circe-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@circe-mcp check this text: It's not just code, it's craft."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
circe-mcp
An MCP server that checks text against the Circe Editorial Standard (CES) v0.3 and returns each violation with its rule ID, line and column, severity, confidence, and a suggested fix. Every check is deterministic pattern matching. The server makes no LLM calls, so a check costs nothing to run and gives the same answer every time.
The standard catches the writing patterns that read as machine-generated: em dashes, hedges, hollow intensifiers, trailing contrast clauses, negate-then-elevate constructions, throat-clearing openers, and the rest. This server is the standard's automated reviewer. The standard is a working draft by Avalynn Circe; explain_rule carries the text of every rule the server enforces.
Install
The server runs with npx, so there is nothing to install.
Claude Code
claude mcp add circe-mcp -- npx -y circe-mcpClaude Desktop
Add the server to claude_desktop_config.json:
{
"mcpServers": {
"circe-mcp": {
"command": "npx",
"args": ["-y", "circe-mcp"]
}
}
}Cursor, Windsurf, and other stdio clients use the same command and arguments. For a hosted deployment, npx circe-mcp http --port 3000 serves the same tools over Streamable HTTP at /mcp.
Node.js 20 or later is required.
Related MCP server: slop-guard
Tools
Tool | Input | Output |
|
| Violations with |
| optional | Every rule with its ID, name, level, and a one-line summary. Optional rules are marked. |
|
| The full rule: pattern, reasoning, replacement strategy, exceptions, a before/after example, and the tokens the checker uses. |
|
| The text with safe fixes applied, the list of edits, and the violations left for the writer. |
|
| Statistics with no verdict: sentence lengths per paragraph, rhythm, opener echo, specificity density, paragraph density, and over-enthusiasm signals. |
Each tool returns a readable report as text and the full result as structured content, so a client can show one and a model can parse the other.
An example
Input to check_text:
It's not just code, it's craft. We built it in practice.Output:
1:6 error CES-C-008: Trailing contrast clause 'not just'. End the sentence at the positive point. (also CES-Q-001, CES-C-001)
1:45 error CES-V-007: Flourish 'in practice'. Delete it or make the claim as a plain statement.
2 error(s), 0 warning(s), 0 suggestion(s).When violations overlap, the most severe one is reported and the others are listed under related. The first sentence trips three rules in one span; the report shows it once.
Rules
Severity follows the standard's tiers. MUST AVOID rules report as error, SHOULD AVOID rules as warning. Each violation also carries a confidence level, because a token match is sometimes a guess.
Rule | Level | Default | Catches |
CES-P-001 No em dashes | error | on | The em dash character (U+2014). |
CES-Q-001 No reflexive hedges | warning | on |
|
CES-Q-002 No hollow intensifiers | warning | on |
|
CES-C-001 No negate-then-elevate | warning | on |
|
CES-C-002 No stylistic triple lists | warning | off | Three single words joined |
CES-C-003 No throat-clearing runways | warning | on |
|
CES-C-004 No trailing bows | warning | on |
|
CES-C-005 No so-what gloss | warning | on |
|
CES-C-008 No trailing contrast clauses | error | on |
|
CES-V-001 No resume cliches | warning | off | The standard's phrase list: |
CES-V-006 No chatty asides | warning | on |
|
CES-V-007 No punchy-for-punchiness'-sake flourishes | error | on | Implied inferior baselines: |
CES-V-008 No gratuitous trailing clauses | warning | on | Trailing minimizers: |
Two rules are off by default because they guess more than the others. Pass includeOptional: true to run them, or name them in rules.
The sub-forms of a rule come back in the form field. An inferior baseline reports as CES-V-007 with form: inferior-baseline, which is where the standard files it.
What it misses
Pattern matching catches stock phrasing. It cannot judge whether a short declarative is a pronouncement or the one line that states a paragraph's point (CES-C-007), whether a contrast carries facts the reader needs, or whether a paragraph is vague perfection (CES-V-004). The flourish and sass detectors catch the signature phrases and will miss cases that need a reader. measure_text reports the numbers behind the judgment calls and leaves the call to the writer.
Rules that need a human stay out of check_text: the deletion test for CES-C-003 through CES-C-007 beyond their stock forms, manufactured intimacy (CES-V-002), scene-painting (CES-V-003), and the sections of the standard that cover factual accuracy, document structure, and typography.
Exceptions
A contrast that carries facts the reader needs is allowed under CES-C-008, such as a migration from one database to another. Mark a justified exception with an inline directive in an HTML comment, which renders as nothing:
<!-- ces ignore CES-C-008 -->
The team moved the reports from MS Access to SQL Server rather than rewriting them.
<!-- ces end -->A directive takes one or more rule IDs, in any case, separated by spaces or commas. Without an end marker it runs to the end of the text. <!-- ces end CES-C-008 --> closes one rule and leaves the others open. <!-- ces off --> and <!-- ces on --> silence every rule for a section. For a single call, pass the rule IDs to skip in the ignore parameter.
Markdown
Fenced code blocks, inline code, link destinations, HTML tags, and comments are skipped. Front matter is checked, because a page description is prose. Pass checkCode: true to lint code comments too (CES-K-003). Line and column numbers refer to the original text.
Fixes
fix_text applies the fixes that have one safe answer. An em dash becomes a comma, with the spaces around it absorbed. A hedge or hollow intensifier is deleted along with one adjacent space, and the next word is capitalized when the deleted word opened the sentence. A fixable word inside a larger construction is left in place, because the construction needs a rewrite, not a deletion. Review an em dash fix that joins two independent clauses; the standard allows a semicolon or a new sentence there, and the fixer cannot tell which.
Every other rule is report-only.
The rules file
rules/ces.yaml is the single source of truth. Each rule carries its ID, severity, description, message, suggestion, exception, a before/after example, and its tokens with per-token severity and confidence. Heuristic rules name a detector in src/detectors.ts. The server loads the file at startup and validates every entry, so a typo in a rule fails fast with a message that names the rule and the field.
Adding a rule is one YAML entry plus a fixture with a text it flags and a text it passes. The test suite refuses a rule without a fixture and a fixture without a rule.
Command line
circe-mcp # MCP server on stdio
circe-mcp http --port 3000 # MCP server over Streamable HTTP
circe-mcp check README.md docs/ # lint files; exit 1 on an error-level hit
circe-mcp fix draft.md --write # apply the safe fixes in place
circe-mcp measure draft.md # statistics as JSONcheck takes --fail-on warning to make warnings fail too, --include-optional to run the optional rules, and --json for machine-readable output.
Development
npm install
npm test # Vitest: fixtures per rule, positions, exceptions, fixes, MCP round trip
npm run typecheck
npm run build # compiles to dist/; the server reads rules/ces.yaml from the package root
npm run check # typecheck, test, buildEvery rule has at least one fixture that it flags and one that it passes, and its own before/after example from the standard is tested the same way. The README and the tool descriptions are checked against the standard in the test suite and in CI, so this file passes check_text with zero violations, optional rules included. A 10,000-word check completes in under 200 ms.
DECISIONS.md records the design decisions and the reasoning behind them, including where each rule's ID comes from and why two rules ship off by default.
License
MIT. See LICENSE.
Available Tools
5 toolscheck_textCheck text against the Circe Editorial StandardARead-onlyIdempotent
Check text against the Circe Editorial Standard (CES) v0.3. Returns each violation with its rule ID, severity, confidence, line and column, the matched text, a message, and a suggested fix. Overlapping violations are reported once, with the others listed under 'related'.
Inputs: text (required), rules, ignore, severity, +2 more. Returns: counts, rulesRun, violations
explain_ruleExplain a ruleARead-onlyIdempotent
Show the full text of one CES rule: the pattern, why it is a tell, the replacement strategy, exceptions, a before/after example, and the tokens the checker uses.
Inputs: ruleId (required). Returns: id, name, rating, tokens, +8 more
fix_textApply safe fixesARead-onlyIdempotent
Apply the safe fixes to a text: em dashes become commas, and reflexive hedges and hollow intensifiers are deleted. Returns the fixed text, the list of edits, and the violations left for the writer.
Inputs: text (required), rules, ignore, includeOptional. Returns: text, applied, remaining
list_rulesList rulesARead-onlyIdempotent
List the CES rules this server enforces, with ID, name, severity, and a one-line description. Optional rules are marked.
Inputs: severity. Returns: rules, version, standard
measure_textMeasure textARead-onlyIdempotent
Report statistics about a text with no verdict: sentence lengths and rhythm per paragraph, opener echo, specificity density (CES-V-004), paragraph density (CES-S-002), and over-enthusiasm signals (CES-V-005).
Inputs: text (required). Returns: wordCount, paragraphs, exclamations, sentenceCount, +3 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool has a clearly distinct role: check_text finds violations, measure_text reports statistics with no verdict, fix_text applies safe fixes, list_rules enumerates rules, and explain_rule details a single rule. The check_text/measure_text boundary is explicitly clarified by 'no verdict' vs violation reporting, leaving no real overlap.
All five tools follow a consistent verb_noun snake_case pattern (check_text, fix_text, measure_text, list_rules, explain_rule). Verb pairing with noun reads predictably across the set.
Five tools is well-scoped for an editorial-linting server, covering the natural verbs (check, fix, measure, list, explain) without redundancy. Each tool earns its place in the workflow.
The core lifecycle is covered: discovery (list_rules, explain_rule), analysis (check_text, measure_text), and remediation (fix_text). Minor gaps exist, such as batch/file-level checking or rule configuration, but agents can work around them.
Related MCP Connectors
DraftCheck: flags AI-writing patterns in prose with fix hints. Deterministic linter, not a detector.
Check AI work against requirements and return structured verdicts, findings, and repair steps.
Free mechanical checks for AI text: unnamed counts, dangling references, bad arithmetic, misquotes.
Checks RAG citations against their sources (CrossSource method), plus claim checks. Read-only, free.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI hosts to check text for grammar, spelling, and style errors, measure readability, and generate tone-adjusted rewrites through structured tool calls. Every tool returns predictable, typed output for reliable routing and results.MIT
- FlicenseNot gradedqualityAmaintenanceEnables agents to analyze text or files and score prose from 0–100 for formulaic AI writing patterns using rule-based heuristics, returning violations, band labels, and concrete advice.-
- FlicenseNot gradedqualityAmaintenanceEnables agents to analyze text or files and receive numeric slop scores, band labels, specific violations with context, and concrete writing advice, using rule-based diagnostics with no LLM judge or API calls.-
- FlicenseNot gradedqualityAmaintenanceEnables agents to score prose from 0–100 for formulaic AI writing patterns using rule-based linting, returning diagnostics such as band labels, violations with context, and advice. It can analyze inline text or files without API calls or LLM judges.-