# Agent Ammo > Agent Ammo hosts expert-written agent skills that your AI agents use over MCP. Each skill learns from your edits, ratings and real results, separately for every account you connect, and every change is tested against your own history before it goes live. Agent Ammo is a subscription platform for agent skills. A specialist writes a base skill. Each subscriber connects one or more accounts (for example an inbox or a brand), and every account gets its own copy that learns from that account's edits, ratings and measured results. Agents use it over MCP (Streamable HTTP) with one bearer token per account. Key facts: - Works with any MCP client over HTTP: Claude Code, Claude Desktop, Cursor, Hermes Agent, OpenClaw, n8n, custom agents. - Launch skill: "SEO articles that book calls". Writes search-optimised articles and learns, per account, what actually gets readers to book a call on Cal.com. - Learning is per account and gated: new lessons run as trials, and a new version goes live only if it does no worse on the account's own past edits. One-click rollback, JSON export. - Pricing: Start free with up to three accounts. After the trial, $29 per account per month. ## Connect - [MCP endpoint](https://www.agent-ammo.com/api/mcp): Streamable HTTP, JSON responses. Header: `Authorization: Bearer `. - [MCP server card](https://www.agent-ammo.com/.well-known/mcp.json): machine-readable server description and tool list. - [Create an account and a token](https://www.agent-ammo.com/signup): a human must sign up and connect an account to receive a token. ## Docs - [Full reference for agents](https://www.agent-ammo.com/llms-full.txt): tools, call order, feedback rules, safety properties, FAQ. - [Home page as markdown](https://www.agent-ammo.com/index.md) - [The SEO skill](https://www.agent-ammo.com/skills/seo-booked-leads): what it does, what each account can tune, approved shared lessons. ## Optional - [Local SKILL.md stub](https://www.agent-ammo.com/api/stub/seo-booked-leads): a thin skill file for agents that load skills from disk; the real skill is served over MCP. --- # Reference for agents ## Call order for one piece of work 1. `get_skill` with the task. Keep the returned `run_id`. Follow the returned skill for this task. 2. Write the piece. Keep the booking links exactly as given: they are signed so a booking is credited to this run. 3. `log_run` with the full text you wrote. Fix any failed hard check before publishing. 4. When the user edits, rates or comments, call `submit_feedback` with the same `run_id`. Send the final text they shipped in `shipped`. Never invent feedback. 5. Optional: `get_scoreboard` to report progress to the user. ## Tools ### get_skill Returns the skill compiled for THIS account (base rules + what it has learned from this account's edits and results) for the given task, plus a run_id. Call it before writing. Follow the returned skill exactly, then call log_run. Arguments (required: task): - `task` (string): What to write, e.g. 'Article for keyword: fractional cfo for startups' ### log_run Send the full output you wrote for a run_id. Returns automatic hard checks (length, headings, meta description, booking link, banned claims). Arguments (required: run_id, output): - `run_id` (string) - `output` (string): The complete text you wrote - `artifact_url` (string): Published URL, if already live ### submit_feedback Send what the user actually shipped (their edited version), a 1-5 rating, and/or a comment for a run_id. This is how the skill learns this account's preferences. Never invent feedback the user did not give. Arguments (required: run_id): - `run_id` (string) - `shipped` (string): The final text the user published or approved - `rating` (integer) - `comment` (string) - `artifact_url` (string) ### get_scoreboard Edit-rate trend, lessons learned, bookings attributed and pending updates for this account's skill. No arguments. ## Server - Name: agent-ammo, version 0.1.0 - Protocol versions: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05 - Transport: Streamable HTTP at https://www.agent-ammo.com/api/mcp (stateless, JSON responses, batches up to 10) - Auth: bearer token per connected account (prefix `esk_`). Clients that cannot send headers may use https://www.agent-ammo.com/api/mcp/. - Rate limits: 120 requests per minute per account. ## What the SEO skill tunes per account - length: short (900-1,200 words: one sharp answer to one query); medium (1,400-1,900 words: the answer plus the two follow-up questions buyers ask); long (2,200-3,000 words: the definitive guide for the query) - intro_style: problem_first (Open by naming the reader's exact problem in their words, then promise the fix); story (Open with a 3-4 sentence scenario the ICP will recognise); stat_led (Open with one sourced number that reframes the problem) - cta_style: inline_twice (Two in-line booking links: after the first insight and before the conclusion); end_box (One clearly separated booking box at the end); both (One in-line link mid-article plus a booking box at the end) - depth: practical (Steps, checklists and examples the reader can apply today); strategic (Frameworks, trade-offs and decision criteria for a buyer choosing an approach) ## How learning works - Each edit is triaged by a small decision model (durable preference, one-off fix, factual correction, or noise). - Durable preferences become short itemised lessons. Lessons are screened for prompt injection; links, emails and phone numbers are removed unless the account owns the domain. - New lessons are on trial: shown in about half of runs, promoted when runs with them need fewer edits than runs without, or when the user confirms them. - A new version is replayed against the account's own past edits and goes live only if it is never worse. Accounts choose automatic updates or approval. - Bookings from Cal.com are attributed through signed links and tune which settings each account favours. - Authors' base skills improve weekly from aggregate evidence, one section at a time, with author approval. Accounts move to a new base only if it beats their own copy. ## Questions ### What is an agent skill? A set of instructions an AI agent follows to do one job well, such as writing an SEO article. Most skills are static files that never change. On Agent Ammo a specialist writes the first version and your copy keeps improving from your own use. ### Which agents does it work with? Any agent that speaks MCP over HTTP: Claude Code, Claude Desktop, Cursor, Hermes Agent, OpenClaw, n8n and custom agents. You paste one command or one JSON block. ### How does it learn? Your agent reports what it wrote and what you finally published. The difference becomes small, itemised lessons such as "avoid the phrase 'in today's fast-paced world'". New lessons run as trials in about half of your runs and are kept only if they help. ### Can it get worse over time? Every update is tested against your own past edits before it goes live, and is held back if it would have done worse. You can approve each update yourself and roll back to any earlier version with one click. ### Does my data help other customers? Your skill learns only from your accounts. By default, an abstract lesson that at least three different customers learned independently can be offered to similar accounts as a trial, after the skill's author approves it. Quotes, facts and links never leave your account, and you can mark any account private when you connect it. ### Can I take my skill with me? Yes. Each account's skill, with its lessons and history, exports as JSON at any time.