AI Humanizer MCP Servers Compared: What the Public Contracts Expose

This dated, non-exhaustive review verifies first-party MCP humanization tools from nine vendors. That count applies only to the vendor-controlled pages checked and listed below; it is not a market total. Their public contracts differ in payload, transport, authentication, execution shape, and request limits.

The checked sample is HumanPen, WriteHuman, Phrasly, Duey AI, Ryter Pro, StealthGPT, Walter Writes, HumanText Pro, and Undetectable AI. Its reproducible boundary is the exact set of vendor-controlled pages, repositories, packages, and schemas named in the source list below. A product entered only when one of those sources named a humanization tool. This is a dated source ledger, not an exhaustive market search. I did not authorize an account, submit text, upload a document, or run a detector.

HumanPen publishes this comparison and is one of the products in it. Treat its product statements as first-party claims, not independent test results.

The meaningful question is not whether a page says MCP. It is whether the published tool contract names a humanization operation and defines what crosses the boundary.

Four labels keep unlike integrations from collapsing into one list:

1. First-party MCP means the product vendor publishes or controls the server and documents a humanization tool. This is the category used for verified rows in this dated sample.
2. First-party API means the vendor exposes a REST or streaming interface. An API can sit behind an MCP server, but an API by itself is not proof that an MCP server exists.
3. Third-party wrapper or actor means another developer operates an adapter around a product or a generic workflow. Its maintenance, authentication, and data route belong to that developer's contract.
4. Label without a verified humanization tool means an MCP package or repository exists, but the checked public schema does not establish a humanize operation.

The rule is deliberately narrow. It does not say an unverified capability is impossible. It says the public material checked on the stated date did not prove it.

Discovery evidence tells you where to look. A tool schema tells you what an agent can actually call.

The table records public contracts checked on August 26, 2026. A published limit is included only when the relevant MCP or product page attached a number to that workflow.

These rows do not measure writing quality, detector outcomes, or institutional acceptability. They describe integration surfaces. Limits can also use different units: words, characters, plan allowances, or no published number at all.

Convert every limit to its original unit before planning a workflow. A 30,000-character input is not a 30,000-word input.

HumanPen and HumanText Pro publish local stdio packages. Ryter Pro also links an optional local stdio package, while recommending its hosted Streamable HTTP endpoint. For a local package, the MCP process starts on the user's machine, reads an API key from the environment, and communicates with the client over standard input and output. That describes where the connector runs. It does not mean the rewrite model or the submitted content stays on the machine.

WriteHuman, Phrasly, Duey AI, Ryter Pro, StealthGPT, Walter Writes, and Undetectable AI publish hosted connections. WriteHuman and Duey AI document OAuth 2.1 with PKCE, while Phrasly uses account OAuth approval. Walter Writes uses OAuth workspace authorization. StealthGPT offers OAuth for web connectors and a bearer token route for IDE or CLI clients. Undetectable AI documents both browser OAuth and a static API key. Ryter Pro uses `X-API-Key`.

Authentication affects setup and credential handling, but it does not answer every privacy question. Before using any server with restricted material, check the vendor's current retention, training, deletion, and subprocessor terms separately. Those policies were outside this contract comparison.

HumanPen's MCP is also different at the payload boundary. Its humanizer tool names DOCX and PPTX file paths and writes the result to a local output path. The other verified humanization tools in this table document text-oriented inputs. For HumanPen setup and the file-path flow, use the separate guide to humanizing a document inside Claude Code.

Local stdio is a connector topology, not a claim that document processing is offline.

A humanization call can return content immediately or return a receipt for work that is still running. That difference determines what the agent must do after the first tool call.

StealthGPT makes the asynchronous step visible in MCP: `humanize_text` returns a `runId`, and `get_run_status` retrieves the eventual result. HumanPen's REST API follows a file-job sequence: multipart upload, job receipt, polling, then editable-file download. Its MCP tool wraps that sequence and writes the completed file to disk.

Among the added contracts, Phrasly returns the result in chat and saves it to the user's Phrasly history. Duey AI lists `humanize_text` separately from its Session scheduling and status tools. Ryter Pro documents `humanize_text` without a separate job or polling contract on the checked MCP page. The separate APIs widen the picture: Undetectable AI's REST v2 route submits text and returns an ID for polling, Walter Writes documents synchronous and asynchronous API jobs, and WriteHuman publishes a separate JSON REST API whose limits should not be substituted for MCP limits.

For a chat answer, a text result may be the deliverable. For a thesis, report, or slide deck, the deliverable is often the file itself. In that case, record the source path, output path, job status, and returned file type, then perform a document-level acceptance check. The fuller Word document review workflow covers that post-run step.

A successful tool call proves that the tool returned successfully. It does not prove that every citation, field, table, note, or slide object survived correctly.

Text2Go's public repository and npm package are named `ai-humanizer-mcp-server`. That title looks decisive in a search result. In the public TypeScript source checked for version 1.0.7, however, the server registered one tool named `detect` and called a text-detection endpoint. I did not find a registered humanization tool in that examined source.

The defensible classification is therefore MCP package exists; humanization not verified. It would be too strong to say the capability can never exist, because the package, remote endpoint, or documentation may change after the verification date.

Community actors are a separate case. The Apify listing `Humanize AI Text` identifies its developer as Neatrat and labels the actor community maintained. It may expose an automation route, but it is not a first-party MCP or API contract from a named humanizer vendor. Claims on the actor page belong to that listing, not automatically to any underlying product.

For any directory result, open the current schema and answer three questions: who operates it, which tools are registered, and which endpoint receives the content. If one answer is missing, preserve the uncertainty in the article instead of filling it from the title.

Not verified is an evidence status, not a verdict about what may be built later.

Start with the artifact and control boundary, then check the product name.

1. Identify the deliverable. Decide whether the agent must return text in chat or an editable DOCX or PPTX file.
2. Read the actual tool schema. Confirm the humanize tool, required fields, minimum input, and output object.
3. Record transport and ownership. Distinguish local stdio, hosted HTTP, and a community-operated actor.
4. Match authentication to the client. OAuth avoids copying a long-lived key into some clients; API-key routes may be simpler for controlled automation.
5. Handle the execution shape. If the first response is a job or run ID, require polling, terminal-state handling, and a clear output path.
6. Keep units attached to limits. Do not convert characters, words, credits, and monthly allowances into one unlabeled number.
7. Define acceptance outside the server. Recheck meaning, citations, numbers, formatting, and the policy that governs AI assistance.

Among the contracts reviewed here, HumanPen is the file-oriented option: its MCP README documents DOCX and PPTX paths rather than a pasted-text field. HumanPen publishes this article, so verify that first-party claim against a disposable file before using it on a real deliverable. Its public workflow also allows selected passages so the review surface can stay smaller.

Eligible results can continue lowering AI for free.

Is an AI humanizer API the same thing as an MCP server? No. A REST API defines application requests and responses. An MCP server exposes tools and resources in a protocol an agent client can discover. A vendor may publish both, one, or neither.

Which checked MCP server accepts document files? HumanPen's public MCP README documents DOCX and PPTX file paths for its humanization tool and a result file on disk. A product's web upload feature does not establish that its MCP tool accepts files.

Does a local stdio MCP keep my document offline? Not necessarily. Stdio describes communication between the client and the local connector. Read the connector source and vendor network contract to see whether content is sent to a remote API.

Is OAuth always preferable to an API key? They solve credential handling differently. OAuth can avoid pasting a static key into compatible clients, while an environment-scoped API key can suit controlled local automation. Neither choice alone establishes data retention or content privacy.

Why is Text2Go not listed as a verified humanizer MCP? The checked package source registered a detection tool, not a humanization tool. The classification can change if a later public version exposes and documents a humanize operation.

Can an MCP humanizer guarantee a detector result? No such conclusion was tested here. The public contracts describe calls, payloads, job states, and limits. They do not establish a future result from Turnitin or any other detector.

These pages and public repositories were checked on August 26, 2026. Reopen the live contract before implementation because endpoints, schemas, limits, and authentication can change.

• HumanPen: developer page, humanizer workflow, REST API reference, and MCP repository README
• WriteHuman: MCP page, API page, and MCP launch documentation
• Phrasly: MCP page
• Duey AI: MCP page
• Ryter Pro: MCP page
• StealthGPT: MCP page, MCP integration reference, and API page
• Walter Writes: MCP overview and developer documentation
• HumanText Pro: MCP page, npm package, and public repository
• Undetectable AI: MCP page, developer page, and Humanization API v2 guide
• Text2Go: public repository and published server source
• Third-party example: Apify Humanize AI Text actor