redline

README

redline

⚠️ Alpha quality. redline is under active development. The core scan pipeline works, tests are green, and the report format is stable — but JavaScript rendering isn't supported, some niche subcommands aren't implemented yet, and breaking changes may land in any release until v1.0.0. Pin a specific tag in production-y use.

redline is a single-binary command-line tool that audits a website's content against your canonical brand messaging and Generative Engine Optimization (GEO) target prompts. It crawls the site, uses a local LLM (Ollama by default) to judge every page, and produces an agent-actionable report — each finding tied to a verbatim quote, a location hint, and a suggested fix — so a downstream LLM can apply UPDATE / REWRITE / DELETE edits across the site. Both report.json and report.md are written on every successful scan.

Other SEO/GEO tools tell you what to add to your site. redline tells you what's already on your site that no longer matches what you're trying to say — and hands the per-page fix-up plan to the next agent in your pipeline.

Why redline

• Messaging-coherence audit, not just monitoring. Every other GEO tool tells you how LLMs currently mention your brand. redline tells you which pages on your own site are dragging that mention quality down — and what to fix.
• Agent-actionable output. The report.md is designed to be eaten by another LLM. Every finding includes verbatim quoted text, a location hint, and a structured edit plan (preserve / remove / rewrite / add). Plug it into an editor agent and let it apply the fixes.
• Local-first. Default invocation uses Ollama on your machine. No cloud API keys required. No page content leaves localhost. Cloud providers (Anthropic, OpenAI, Voyage) are opt-in for users who want speed or model quality.

What the report looks like

Excerpt from a real scan — every non-Aligned page gets a section like this in report.md:

### 1. REWRITE — https://example.com/products/container

**Priority:** 56.3 | **Confidence:** 0.90 | **Word count:** 44

**Primary label:** Contradictory
**Secondary labels:** Stale

#### Page summary

- **Current focus:** Positions Container Security as a runtime threat-detection product.
- **Should focus on:** Container handles image and Kubernetes vulnerability scanning;
  runtime threat detection is out of scope.

#### Findings

##### Finding f1 — contradicts_canonical (severity: HIGH)

**Quoted text on page:**
> Continuous runtime threat detection across AWS, Azure, and GCP.

**Suggested fix:** Delete this sentence entirely. Runtime CSPM is not in scope.

#### Edit plan

**Remove:** The runtime threat-detection sentence
**Rewrite:** H1 headline → "Container Security: Image and Kubernetes vulnerability scanning"
**Add:** A cross-link to AppRisk for compliance mapping

The quoted_text field is guaranteed to be a literal substring of the page body, so a downstream editor agent can locate it with a string match.

The full report.json is the machine-readable equivalent — validates against an embedded JSON Schema, deterministic ordering, every collection explicitly sorted at the boundary.

Quickstart

redline is local-first. The default invocation needs zero cloud API keys; it talks to a local Ollama server.

1. Install redline

Via Homebrew:

brew install rdegges/tap/redline

Via go install:

go install github.com/rdegges/redline/cmd/redline@latest

Or grab a pre-built binary from the releases page.

2. Install Ollama (one-time)

# macOS
brew install ollama
ollama serve         # leave running in another terminal

# Linux
curl -fsSL https://ollama.com/install.sh | sh
ollama serve

Minimum supported version: 0.5.0.

3. Pull the default models (one-time)

ollama pull qwen3:30b           # judge model (~18GB)
ollama pull nomic-embed-text    # embedding model (~275MB)

4. Author a prompts.yaml

version: "1"

prompts:
  - id: category-leader
    text: "Who are the leading vendors in [your category]?"
    weight: 1.5
  - id: best-for-developers
    text: "What is the best developer-first [category] tool?"
    weight: 1.5

canonical_messaging:
  - title: "What we are (2026)"
    body: |
      [Your brand] is the [category] platform for [primary ICP].
      We help [who] [do what] by [how].
  - title: "What we are NOT"
    body: |
      We are not a [adjacent category] tool, not a [another category]
      product. Retired names like "[old product name]" should be flagged
      as STALE.

Working templates for three verticals (SaaS product, dev tool, security vendor) live in examples/.

5. Run the scan

redline scan --site https://example.com --prompts prompts.yaml

The default invocation writes both report.json and report.md to ./redline-report/. Add --format csv to also produce report.csv.

If something's off, run the pre-flight diagnostic:

redline doctor

Where redline fits in the landscape

Tool	Crawls site	Brand-msg input	Per-page edit plan	Machine-readable output	Hosting	License
redline	✅ sitemap + BFS	✅ explicit prompts + canon	✅ UPDATE/REWRITE/DELETE	✅ JSON + Markdown	Local-first / self-host	MIT (OSS)
Profound	Partial	Implicit	❌ generates new content	Unclear	Cloud only	Proprietary
AthenaHQ	✅	Implicit	Partial	❌ dashboards/BI	Cloud only	Proprietary
Peec.ai	❌ tracks LLM outputs	❌	❌	✅ CSV + API	Cloud only	Proprietary
Otterly	Partial (crawlability)	❌ prompt library	Partial	Unclear	Cloud only	Proprietary
Writer.com	❌ governs new content	✅ voice / style guide	❌ drafts, not edits	Partial	Cloud only	Proprietary
MarketMuse	✅ content inventory	Partial (topic, not voice)	✅ optimize briefs	Unclear	Cloud only	Proprietary
Adobe LLM Optimizer	✅ crawl + index	Implicit	✅ auto-suggest	✅ A2A + MCP	Cloud only	Proprietary

Honest read: the closest direct competitors to redline are Adobe LLM Optimizer and AthenaHQ — both crawl your site and emit page-level GEO recommendations, but they're cloud-only SaaS at mid-to-enterprise price points. Profound / Peec.ai / Otterly are a different category: they monitor how LLMs currently mention your brand rather than auditing your site for fixes. Writer.com is the most orthogonal — it governs new content generation against a brand voice; it doesn't crawl and rewrite existing pages. redline's distinct wedge is the combination of (1) crawl-the-whole-site, (2) user-supplied prompts + canonical messaging as first-class inputs, (3) structured agent-consumable edit plans, and (4) MIT-licensed local-first execution.

Command reference

redline scan       Crawl, judge, embed, and write a report (full pipeline)
redline crawl      Run only the crawl stage (no LLM calls)
redline judge      Run only the LLM judge stage against an existing DB
redline embed      Run only the embedding + dedup stage
redline report     Regenerate a report from an existing DB
redline doctor     Diagnose runtime, DB, and provider readiness (use `--run` to also inspect a scan run)
redline models     Local Ollama model info (list, recommend)
redline version    Print version information

Selected flags

Flag	Default	Description
--site URL	required	Root URL to crawl
--prompts PATH	required	Path to prompts.yaml
--output DIR	./redline-report	Output directory
--format	json,markdown	Add csv to also emit a CSV
--llm-provider	ollama	ollama (default) or anthropic
--model	qwen3:30b	Judge model (provider default)
--ollama-url	http://localhost:11434	Ollama HTTP endpoint
--find-duplicates	true	Run Pass 2 (embedding similarity dedup)
--max-pages	5000	Crawl page cap (0 = unlimited) — bounds cost for cloud runs
--dry-run	false	Crawl + estimate only; no LLM/embed calls
--resume	true	Resume an interrupted run against the same DB

Run redline <cmd> --help for the full flag set on any subcommand.

Cloud (Anthropic) mode

export ANTHROPIC_API_KEY=sk-...
export OPENAI_API_KEY=sk-...      # only if --embedding-provider=openai

redline scan \
  --site https://example.com \
  --prompts prompts.yaml \
  --llm-provider anthropic \
  --model claude-sonnet-4-6 \
  --embedding-provider openai

Cloud spend is bounded by --max-pages (default 5000). For finer cost control, use your provider's account-level spending limits (Anthropic Console / OpenAI usage settings) — redline does not maintain its own pricing tables. Token counts per call are recorded in api_calls and surfaced in the report.

Pause / resume

Long scans (4–8 hours against a 2000+ page site) are designed to survive Ctrl-C, machine sleep, network blips, and Ollama restarts. State is committed to the SQLite DB at per-item granularity. Re-invoke the same command against the same --db path; already-fetched URLs and already-judged pages are skipped automatically.

Reports

Every successful scan writes:

• report.json — primary machine-readable artifact, validates against the embedded JSON Schema.
• report.md — agent-actionable playbook with quoted findings, edit plans, and rationales for every non-Aligned page.
• report.summary.txt — 10-line stdout summary, also printed to stderr.
• copies/report.json.sha256, copies/report.md.sha256 — sidecar checksums.

--format csv adds report.csv (RFC 4180, pipe-delimited multi-value columns).

Architecture

redline is a five-stage pipeline (crawl → extract → judge → embed/dedup → report) backed by a single SQLite database for state, resume, retries, and audit. Every stage is idempotent — re-running a stage on the same DB doesn't duplicate work.

For a deeper dive — package layout, lifecycle of a scan, load-bearing design decisions, where to start reading code — see docs/architecture.md.

Troubleshooting

• ollama unavailable / connection refused → Start Ollama: ollama serve.
• ollama model not pulled → ollama pull qwen3:30b and ollama pull nomic-embed-text.
• Configured prompts.yaml didn't validate → redline scan --print-schema shows the embedded JSON Schema.
• A run aborted mid-scan? Inspect what stage it stopped at with redline doctor --run latest — it summarizes counts and suggests the right resume command. Re-running redline scan with the same --db path picks up where it left off automatically. Add --json for scriptable NDJSON output: the per-check lines are followed by a single {"kind":"run_inspection",...} record covering the run's state.

Contributing

PRs welcome — see CONTRIBUTING.md for local dev setup (Go via asdf, optional Ollama for live tests), the test/lint/coverage make targets, the conventional-commits style, and the PR workflow.

For questions and general discussion, use GitHub Discussions. For bug reports and feature requests, use the issue templates.

Security

redline processes third-party page content through LLMs. That combination has inherent risk — prompt injection, data egress when using cloud providers, etc. See SECURITY.md for the full threat model, mitigations, and how to report vulnerabilities privately.

License

MIT — Copyright © 2026 Randall Degges. Built with Cobra, modernc.org/sqlite, PuerkitoBio/goquery, and a healthy respect for the Ollama project.