to 100,000 items with immutable manifest pages of 1–100 items and bounded incremental child admission.
- Storage-aware admission through
GET /v1/storage, declared input/output/scratch envelopes and durable storage reservations. - Resume and failed-item retry for large workloads without resubmitting successful items.
- Crash/restart recovery for scheduler state, reservations, checkpoints and provider-intent reconciliation.
- Fair scheduling across companies while preserving the current ceiling of 4 active platform jobs globally.
- Existing 100-item batches remain available for simpler workloads.
Start with Bulk processing guide or the Python bulk quickstart. This repository contains the public developer kit. The hosted processing service requires a Skavio account and credits.
First result
- Register, create a company profile and generate a read/write key in the company workspace. Copy it once and keep it server-side.
- Upload a source using multipart field
files. - Call
POST /v1/estimatewithout starting processing. - Set
max_creditsto the returnedestimated_credits. - Submit
POST /v1/jobswith a stableIdempotency-Key. - Poll the returned job ID until completed or failed; download through the authenticated result paths.
The included invoice is synthetic, contains no customer data, and can be used for an extraction test.
Python 3.10+
Node.js 22+
Both examples use sample-invoice.txt; set SKAVIO_SOURCE to a PDF/image/text source to change it. They estimate cost before submission and write JSON output into results/.
For intentional processing of a different source or changed payload, use a new idempotency key.
These examples are server-side starters. For the official Python and JavaScript SDK clients, see sdk/README.md; durable workflow pipelines are documented in workflow-guide.md.
Postman
Import the collection or fork its public workspace copy. Set your own local/private environment variable api_key; never save a real key in the public collection. Set source_file to a local source. Run Upload → Estimate → Submit → Job status → Download. Poll every 3–10 seconds until the job is terminal before downloading.
Requests cover the API-key processing endpoints; administration, key creation, billing and webhook registration use an interactive company administrator session in the dashboard.
Before a new business request, clear upload_id, job_id, download_path, max_credits and idempotency_key in your private environment. Preserve the idempotency key when retrying the same submission.
What can I process?
Capabilities snapshot lists exact operation names. Query GET /v1/capabilities for current prices and limits.
Normal file jobs accept one source; PDF merge accepts 2–20 PDFs; Flow can process multiple document sources. Transcribe and Meeting accept one media source. Batch processing supports up to 100 items. API 1.5.0 bulk manifests support up to 100,000 items with incremental dispatch; see bulk-guide.md. Standalone jobs do not support arbitrary operation chaining. Durable linear process/review/archive pipelines are available; see workflow guide.
Credits and price control
All web tools and API share a prepaid company wallet. No unlimited processing entitlement is advertised in this kit.
The server measures bytes, pages and media duration. It reserves the quoted amount on submission, charges that fixed quote on success, and releases the reservation on failure.
Use max_credits to reject unexpectedly expensive submissions and GET /v1/wallet to track balance and reservations.
Current base value: 1 USD = 1,000 credits. Use live pricing and the authenticated dashboard for exact supported packs and taxes. Retired EUR pack examples are not current.
Subscriptions and allowances are product-specific: do not assume former EUR plan tiers apply to MCP; verify the latest live pricing.
Web tools and API share one wallet. Submit reserves the measured quote; success charges it once; a permanent failure releases it. Temporary provider outages keep jobs queued with durable checkpoints and reserved credits. Existing quoted jobs and purchased checkout/subscription offers retain their original amounts.
Canonical live catalog: GET /v1/pricing. Always estimate before submission and use max_credits rather than hard-coding a rate.
Reliability and integration rules
queued → processing → completed | failed; submission returns HTTP 202.- Replaying the identical job request with the same idempotency key returns the same job. Changing its payload returns 409.
- Persist business request, exact payload, key and job ID in your system. If a response is lost, retry with the original key.
- A failed job releases credits; an intentional new attempt uses a new key.
- Provider outages and service restarts preserve durable jobs/checkpoints and reservations; permanent failures release them.
- Webhooks deliver at least once; verify HMAC on the raw bytes and deduplicate event IDs. Polling remains supported.
- Results/source retention defaults to 7 days. Storage Fabric policy, archive and hold control object retention; inspect retain_until and preserve results before expiry.
- Extraction can be wrong. Check evidence and review flags; review line items against the source before downstream accounting.
Limits (current deployment)
120 requests/minute per key/session; up to 100 files/500 MB per upload; 2 GB source storage/company; 100 pages/PDF; 100 queued jobs/company; 4 active platform jobs/company and 4 globally. Transcribe/Meeting input can be up to 10 hours; File Toolbox media up to 2 hours. These are ceilings, not guaranteed throughput or an SLA. Back off on 429 and transient 5xx responses. Current plans share the processing limits.
Errors
JSON errors retain detail and add error_code, request_id and retryable; X-Request-Id provides correlation. Use stable codes rather than matching error text. Validation does not echo secret inputs.
Security and support
Use Authorization: Bearer <your-key> over HTTPS. Never expose keys in frontend apps, URLs, screenshots, commits or shared Postman environments. Company-scoped object IDs cannot access another company.
Create/revoke keys and register webhooks in the dashboard. PDF passwords are secrets; do not log payloads.
Review Privacy and Terms before regulated/sensitive inputs. Source processing can involve external providers; backup/provider retention is separate from platform deletion.
Report integration issues through this repository with synthetic reproductions only, or contact [email protected] privately. Never attach customer documents, keys or billing data to a public issue.
Versioning
Compatible additions remain under /v1; clients should ignore unknown response fields. Breaking changes require a major API version or an announced migration. Changelog.
License
Example code and documentation in this developer kit are MIT licensed. This does not license the hosted backend or grant free use of the hosted service; service use follows its Terms and credit pricing.
Batch processing — API 1.5.0
Submit up to 100 items for any advertised operation through the same processing pipeline. A batch uses one operation/options set and creates ordinary child jobs. Processing remains 4 active jobs per company and 4 globally; a batch does not promise 100 parallel executions.
- Upload files with
POST /v1/uploads; retain their upload IDs. - Quote
POST /v1/batches/estimateusing{"operation":"jpg","upload_ids":["UPLOAD_A","UPLOAD_B"]}. - Submit the same payload to
POST /v1/batcheswithmax_creditsset to the total quote and a persistedIdempotency-Key(8–128 allowed characters). - Poll
GET /v1/batches/{batch_id}. Eachitems[].jobincludes status, credits and result download paths.progressranges from 0 to 1; terminal states arecompleted,partial,failed. - Download
GET /v1/batches/{batch_id}/results.zipafter completion. The archive includes successful results andmanifest.json. Missing/expired files return 410. ZIP output is limited to 1 GB; use individual authenticated downloads for larger results. - Use
POST /v1/batches/{batch_id}/retry-failedwith a new persisted key to create a separately billed batch of failed items. Retry requests are bound to the original batch. Download successes before retention expires.
Credits reserve atomically for the complete batch. Every child has separate accounting; successful items are charged and failed items release their reservation. Replaying the same submission does not reserve credits again. Changing a request while reusing its key returns 409. Batch items count toward the 100 queued jobs per company.
Use either upload_ids (one child per source) or items (explicit groups), never both. PDF merge requires groups with 2–20 PDFs each:
Grouped Flow inputs use the same items shape. Optional webhook_id must belong to your company. webhook_mode accepts batch (parent events), items (child events), or both. Parent events are batch.completed, batch.partial, batch.failed; verify the same signature and deduplicate event IDs. Parent delivery IDs are distinct from child delivery IDs.
GET /v1/batches lists recent company batches. DELETE /v1/batches/{batch_id} deletes terminal result/source specifications while preserving accounting. Source/result retention is seven days. A repeated-failure guard may return 429; contact support rather than continuously resubmitting failing sources. Save exact payloads, keys and returned IDs in durable storage.
Batch starters: Python · Node.js · Batch guide. Set SKAVIO_API_KEY, SKAVIO_UPLOAD_IDS (JSON array from upload), and a persisted SKAVIO_BATCH_IDEMPOTENCY_KEY. The examples poll and print authenticated result paths; archive results before retention expires.
Durable workflow pipelines
Linear process/review/archive steps, checkpointed execution and per-step credit settlement. See workflow guide.
Official SDK clients
Python and Node.js clients are in sdk/README.md. Install from this repository with python3 -m pip install ./sdk/python or npm install ./sdk/javascript. They cover all 92 current OpenAPI operations, durable submission keys, resumable polling, signed webhook verification and authenticated result downloads. Packages are published: pip install skavio-sdk==1.1.0 and npm install @skavio-llc/[email protected].
API 1.9.0 controls and notifications
Bulk pause and cancel require Idempotency-Key and enforce company isolation. Pause prevents new child admission; accepted jobs drain and settle once. Cancel is final and prevents further dispatch. GET /v1/notifications and GET /v1/notifications/settings are tenant-scoped. PUT /v1/notifications/settings requires a company administrator web session. Notifications are disabled by default; email requires configured transport. Frozen events are deduplicated and retried within a bounded delivery window. Email follows the current company billing address and supports cs/de/en.
USDC payment verification for AI agents
Skavio USDC Payment Verification — x402 API inspects native Circle USDC transfers, matches an exact recipient and decimal amount, and generates blockchain evidence PDFs. Use it for merchant payment reconciliation, accountant transaction review and agent workflows. PDF evidence is not a tax invoice or a guarantee of third-party authorization.
- HTTP Verify: 0.02 native USDC; Match: 0.03 USDC; PDF: 0.05 USDC. Payment settles separately on Base through x402 v2. Inspection supports Base, Polygon, Arbitrum, Optimism, Ethereum, Avalanche and finalized Solana.
- API quickstart, schemas, errors and client configurations · three-route discovery · OpenAPI.
- Skavio USDC Payment Verification & Proof MCP, version 1.1.5. Stable registry namespace and tool IDs are unchanged. Connect using
https://getskavio.com/mcp(Streamable HTTP). - MCP has four paid blockchain/domain tools (
verify_native_usdc,match_native_usdc_payment,usdc_transfer_evidence_pdf,domain_dns_tls_audit), two free discovery tools, and twelve separate authenticated AgentTools processing/job/wallet relays. The DNS/TLS audit is 0.05 USDC and is not a vulnerability scan. - A standard MCP client without a wallet x402 extension cannot automatically pay. Discovery is free. Paid MCP retries preserve
params._meta["x402/payment"]; HTTP retries usePAYMENT-SIGNATURE. Never send a private key or seed phrase. No custody or automatic spending is created by adding a server. - PayAPI publishes separate live listings and historical settlement receipts for HTTP Verify, 0.02 USDC, exact recipient/amount Match, 0.03 USDC, and blockchain PDF evidence, 0.05 USDC. Each route was independently read back on 2026-10-17, and each prior Base USDC receipt was checked read-only. No new payment was made for this documentation update. This is blockchain evidence, not a tax invoice.
- RPC availability and finality affect results. No fixed latency SLA is promised. EVM requests use a 32-byte 0x hash; Solana uses its transaction signature. Match takes a recipient wallet and a positive exact USDC decimal string with at most six decimal places.
Cursor .cursor/mcp.json:
Claude Code:
Claude Desktop with Node.js and a reviewed stdio-to-HTTP bridge:
The bridge may download a third-party package and does not supply wallet payment support. Current Desktop remote connectors may accept the HTTPS URL directly.
First unpaid request (no signing or transfer):
Expect HTTP 402 and decode PAYMENT-REQUIRED. Inspect the exact amount, Base USDC asset and recipient before authorizing an x402 wallet client; only then retry the same request. After an ambiguous response, inspect recovery state instead of blindly paying again.
Separate Skavio PDF, OCR & Document Processing MCP, version 0.2.8, uses company OAuth at https://getskavio.com/processing/mcp and shared prepaid credits. Processing docs explain company ownership and measured jobs. Optional AgentTools relays require the caller's own key and current backend availability; preserve the idempotency key, poll status, download results and verify SHA256. HTTP /x402/v1/processing stays disabled.
来源:README.md,提交 365affd
工具
0版本历史
5- v0.2.9最新Nov 7, 2026
- v0.2.8Oct 29, 2026
- v0.2.7Oct 17, 2026
- v0.2.6Oct 14, 2026
- v0.2.3Oct 14, 2026
评分
已满足适用检查项的 85%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通必需25 分
- 已声明认证方式必需10 分
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分