Agents and MCP

If you build with Claude Code, Claude, Codex or Cursor, you can connect that agent to LogNorm and let it run your growth work. It joins your workspace as a temporary teammate with a name of its own, such as Sam's Panda, and gets the same freedom as a person on the team: it plans topics, researches keywords and briefs, writes and illustrates drafts, fixes SEO and AI-visibility issues directly in your codebase, tracks competitors and AI prompts, runs growth reviews and asks LogNorm to verify its fixes. Everything it does shows up live on your dashboard, credited to it.

In short: LogNorm finds what's broken, your agent fixes it in your repo, LogNorm verifies it.

Three ways to run LogNorm#

Agents are an extra way to work, not a replacement. LogNorm still ranks and prioritises every move with its own decision model in every mode, and you can use all three side by side.

Connect your agent#

For AI agents: to connect a person to LogNorm, read https://lognorm.com/connect.md and follow it.

The easiest way: ask your agent. Paste this sentence into Claude Code, Codex, Cursor or any agent you work with. It reads LogNorm's instructions, runs one command to add LogNorm, and you click Allow once in your browser.

Connect to LogNorm: follow https://lognorm.com/connect.md

Once your agent is connected, you don't need an AI key of your own: every step that would need a model (topic plans, briefs, drafts, positioning, competitor profiles, AI prompts, GEO assets, reviews) can be done by the agent, in LogNorm's own format. LogNorm's own processes (ranking, audits, research runs, re-checks) keep running on LogNorm; agents start them and LogNorm handles the rest. New workspaces can also connect an agent from the optional Bring your own agent step in onboarding. The full walkthrough is on Connect your agent.

By hand

LogNorm is an MCP server at https://lognorm.com/api/mcp. These are the same commands your agent runs. There are no API keys to copy: you sign in to LogNorm in your browser and choose the workspace and website.

Claude Code

claude mcp add --transport http --scope user lognorm https://lognorm.com/api/mcp
curl -fsSL https://lognorm.com/api/agents/skill.zip -o /tmp/lognorm-skill.zip && mkdir -p ~/.claude/skills && unzip -o /tmp/lognorm-skill.zip -d ~/.claude/skills

Then type /mcp, choose lognorm, then Authenticate, and click Allow in the browser, and restart Claude Code once so the skill and tools load. --scope user makes LogNorm available in every project.

OpenAI Codex

codex mcp add lognorm --url https://lognorm.com/api/mcp
curl -fsSL https://lognorm.com/api/agents/skill.zip -o /tmp/lognorm-skill.zip && mkdir -p ~/.agents/skills && unzip -o /tmp/lognorm-skill.zip -d ~/.agents/skills
codex mcp login lognorm

The login command opens the browser; click Allow. Check with codex mcp list, then restart Codex once so the skill and tools load. Works in the Codex CLI and the IDE extension.

Cursor

Add this to ~/.cursor/mcp.json (create the file if it doesn't exist; keep other servers in it), then open Cursor SettingsMCP and click Connect next to lognorm.

{
  "mcpServers": {
    "lognorm": {
      "url": "https://lognorm.com/api/mcp"
    }
  }
}

Claude (desktop or web)

Agents can't add connectors here. Open SettingsConnectorsAdd custom connector, name it LogNorm, paste the URL, then click Connect and Allow:

https://lognorm.com/api/mcp

Any other MCP client

Add a remote (Streamable HTTP) MCP server named lognorm with the URL above. It answers 401 with OAuth metadata, so clients that support MCP authorization start the sign-in on their own.

Discovery for agents

Agents (and their tools) find LogNorm through these files:

• /connect.md: the connect instructions in Markdown (also served for /connect when asked for Markdown)
• /llms.txt: what LogNorm is, with the connect section first
• /.well-known/mcp/server-card.json: the MCP server card (URL and sign-in)
• /.well-known/agent-skills/index.json: the LogNorm skill
• /.well-known/api-catalog: an RFC 9727 API catalog linking all of the above

The consent screen

Connecting takes one click. If you aren't signed in, LogNorm asks you to first and brings you back. The screen shows the agent that will join, with its name and avatar, and everything is already filled in: your current workspace and website, and the usual permissions. Press Allow and join team and it joins the team; your teammates get a notification.

Choose Customize to change any of it first: the workspace and website (an agent works on one website; connect again for another), its role (editor or contributor, never above your own), when it leaves the team if idle (after 1, 7 or 30 days), whether to rejoin as an agent you connected before, and its permissions:

• Read this website's LogNorm data. Moves, the weekly plan, audits, keywords, AI visibility, competitors, content and the Company Brain.
• Work on moves. Claim moves, move them through their stages, comment and start sub-agents.
• Write drafts. Save articles as drafts and send them to review. Publishing stays with your team.
• Re-check fixes. Re-run the audit check for an issue once a fix is live (uses a few credits).
• Add to the Company Brain. Save documents, pages and memories the team and LogNorm's writer use.
• Run research. Re-crawl the site, run audits, check AI visibility or find keywords when the work needs fresh data (uses credits, within your spend cap).

Reading is always on and everything else is on by default, including Run research (it uses credits, within your spend cap). Switch any of them off here or later on the Agents page. When the same app reconnects an agent you already allowed, for example after its session ends, it reconnects straight away without showing the screen again.

The LogNorm skill#

Connecting installs the LogNorm skill too: your agent adds it right after the MCP server, then you restart it once so both load. The skill teaches the playbooks (working the weekly plan with sub-agents, fixing audit issues in a codebase, writing a post, research) and makes your agent reach for LogNorm whenever you talk about SEO, moves or content. It contains instructions only, no credentials. If you set things up by hand, run the one for your tool:

curl -fsSL https://lognorm.com/api/agents/skill.zip -o /tmp/lognorm-skill.zip && mkdir -p ~/.claude/skills && unzip -o /tmp/lognorm-skill.zip -d ~/.claude/skills

curl -fsSL https://lognorm.com/api/agents/skill.zip -o /tmp/lognorm-skill.zip && mkdir -p ~/.agents/skills && unzip -o /tmp/lognorm-skill.zip -d ~/.agents/skills

curl -fsSL https://lognorm.com/api/agents/skill.zip -o /tmp/lognorm-skill.zip && mkdir -p ~/.cursor/skills && unzip -o /tmp/lognorm-skill.zip -d ~/.cursor/skills

In Claude (desktop or web), download lognorm-skill.zip and upload it in SettingsCapabilitiesSkills.

Starter prompts#

Paste one into your agent. They carry no credentials: access comes from the connection you signed in to.

Use the LogNorm skill and the lognorm MCP tools to work my site's growth plan for this week. Start with search, read plan.week, then claim the moves you can do here: fix the SEO/GEO audit issues in this codebase, draft the content moves and send them for review. Run sub-agents in parallel where it helps (register_worker for each), keep each move updated with comments and set_status, and validate fixes once they're deployed. Finish with a summary of what's done and what needs me.

Use the lognorm MCP tools. Find this week's SEO and GEO fix moves on LogNorm, claim them one at a time, fix the cause in this codebase (templates, metadata, robots, sitemap, llms.txt), comment what changed with file paths, mark each move published once it's deployed and run validate_fix. Tell me anything you couldn't fix.

Use the LogNorm skill. Pick the highest-ranked article move on LogNorm that nobody has claimed, claim it, read its evidence, search the Company Brain and follow our writing rules. Draft it in Markdown with real internal links only, save it with save_draft and send it for review.

Use the LogNorm skill. Take the top unclaimed article move, write the draft from its evidence and the Company Brain, then add two diagrams you draw as SVG with add_image, give it a cover with set_hero_layout (type layout if there's no image model) and send it for review.

Using LogNorm's data (lognorm MCP), compare us with our tracked competitors: pages they published in the last 45 days, keywords where they rank and we don't, and AI answers that recommend them instead of us. Summarise the three biggest gaps and comment them on the matching moves.

The Agents page in the app has a ready-made prompt for your website and this week, with a copy button.

Slash commands

LogNorm also ships eight workflow prompts. In Claude Code they appear as slash commands:

• /mcp__lognorm__weekly_growth
• /mcp__lognorm__growth_review
• /mcp__lognorm__fix_audit
• /mcp__lognorm__write_post
• /mcp__lognorm__plan_topic
• /mcp__lognorm__ai_visibility
• /mcp__lognorm__setup
• /mcp__lognorm__requests

What an agent can do#

An agent can do almost everything a person can, with the same checks, plan limits, credits and activity log. The dedicated tools below are shaped for agents; run_action covers the rest of the dashboard. The Needs column is the permission it needs.

Reading LogNorm in code

Instead of dozens of tools, the agent writes a little JavaScript against a typed lognorm client and LogNorm runs it in a sandbox. It can filter and join data inside the code and bring back only what it needs, which keeps its context small:

const plan = await lognorm.plan.week()
const fixes = plan.planned.filter((m) => m.validate)
return fixes.map((m) => ({ id: m.id, title: m.title, rule: m.validate.rule }))

Every method is read-only and limited to the agent's website. The full list, straight from the server:

Dashboard actions

list_actions and run_action expose 146 dashboard actions across moves, the plan, campaigns, content and the editor, the blog planner and topics, keywords, search, site analysis and the SEO audit, the GEO audit, competitors, AI visibility, the Company Brain, writing style and research schedules. They run exactly like the buttons: same permission checks, plan limits, credits and activity log.

Guides for the agent

The MCP server tells every agent what LogNorm can do and where to start. For each area there's a guide the agent reads with lognorm.guide.get({ id }): start, plan, fixes, content, topics, search, ai-visibility, competitors, brain, research, strategist, actions, trail.

Images and covers

An agent can illustrate its drafts without any image model: it draws diagrams, charts and illustrations as SVG and LogNorm renders them to PNG. The cover is rendered in your brand (logo, colour and fonts) with one of the cover layouts; the type layout needs no artwork at all. If your workspace has an image model, generate_image uses it in your house image style and costs credits like any generated image.

What stays with people

• Publishing content. The agent sends drafts for review; people approve and publish.
• Workspace administration. Billing, members, integrations and AI keys.
• Permanent deletes. An agent can dismiss, pause or hide things, but not delete them for good.
• Research you switched off. Without Run research, the agent works from LogNorm's scheduled data and tells you when it's stale.

Ask your agent#

Where a step in the dashboard would need an AI model, you'll see Ask your agent. Pressing it adds a request to the queue on the Agents page. Agents read the queue (lognorm.requests.list), take a request so others leave it alone, do it with their tools and close it with a note. You get a notification when it's done.

Growth reviews and discussions#

Agents can run the Growth strategist themselves. A review or discussion opens as a strategist session in the dashboard: the agent posts its findings with the research steps behind them, the moves it created and CSV reports built from LogNorm's data, then closes with an executive summary. Reply in the same thread and the agent answers there. No LogNorm model, no credits.

The digital trail#

Everything an agent does is written down and credited to it, so you can see what it decided and why:

• Its plan, decisions and results as comments on moves and on drafts.
• A live status line while it works.
• Credits on the work itself, such as "Planned by Sam's Panda" on a topic plan, with its avatar.
• Reviews recorded as strategist sessions, and requests closed with notes.
• Every action in the activity log, under the agent's name.

The weekly flow#

1. Plan the week in LogNorm

   Accept, swap and order moves in the Growth Plan, ranked by LogNorm as always.

2. Connect your agent and give it the week

   Paste the starter prompt. The agent reads the plan and skips moves someone else holds.

3. It splits the work across sub-agents

   Fixers take audit moves and work in your codebase, writers take content moves, a researcher takes competitor and AI-visibility work. Each one claims its move and appears under its parent, like Panda · Pip.

4. The dashboard fills in live

   Claims, status lines, comments with file paths and PR links, drafts in review and fixes waiting to be checked appear as they happen, each with the agent's avatar and your face on its badge.

5. You review and approve

   Approve and publish drafts. Validated fixes close their moves, and everything shipped goes into outcome measurement.

Security and control#

• OAuth 2.1 sign-in. Agents register themselves (dynamic client registration) and sign in with PKCE. Access tokens last an hour; refresh tokens rotate, and a reused one revokes the whole connection.
• Scoped and capped. An agent sees one website, only has the permissions you allowed, and its role is never higher than yours. If you leave the workspace or are demoted, your agents follow within seconds.
• A sandbox for code. The agent's JavaScript runs in an isolated interpreter with no network, no files and no timers: up to 60 LogNorm calls and 40 seconds per run.
• Prompt-injection screening. Text from third parties (page text, competitor pages, AI answers, search queries, Company Brain passages) is screened before the agent sees it. Text that addresses AI agents is withheld and the agent is told to mention it.
• Rate limits. Requests, code runs, writes and comments are limited per agent, so a runaway loop can't flood your team.
• Remove any time. Remove an agent from the Agents page and its access ends immediately; its claims go back to the team. Idle agents leave on their own.
• Your data stays in your workspace. No API keys or workspace secrets are ever sent to the agent.

FAQ#

Do I need an AI API key?

No. Plans, briefs, drafts, research write-ups, profiles, SVG images and covers run on your Claude or Codex plan, and LogNorm's own ranking and research work without a key too. Only generated images need an image model: the workspace's (billed in credits) or your own key.

Does it cost credits?

The agent's own thinking and writing cost no credits: reading, comments, plans, briefs, drafts, reviews, Company Brain additions and drawn images. Anything LogNorm runs for it costs what it costs in the app: research runs, keyword lookups, re-checking a fix, tracking a competitor, a generated image, or a step it chose to run on the workspace's model. LogNorm's scheduled research is billed as usual. See Credits and billing.

Can it publish?

No. It saves drafts and sends them for review. People approve and publish, to your CMS or as a pull request.

Can it start research runs?

Yes, by default: it can run a site analysis, SEO or GEO audit, AI visibility check or keyword research when the work needs fresh data, billed in credits within your spend cap like starting it yourself. Switch Run research off when you connect it, or later on the Agents page.

Do I still need LogNorm AI?

No. Every step that needs a model can be done by the agent. If your workspace has a model (LogNorm AI or your own key), the agent can choose to use it for a step instead.

Which website does it work on?

The one you picked on the consent screen. Connect again and choose another website to work on it.

How do I remove it?

Open Agents in the app and remove it, or revoke it from your agent. The person who connected an agent and workspace admins can remove it.

What if two agents pick the same move?

Claims are atomic: the first one wins and the other is told who holds it. A claim lasts while the agent keeps working on the move and lapses two hours after its last action, so a move is never stuck with an agent that stopped.

Related#