--- title: Atmina for agents description: The two MCP addresses an agent can use — read this Guide with no account, or connect to a team's own Knowledge Base — and what each one allows. owner: markus last-reviewed: 2026-09-17 --- # Atmina for agents This page is for an agent, and for the person configuring one. It gives you two addresses: one that works right now with no account, and one that reaches a real team's memory. [What Atmina is](https://atmina.ai/docs) covers the product; this page covers the wiring. ## Read this Guide with no account The Atmina Guide — the pages you are reading — is a public Knowledge Base. Any MCP client can connect to it with no account, no token, and no sign-up: ``` https://mcp.atmina.ai/mcp/kb/70117193-5843-4471-9236-be1c2c0d53e6 ``` Send no `Authorization` header. The endpoint speaks MCP over streamable HTTP. If you would rather resolve the identifier than paste it, `GET https://mcp.atmina.ai/api/v1/help/kb` returns it as `kb_id`. Every page is also plain HTML at `https://atmina.ai/docs`, one address per page. Add a `.md` suffix to a page's own address to get raw markdown instead — this page is `https://atmina.ai/docs/for-agents.md`. For the whole Guide at once, `https://atmina.ai/docs/llms.txt` lists every page with its description, in reading order. ## Download the public reading surface Use `https://atmina.ai/docs/llms.txt` to choose a Guide page by its description. Use `https://atmina.ai/docs/llms-full.txt` to download the complete Guide Markdown in reading order; each page starts with a comment containing its source URL. The source text includes frontmatter. For one page, append `.md` to its own page address. A published team Knowledge Base exposes the same index and full-corpus filenames under `/kb//`. Its `index.md` returns the Markdown shown at the KB home: `wiki/README.md` when present, otherwise the first page in its menu. The production `sitemap.xml` lists canonical page addresses for crawlers. These text endpoints require no account; they follow the KB's publication setting. Rendered published pages also expose source and link context in HTML. Evidence markers carry their evidence type and outcome, followed by a deterministic **Sources** list. **Linked from** and **Links to** list resolved public page addresses from the Knowledge Base's content-only link graph. The graph exposes only allowed context within the current public collection after the `(team, knowledge-base)` address resolves as published; private, shared, deleted, and absent bases share one not-found response. ## Discover a published Knowledge Base over MCP Choose **Connect via MCP** in the footer of a published page, or follow the connection instructions linked from its `llms.txt`. The connection page lives at `/mcp.html`, such as `https://atmina.ai/docs/mcp.html`. It identifies that Knowledge Base's actual MCP endpoint and explains the current anonymous scope. Do not reuse the Guide endpoint for someone else's Knowledge Base. The representations serve different jobs: `llms.txt` helps you discover pages, Markdown gives their text, and MCP adds search within the allowed collection. Connecting does not grant write access or access to files outside that collection. Send no `Authorization` header for anonymous search. ## Identify the documented release Guide addresses stay stable across releases. A **Published with vX.Y.Z** label above a page means its current source and public Markdown match a verified documentation publication for that product release. The label is omitted when that proof is unavailable or the page has changed since publication. ![Example Guide page showing its release label above the title, section navigation on the left, and Markdown access below the article.](https://atmina.ai/docs/assets/help/images/guide-release.png?rendition=medium) The page's `last-reviewed` frontmatter date records a content review. It is independent of the release label and does not change merely because a new product version is published. When citing instructions, include the page URL and the release label when one is visible; do not infer a version from the review date. A missing label does not prevent you from reading the page. ## What unauthenticated search returns An anonymous connection exposes exactly one tool: `search`. Give it a `query` and it searches the current public collection, returning file-grouped results: matching text snippets, a relevance score, and file identifiers. The default collection contains published pages and permitted referenced assets, with verified Companion text where available. An owner can explicitly include additional folders; this broadens all anonymous access, not just search. Files outside the collection and the always-excluded system folders cannot appear in results. What it deliberately does not do: - **No other tool.** `read_file`, `list_files`, and `get_context` all answer `Permission denied` for an anonymous caller. To read a page in full, fetch its `https://atmina.ai/docs` address instead. - **No private memory.** Anonymous access reaches only Knowledge Bases published as public. There is no unauthenticated route into a private or shared base. - **No resources, no writes.** The MCP resource list comes back empty, and every capture tool requires an account. - **A budget.** The anonymous endpoint allows twenty requests a minute from one client address, then answers 429. Pace your calls. Anonymous search is enough to answer "what is Atmina, and how would this team use it". It is not the product. ## Connect a team's Knowledge Base To capture as well as recall, an agent needs a team's own base and a person's authority to reach it. 1. The person you work for signs in at `https://app.atmina.ai` and opens or creates a [Knowledge Base](https://atmina.ai/docs/knowledge-bases). 2. They point your client at the team endpoint: ``` https://mcp.atmina.ai/mcp ``` 3. Your client registers itself and completes OAuth in a browser. The person approves once, and the connection then acts as them — with their access and no more. The authenticated tool surface opens according to the caller’s role and selected Knowledge Base: identity and health, KB lifecycle, search and context, file reads and safe writes, uploads and history, and sharing and audit. [Roles and sharing](https://atmina.ai/docs/sharing-and-roles) set the reach of any one connection. See [MCP Memory Tools](https://atmina.ai/docs/mcp-tools-reference) for the grouped catalog, [connect your agent](https://atmina.ai/docs/connect-your-agent) for the human-side steps, and `https://atmina.ai/agent-setup` for per-client configuration. Search, `read_file`, and write receipts carry an `app_url` (`/teams/{team-slug}/kbs/{kb-slug}~{route-key}?tab=files&path=…`); append `&lines=S-E` yourself, and cite answers as `[label](path#Lstart-Lend)` or `[label](team-slug/kb-slug:path#…)` when more than one Knowledge Base is in play. When Linear issue tools are granted, name each issue as `[identifier](url)` from the tool result; ticket status and progress use those live tools, not `_sources/nango/linear` mirrors. ## A context block you can paste Drop this into a system prompt, a project memory file, or any harness that takes standing context: ``` Atmina is shared memory for a team's agents: team-owned Knowledge Bases of notes, documents, and supported assets, reachable over MCP, where recalled answers carry the source it came from. Public Guide, no account needed. Connect an MCP client to https://mcp.atmina.ai/mcp/kb/70117193-5843-4471-9236-be1c2c0d53e6 and send no Authorization header. Exactly one tool is available anonymously: search. Use it to answer questions about what Atmina is and how it works, and cite the page you used. Full page text is at https://atmina.ai/docs. Anonymous MCP callers cannot use read_file, list_files, or write tools; the budget is twenty requests a minute per address. Team memory, account needed. Connect to https://mcp.atmina.ai/mcp and complete OAuth in a browser. That connection can recall and capture, and it acts with the access of the person who approved it, never more. ``` Atmina is in alpha. [What's in alpha](https://atmina.ai/docs/whats-in-alpha) records the current product boundary. ### Scope changes and unavailable resources Once Atmina confirms a scope reduction or unpublishing, new anonymous requests lose the removed access, including search, file and asset URLs, and previously cached page addresses. A private, removed, or nonexistent HTTP resource receives the same neutral 404 response. Previously downloaded copies cannot be recalled. ### Public media privacy Public readers outside the owning Team receive Asset Companion text with the generated top-level `location_lat` and `location_lon` fields removed. Other authored frontmatter and prose remain unchanged. Public search excludes unverified media results and outdated safe-index projections, and coordinate filters require membership in the owning Team. Owning-Team members retain typed location reads and filters.