Reliquary Alpha
Docs menu

Docs / Concepts

Links

A link is a vault's credential to another remote MCP server, like Stripe, Linear or Supabase, so a team can share a platform account with their agents without sharing the key.

A vault's owner adds, edits and deletes links from its Links page in the web app: the page lists the vault's links, Add link at the top opens the form, and each row's menu has Edit, Grants, Replace credential, Refresh tools and Delete. An agent can list them (name and url, never the credential) with list_links. Adding a link also discovers its tools, from the upstream server's own tool list; an owner grants them per role from the link's Grants page; and a granted tool is callable, as its own MCP tool. Step by step: Add a link.

  • A name and a url. The url is https://, a host and an optional path, with no user name or password (user:password@), query string (?) or fragment (#): every member and agent can read a link's url, and it is kept in the vault's activity log, so a key or token goes in the credential instead. A link saved before this rule keeps its url until it is next edited. Letters, digits and underscores for the name, unique in the vault. A save that is refused (a bad name or url, a name another link already has) shows its reason in the form, with a reference, and keeps what you typed, so you correct it and save again. A refused edit stays an edit.
  • A credential, encrypted the same way as an environment variable: nobody, including an agent, ever reads it back. It is attached to a call at the last moment, server-side, and re-checked for safety (see below) on every call, not just when the link was added. It is sent in the link's Header: Authorization: Bearer <credential> by default, or, for a service that reads its key from another header such as X-Api-Key, as that header's whole value. The header's name isn't secret: members read it beside the url, and changing it is logged. Editing a link changes its name, url and header only; the credential isn't shown or changed there. To replace it, an owner chooses Replace credential in the link's menu: the new one is stored in place of the old, the link keeps its name, url and grants, its tools are discovered again with it, and the replacement is logged in Activity without the credential. Changing the url keeps the credential, which is then sent to the new address, and the tools already discovered.
  • Discovered tools. Adding a link calls the upstream server's own tool list and stores each tool's name, description (as one line of at most 300 characters), arguments and whether it writes (from the tool's own readOnlyHint; anything but an explicit yes is treated as a write tool). A slow or unreachable server doesn't stop the link from being added: it's flashed as a warning, and the link's tools stay empty until an owner chooses Refresh tools in its menu. Refresh tools also picks up tools the server added since: it runs discovery again with the stored credential, gives a new tool the starting grants below, keeps the grants an owner set, and removes a tool the server no longer lists. It says which tools were added and removed; if it fails, the tools stay as they were and the warning says why. If the server answers with an error of its own, the warning gives only the error's code, never the server's words: they can be written to read as instructions, and servers often repeat the credential they were sent. What it said is kept in the web app's server log under the warning's reference, with the credential taken out.
  • Grants per role. An owner decides which of a link's tools each role may use, from the link's Grants page. Read tools default on for editors and owners; write tools stay off until an owner turns them on there.
  • A call log, append-only: who, which agent, which tool, how it went and the size of the answer, kept as a record even after the link itself is deleted. Never the arguments or the result, only their hashes. See A link's calls.

Calling a tool

A granted tool shows up in tools/list as <link>.<tool> (a link named stripe with a granted create_invoice tool becomes stripe.create_invoice; see MCP tools. Which ones appear depends on the connection: a role's grant, and for a write tool, a write-capable connection too (a read-only token never reaches one, whatever its role is granted). Reliquary makes the call server-side and attaches the credential itself; it's never sent to, or held by, the agent. The result comes back fenced as data from the upstream server, the same as any other text an agent reads through Reliquary, never as instructions.

The upstream server's description of a tool is its own text, not Reliquary's, so tools/list shows it last, in quotes, after a sentence saying it is data and not instructions, as one line of at most 300 characters. A tool whose name isn't made of letters, digits, underscores, dots and hyphens (up to 128) isn't listed.

If two vaults you belong to each have a link with the same name, and both offer a tool with the same name, tools/list offers only one of them, the same one every time, and leaves the other out. Rename one of the links to reach both.

When a call fails, the agent is told what was being done, where it broke and why, with a reference. It has two, the MCP server's and the web app's, since the detail is in the web app's log. If the linked server refuses with an error of its own, the agent gets only the error's code, as the warning shown when you add a link does.

To see what a shared account is being used for, open the link's Calls page: on the Links page, pick Calls in the link's menu, or select its Last call (an editor, who has no menu, uses that). It lists every call through the link, newest first, 100 a page (Older for the rest):

  • When, and Who: the person, and the agent that acted for them ("you via Claude Code").
  • Tool: the tool called, as <link>.<tool>.
  • Outcome: OK, Error (the upstream server answered with an error, or couldn't be reached), or Refused (the link's grants or the caller's connection didn't allow it, so nothing reached the server).
  • Size: how much the server answered. A dash when there was no answer.

The arguments and the answer are never stored, so they can't be shown. The Links page also shows each link's Last call and how many calls it had in the Last 7 days.

Owners and editors see a link's calls; viewers don't, the database returns them none.

Who may do what

Adding, editing or deleting a link, replacing its credential, refreshing its tools and setting its grants is an owner's, in person, same as the ceiling: no agent, token or connection may do it, whatever surface offers the call. Calling a granted tool is different: any connection whose role is granted it may, agents included, the same as any other tool. Members read a vault's links and their grants within their role; an agent does the same over MCP with list_links, read-only connections included.

Not to be confused with a connection

A connection is a client reaching into Reliquary: Claude Code, a token, the CLI. A link is Reliquary reaching out, to an upstream MCP server on the vault's behalf. The two are opposite directions and unrelated, beyond both letting an agent do more with less it has to hold itself.

For agents: this page as Markdown, and every page at /llms.txt.