# MCP tools for agents

> GLLmolfar has a Model Context Protocol server inside it. An agent such as Claude Code, Codex or Hermes connects to it on this computer, starts crawls, reads the results and saves reports. Everything it does happens in the open window, where you can see it.

Describes GLLmolfar 0.9.22, checked 2026-11-01. Canonical page: https://www.getlooploop.com/gllmolfar/docs/mcp-tools

- **Where:** Settings › General, and Help › Connect an AI agent…
- **Reachable from:** This computer only, with a token
- **Transport:** Streamable HTTP, or a command for clients that start a server
- **Default:** Off until switched on

## Connect an agent

- In [Settings › General](https://www.getlooploop.com/gllmolfar/docs/settings-general), switch on **Let AI agents use GLLmolfar on this computer (MCP)** and press **Save**.
- Press **How to connect an agent…** beside it, or **Help › Connect an AI agent…**.
- Copy the block for your agent with **Copy** and run or paste it where the dialog says.

| Block in the dialog | Where it goes | What it holds |
| --- | --- | --- |
| Claude Code | Run once in a terminal: | A `claude mcp add --transport http gllmolfar …` command with the address and the token. |
| Codex | Add to ~/.codex/config.toml: | A `[mcp_servers.gllmolfar]` entry that starts GLLmolfar's own bridge as a command. |
| Hermes Agent | Add to ~/.hermes/config.yaml: | The same bridge, as `mcp_servers: gllmolfar:`. |
| Any other MCP client | Streamable HTTP, or the command above for clients that start a server themselves. | The URL and the header `Authorization: Bearer <token>`. |

The address is `http://127.0.0.1:<port>/mcp`. The port is normally `47863`; when another program already holds it, the app takes a free one, and the dialog shows the address in use. The bridge reads the port and the token from the data folder, so a Codex or Hermes setup needs no token pasted in.

> **The token.** **Only programs on this computer can reach it, and only with the token. Anyone with the token can drive the app, so keep it to yourself; a new token cuts off every agent set up with the old one.** **New token** at the foot of the dialog makes one. The server also refuses requests that come from a web page in a browser.

A lock means it is not open on this account; pressing it says what opens it.

## Crawls

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `get_status` | What GLLmolfar does now: the open crawl, its progress, the account level and its URL limit, and whether Search Console data is there (`searchConsole`, with `say`: what to do). Agents are told to call it first. | Reads |
| `list_crawls` | The saved crawls on this computer, newest first. Can filter by site. | Reads |
| `open_crawl` | Opens a saved crawl in the window; the other tools then read it. | Changes the window |
| `start_crawl` | Starts a crawl in the window and returns at once. | Fetches pages |
| `wait_for_crawl` | Waits until the crawl, the embeddings and the topics are done, or the timeout. Returns the status. | Reads |
| `pause_crawl` | Pauses the running crawl. The queue stays. | Changes the crawl |
| `resume_crawl` | Continues a paused crawl from its queue. | Changes the crawl |
| `stop_crawl` | Stops the running crawl. The pages crawled so far are kept. | Changes the crawl |
| `get_history` | The saved crawls of one site with pages, errors, non-indexable pages and health, for trends. | Reads |
| `compare_crawls` | Two crawls of one site: URLs found and lost and why, changes per URL, fixes that grew or went, robots.txt, AI bots, health and alerts. Without arguments: the open crawl and the one before. | Reads |

## Findings and fixes

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `get_overview` | The open crawl in short: site, pages, health (0–100, GLLmolfar's score, not a Google score), topics, the number of fixes, and `searchConsole`. | Reads |
| `get_fixes` | The [fix list](https://www.getlooploop.com/gllmolfar/docs/fix-list), most important first. Each fix: key, importance, type, URL count, why, how, and `effect` (index, rank, ctr, users or none). `notCounted` names service pages a content check left out. Translation-only items come last, apart from health. | Reads |
| `get_fix_urls` | The URLs of one fix key, also shown in the window. For broken links, redirects and similar keys, `linkedFrom` gives the pages that link to them, with anchor and position. The key `links` gives link opportunities. | Reads |
| `explain_fix` | One fix key for a person who does not know the app: what it is, why it matters, how to correct it, how to verify it, with example URLs. | Reads |
| `create_jira_task` | One fix as a full Jira ticket: fields, description in Jira markup and Markdown, a block for a coding agent, steps and Given/When/Then criteria, and attachments written as files. GLLmolfar sends nothing to Jira. | Writes files |
| `find_risky_content` | Pages with risky content: words, foreign script and meaning unusual for the site. No model is used. The crawl must be stopped first. | Reads |
| `get_ai_readiness` | The AI-readiness score (0–100) and its four parts. With `run_probe`, it fetches about ten pages as each AI bot and reports robots.txt per bot, llms.txt and Markdown for agents. | Fetches pages |

## Pages and meaning

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `get_pages` | The open crawl's pages as a table: chosen columns, filters (fix key, topic, URL pattern, status, indexable, language), paged with `next_offset`. The column `fixes` lists each page's fix keys. | Reads |
| `count_pages` | Page counts by one field, without rows: status, indexable, depth, topic, folder, template, language, page type, sitemap or fix. | Reads |
| `audit_url` | One live page without a crawl: status, indexability, time, headers, redirects, JavaScript changes, robots.txt per AI bot, accessibility, headings, JSON-LD and the start of the text. Full text or HTML on request. | Fetches pages |
| `semantic_search` | Pages and sections of the open crawl that answer a question, with scores. Needs embeddings. | Reads |
| `get_topics` | The [topics](https://www.getlooploop.com/gllmolfar/docs/topics): label, terms, pages, cohesion, link coverage, cannibalization pairs, pillar page and example URLs. Needs embeddings. | Reads |
| `get_cannibalization` | Groups of pages with almost the same meaning, with their pairs and similarity. The first URL of a group is the page to keep. Needs embeddings. | Reads |
| `measure_text` | Measures a draft with embeddings on this computer, without a language model: its passages, the best passage per query, the nearest page of the crawl, and passages that lean on an earlier one. | Reads |
| `get_search_console` | Search Console data for the open crawl's site: totals against the period before, top queries and pages, this week's movement, shared queries and pages that lost clicks. Search Console must first be connected in the app's [GSC](https://www.getlooploop.com/gllmolfar/docs/search-console) section. | Fetches data |

## Reports

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `list_reports` | The reports of the [Reports](https://www.getlooploop.com/gllmolfar/docs/reports) section, each with an id, its direction, `ready` and, when not ready, `why`. | Reads |
| `get_report` | Opens one report in the app and returns its verdict, figures and findings. Writes the same page as HTML or PDF to Documents\GLLmolfar\Reports, with a new name each time. Takes a `note` shown under the report's head, and the person's own projects to compare. | Writes files |
| `save_report` | Saves the classic summary report of the open crawl to Documents\GLLmolfar\Reports and returns its path. Never replaces an older file. | Writes files |

## Planned crawls

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `list_schedules` | The [planned crawls](https://www.getlooploop.com/gllmolfar/docs/schedule) with their next run, the queue, and the runs of the last 30 days. | Reads |
| `save_schedule` | Adds a planned crawl or changes one by id: start as local time and repeat. Several URLs make one schedule each. | Changes schedules |
| `run_schedule` | Runs one planned crawl now; if a crawl is running, it waits after it. | Fetches pages |
| `delete_schedule` | Deletes one planned crawl by id. Its queued runs are skipped; saved crawls stay. Agents are told to ask first. | Deletes |

## Reverse prompts

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `reverse_prompts_read` | Step 1 of 2. Reads a public page and splits its text into candidate blocks. Returns a run id, the brief, a draft analysis and the rules. | Fetches pages |
| `reverse_prompts_build` | Step 2 of 2. Takes the finished analysis, checks it and builds a prompt and a measured example for each block, or names the lines to correct. | Changes the run |
| `reverse_prompts_get` | One run by id: what it waits for, or the page profile and each block's prompt and example. | Reads |
| `reverse_prompts_list` | The runs on this computer, newest first, with URL, status and run id. | Reads |

## Tools that use your own model

These three use the model connected in [Settings › AI models](https://www.getlooploop.com/gllmolfar/docs/settings-ai-models). The server tells agents to call them only when the person asks.

| Tool | What it returns or does | Kind |
| --- | --- | --- |
| `plan_fixes` | An action plan from the open crawl's fixes. GLLmolfar checks each task against its fixes and page counts. The result shows on the [Agent · test](https://www.getlooploop.com/gllmolfar/docs/agent-test) screen. | Uses a model |
| `write_content` | Rewrites a page, or writes a blog post from it. Each draft is measured with embeddings, sent back for another round if it fails, and fact-checked against the page. Can take minutes. | Uses a model |
| `write_social` | Posts about one page for each network asked for, each with its length, the network's limit and any numbers not in the page. | Uses a model |

## Recipes

The server also offers named jobs as MCP prompts. Agents that support prompts show them as slash commands; **Help › Agent skills…** lists the same recipes in the app.

| Prompt | Title | When |
| --- | --- | --- |
| `technical-audit` | Full technical audit → action plan | First contact with a site, or a quarterly audit. |
| `ai-readiness` | AI-readiness audit of the whole site (GEO, AEO) | Can AI assistants read and cite the site? |
| `page-agent-readiness` | Agent-readiness check of one page | Before launching a key landing, signup or pricing page. |
| `citability` | Citability review and rewrite plan for one page | A page that should be cited for a set of questions but is not. |
| `question-coverage` | Question fan-out coverage test | Planning content for a topic: which questions does the site not answer? |
| `cannibalization` | Cannibalization cleanup | The cannibal check appears, or two pages split one query in Search Console. |
| `internal-links` | Internal linking plan | Orphans, deep pages, or a new page that needs links. |
| `llms-txt` | llms.txt draft | The site wants an llms.txt, or GLLmolfar flagged problems in the existing one. |
| `refresh-brief` | Content refresh brief for one page | A page losing clicks, stale, or beaten by a competitor. |
| `stale-content` | Stale, undated and unsourced content sweep | Quarterly E-E-A-T hygiene. |
| `jira-backlog` | Jira backlog from a crawl | Handing technical work to developers. |
| `release-check` | Release check: did this deploy break anything? | After a deploy, migration or redesign. |
| `monthly-report` | Monthly client progress report | Agency reporting. |
| `structured-data` | Structured data plan | Schema errors, or no schema on key templates. |
| `fix-and-verify` | Fix it in the codebase, then verify | The agent is working inside the website's repository. |
| `competitor-passages` | Competitor passage comparison | Why are they cited and we are not? |
| `gsc-opportunities` | Search Console × crawl: quick wins | Quick wins with real demand behind them (Search Console connected in GLLmolfar). |
| `reverse-prompts` | Writing prompts for a page type (reverse prompts) | Briefing writers or a model to produce more pages like the best one. |
| `checked-audit` | Audit with live checks: do, do not, check first | Before you give a fix list to a person or a developer. |
| `verify-fixes` | Verify the fixes after a deploy | After the developers say the fixes are live. |

## States and messages

- **On** — the app says: "On: http://127.0.0.1:47863/mcp" The server is listening at that address.
- **Starting** — the app says: "Starting…" Switched on; the server is not listening yet.
- **Off** — the app says: "Off. Switch it on in Settings › General." No agent can connect.
- **Not running** — the app says: "GLLmolfar is not running, or its agent connection (Settings › General) is off." From the bridge, to an agent that starts it as a command.
- **Token changed** — the app says: "GLLmolfar answered 401: the token changed; restart the agent." From the bridge, after **New token**.

## Related pages

- [Settings › General](https://www.getlooploop.com/gllmolfar/docs/settings-general.md): Settings › General: the interface language, term explanations, the interface mode, the default crawl size and mode, updates, statistics, the tray, MCP, sound.
- [Settings › AI models](https://www.getlooploop.com/gllmolfar/docs/settings-ai-models.md): Settings › AI models: connect your own language models, on this computer, on your network or in a cloud, and choose which model does which analysis.
- [Agent · test](https://www.getlooploop.com/gllmolfar/docs/agent-test.md): The Agent · test screen: your own models plan tasks from the fix list, rewrite pages, write social posts and review risky pages; GLLmolfar measures the results.
- [Reports](https://www.getlooploop.com/gllmolfar/docs/reports.md): The Reports tab: eleven reports in four directions, each answering one question about the open crawl, with a language filter and HTML or PDF export.
- [Fix list](https://www.getlooploop.com/gllmolfar/docs/fix-list.md): The Fix list tab: what to fix in the crawled site, numbered, most important first, each with why it matters, how to fix it and the URLs.
- [Schedule](https://www.getlooploop.com/gllmolfar/docs/schedule.md): The Schedule tab: a month calendar of planned and past crawls, the list of schedules, and the dialog that plans a crawl once or on repeat.