Red Pen Editor

A free, local-first editorial judgment skill, npm CLI, and MCP server. Version 0.2.1. MIT licensed. No hosted service, paywall, telemetry, or provider API key.

The main question is: What can be deleted without diminishing the piece's useful meaning, accuracy, or voice?

This project focuses on substantive cuts: repetitive claims, unnecessary paragraphs, empty introductions, repeated conclusions, and excessive explanations. It does not impose a word-count target or automatically ban stylistic phrases.

Three ways to use it

1. Agent Skill: no model installation needed

Copy skills/red-pen-editor/ into the skills folder of a compatible agent. The SKILL.md contains the editorial method and works using the agent's existing model. The included code is optional.

Example for an agent with a project-level .agents/skills folder:

mkdir -p .agents/skills
cp -R skills/red-pen-editor .agents/skills/

2. npm CLI: review or edit a file locally

Requires Node.js 20+. Install from npm:

npm install -g red-pen-editor
red-pen-editor inspect draft.md
red-pen-editor brief draft.md

Or run it once without installing:

npx red-pen-editor inspect draft.md

For actual contextual editorial review, use a local Ollama model:

ollama pull llama3.2
red-pen-editor review draft.md --model llama3.2
red-pen-editor edit draft.md --model llama3.2 > revised.md

Ollama must be running. inspect is only a surface-level diagnostic, not a substitute for an editor. It flags near-identical sentences and paragraphs that share most of their content words. review returns the model's proposed cuts and accepted/rejected operations; edit prints the revised text and does not overwrite input. Keep a human review before publication. You can override the local model with --model or EDITOR_MODEL.

3. MCP server: editorial judgment from any connected agent

Add the dependency-free stdio server to your MCP host. Example client config (adapt the fields to your host):

{
  "mcpServers": {
    "red-pen-editor": {
      "command": "npx",
      "args": ["-y", "-p", "red-pen-editor", "red-pen-editor-mcp"]
    }
  }
}

If you installed the package globally, use "command": "red-pen-editor-mcp" with no args.

| Tool | Function | Model requirement | | --- | --- | --- | | editorial_brief | Returns a substantive editorial rubric and possible duplication signals | Host agent only | | editorial_apply | Applies host-proposed, exact-span edits with safety checks | Host agent only | | editorial_review_local | Uses Ollama to make and explain edit proposals | Local Ollama | | editorial_edit_local | Uses Ollama and returns a guarded edited version | Local Ollama |

The host agent can make editorial judgments and call editorial_apply itself. No second model, subscription, or outside API is needed for that path. The stand-alone CLI uses a local model for substantive editing. The MCP server supports the 2025-11-25 legacy handshake and the 2026-07-28 stateless stdio request pattern.

Editing behavior

1. Read for the audience and purpose.
2. Identify paragraphs or claims with no unique contribution.
3. Recommend deletion before sentence-level compression.
4. Require an exact source span for each cut or replacement.
5. Reject ambiguous spans, overlapping edits, and edits that would remove the last occurrence of a protected number, written or ISO date, URL, email, or explicit preserve string. Literals are matched as whole tokens, so 1 does not count as kept because 10 survives.
6. Allow a justified cut of a protected detail only when the edit names it in allowLoss. The accepted edit records what it released in allowedLoss. allowLoss cannot release a preserve string, and edits proposed by a local model have it stripped.
7. When a deletion removes the start of a sentence, capitalize the word that now opens it.
8. Return an edited version and rejected-edit log; review meaning and grammar before publishing.

Host-agent example

The host sends this edit proposal to editorial_apply:

{
  "text": "There are several factors to consider. The change takes effect May 1.",
  "edits": [{
    "before": "There are several factors to consider. ",
    "after": "",
    "kind": "throat-clearing",
    "reason": "The preface adds no relevant information."
  }]
}

The returned text is The change takes effect May 1.

Cutting a paragraph that holds a unique detail

A background paragraph can be worth cutting even though it holds the only mention of a year. Name the detail in allowLoss:

{
  "before": "The program began in 2019 as a pilot in two neighborhoods. ",
  "after": "",
  "kind": "overexplaining",
  "reason": "Residents need the current rule, not the program's history.",
  "allowLoss": ["2019"]
}

Without allowLoss, the edit is rejected with the detail it would remove.

What the guard can and cannot prove

The package checks exact source spans and literal anchors. It cannot prove semantic equivalence, detect every changed implication, or know whether a distinctive passage was intentional. A local model can also make poor editing decisions. editorial_apply operates on text supplied by the caller; it does not access a file system or send content to a server. Only the optional local Ollama call uses HTTP, by default restricted to loopback.

Privacy

The skill and host-agent MCP tools do not contact a third-party service. The Ollama integration sends text to a loopback HTTP endpoint only. If the host agent itself uses an online model, its existing data-handling arrangements still apply. Review source confidentiality before sending sensitive content to any model.

Development

From a clone of this repo, with no npm install:

npm test
npm run demo
npm pack --dry-run

Tests include semantic-risk guardrails, span validation, deterministic review fixtures, and an MCP JSON-RPC smoke test. The optional local-model integration can be tested against an Ollama installation. There are no npm dependencies.

Roadmap

• Add a human-reviewed benchmark of paragraph-level substantive cuts.
• Evaluate usefulness and meaning retention, not just words removed.
• Add a document-aware span format for prose, Markdown, and citations.
• Test with multiple agents and local models.

MIT License.