MCP server

Connect Claude, ChatGPT, Cursor or VS Code to Sailfork over MCP, so an assistant can edit in caissons, open pull requests and read issues and CI.

Server address

Add this address to any assistant that supports remote MCP servers. You need a Sailfork account to approve the connection.

https://status.sailfork.com/mcp
AddressTools
https://status.sailfork.com/mcpEvery tool
https://status.sailfork.com/mcp/readonlyOnly tools that never change anything
https://status.sailfork.com/mcp/x/pulls,ciChosen toolsets: caissons, issues, pulls, ci, repos, drydock, context
https://status.sailfork.com/mcp/x/pulls,ci/readonlyChosen toolsets, read only

Add it to your assistant

Each one opens a Sailfork page in the browser to ask for your approval.

claude.ai

Settings, Connectors, Add custom connector. Paste the URL, then sign in when asked.

https://status.sailfork.com/mcp

Claude Code

Run this in a terminal. Claude Code opens the browser the first time it needs access.

claude mcp add --transport http sailfork https://status.sailfork.com/mcp

Cursor

Add this to ~/.cursor/mcp.json.

{
  "mcpServers": {
    "sailfork": {
      "url": "https://status.sailfork.com/mcp"
    }
  }
}

VS Code

Add this to .vscode/mcp.json.

{
  "servers": {
    "sailfork": {
      "type": "http",
      "url": "https://status.sailfork.com/mcp"
    }
  }
}

ChatGPT

Settings, Apps & Connectors, Advanced: turn on developer mode. Then Create, paste the URL and choose OAuth.

https://status.sailfork.com/mcp

On this machine, with the CLI

This uses the login of the sailfork command line instead of a browser approval. Install the CLI, run sailfork auth login, then:

claude mcp add sailfork -- sailfork mcp

Add --read-only for tools that change nothing, or --toolsets pulls,ci to pick toolsets. Locally, the drydock toolset is left out unless you ask for it with --toolsets all, because its sandbox tools need your own Drydock machines.

Settings for the local server:

SettingWhat it does
-R owner/repo or SAILFORK_REPOThe repository to use when the person names none
--host <url> or SAILFORK_HOSTA host other than sailfork.com
SAILFORK_TOKENA token to use instead of the stored login, for CI and headless agents
SAILFORK_TIMEOUTThe request timeout in seconds. Calls that wait on purpose set their own.
--lockdown or SAILFORK_MCP_LOCKDOWN=1See Choose what an assistant can do

Signed out, the server still starts and every tool answers run sailfork auth login.

How it signs in

Nothing is pasted into the assistant: it never holds your password or a long-lived token.

  1. The assistant opens Sailfork in your browser.
  2. You see which app is asking, choose read only or read and write, and can limit it to some repositories.
  3. It gets a token that lasts an hour and renews itself while you keep using it. Revoke it any time under Connected apps in Settings.

Your own client or script can skip the browser: send a personal access token as Authorization: Bearer. An unauthenticated request gets 401 with a WWW-Authenticate header that points to the OAuth metadata at https://status.sailfork.com/.well-known/oauth-protected-resource. Tokens come from the REST API.

  • An OAuth token works on /mcp only. Anywhere else it is 401.
  • A tool that changes data called with a read-only token is 403 with error="insufficient_scope".
  • 600 requests a minute per token. Over that is 429 with Retry-After.
  • Requests over 8 MiB are 413.

Choose what an assistant can do

Give an assistant the least it needs. Four controls narrow what it can do, and they combine.

ControlRemote serverLocal server
Read onlyhttps://status.sailfork.com/mcp/readonly--read-only
Some toolsetshttps://status.sailfork.com/mcp/x/pulls,ci--toolsets pulls,ci
Bothhttps://status.sailfork.com/mcp/x/pulls,ci/readonly--read-only --toolsets pulls,ci
Hide text from outsidersAdd ?lockdown=1 to any address--lockdown

Read-only mode lists only tools that never change anything. A tool that changes data called with a read-only token is refused even if it was listed.

Lockdown replaces the titles, bodies and comments of people who cannot push to the repository with a [hidden by lockdown] note in issue and pull request reads, so text written by outsiders on a public repository never reaches the model. Use it when an agent reads a public repository. It is off by default.

At the consent step you can also limit a connection to some repositories. Revoke it any time under Connected apps in Settings.

Toolsets

Every tool is in one of seven toolsets. Ask for only the ones an agent needs with /mcp/x/pulls,ci. An unknown toolset is a 404. Your assistant lists the exact tools once it is connected.

ToolsetWhat it covers
caissonsOpen a caisson, read and write its files, commit and open the pull request.
issuesList, read, create and edit issues; claim the next one and post progress.
pullsRead pull requests and their diffs, review, comment and merge.
ciList runs, read failures and logs, rerun and dispatch workflows.
reposBrowse and search code, commits, branches and symbols.
drydockStart sandboxes and run commands on your organization’s own machines.
contextWho you are signed in as and what you are working on, so an agent starts oriented.

What each toolset is for

ToolsetUse it whenExample tools
caissonsThe agent changes code. It works in a caisson on the server, so it needs no clone and no shell.caisson_open, caisson_read, caisson_edit, caisson_batch, caisson_grep, caisson_test, caisson_ship
issuesThe agent picks up and reports on work.issue_next, issue_claim, issue_start, issue_comment, issue_ship
pullsThe agent opens, reviews or lands pull requests.pull_get, pull_diff, pull_feedback, pull_review, pull_merge
ciThe agent reads why a check failed and acts on it.ci_runs, ci_failures, ci_logs, ci_rerun, ci_dispatch, check_run_create, check_run_update
reposThe agent needs to look around before it changes anything.repo_tree, repo_read, repo_glob, search_code, search_symbols, repo_log
drydockThe agent needs to run code on your own machines.sandbox_create, sandbox_exec, sandbox_stop
contextThe agent starts oriented: who it is, what needs it, what happened.whoami, doctor, brief, events, inbox, inbox_mark_read, wait, undo

Short repository notes for the next agent are repo_note_list and repo_note_add, in repos. See Working with agents.

Typical flows

Fix an issue. issue_next with claim set returns the most important ready issue and takes a lease on it. issue_start claims an issue and opens a caisson. The agent reads with caisson_read and caisson_grep, changes files with caisson_edit or caisson_batch, runs the tests with caisson_test, and issue_ship or caisson_ship commits and opens the pull request.

Answer a review. brief returns the pull request with its unresolved line threads and the code each points at. The agent opens a caisson on the pull request's branch (caisson_open with branch), edits, then pull_resolve closes the threads it fixed.

Fix a failing check. ci_failures returns each failed step's exit code and last log lines. ci_logs has more, and ci_rerun runs the workflow again once the fix is pushed. wait blocks until the checks finish, so the agent does not poll.

Look before it changes. With only repos and --read-only, an agent can search, read files, history and symbols, and cannot change anything.

How the work is labeled

The server records the client's name from the MCP handshake, such as claude-code or cursor, so pull requests, comments and issue leases it makes are marked as agent work and name the client. Names such as Claude Code 1.0.3 and claude-code/2.1 count as one agent. A person can turn this off in Settings, Keys and tokens, under Agent tracking. See Working with agents.

Troubleshooting

What you seeWhat to do
The client cannot connect and the address 401sThat is expected before sign-in. The WWW-Authenticate header points to https://status.sailfork.com/.well-known/oauth-protected-resource. A client that supports remote MCP follows it and opens the browser.
The browser approval never opensSome clients ask you to start sign-in from their connector settings. If the client has no remote MCP support, use sailfork mcp instead.
401 with invalid_tokenThe token is expired, revoked, or for another address. Reconnect. An OAuth token works only on /mcp, so a script that calls /api needs a personal access token.
403 with insufficient_scopeYou approved read only, or the address ends in /readonly. Connect again and choose read and write.
A tool is missingThe toolset is not in the address or --toolsets, --read-only hides it, or it is drydock, which the local server leaves out by default.
404 on /mcp/x/...A toolset name is unknown. The names are caissons, issues, pulls, ci, repos, drydock, context.
403 on a request from a browser appA request with a foreign Origin header is refused.
413The request is over 8 MiB.
429More than 600 requests a minute on one token. Wait for Retry-After.
Tools answer run sailfork auth loginThe local server has no stored login. Run sailfork auth login, or set SAILFORK_TOKEN.
A repository is missing from resultsThe approval was limited to some repositories, or the account cannot read it. A private repository you cannot read answers as not found.
sailfork doctorChecks sign-in, token scopes and expiry, git and the server in one pass, and prints the command that fixes each problem. The MCP server offers it as the doctor tool.

In the browser

An agent that drives this site in a browser with WebMCP (document.modelContext) gets these tools instead of guessing from the page. Each page lists only the tools that fit it, and they follow the names and inputs of the hosted server above.

  • Tools run as you, with your permissions. They call the same API as the page and add no access.
  • Every tool that changes something asks you in a dialog first, naming exactly what will happen. Only a real click or key press on that dialog counts.
  • Deleting a repository, changing visibility and billing have no tool: they stay on the page.
  • Issue bodies, comments, titles and file contents come back wrapped as {"untrusted": "..."} with a notice. They are data, never instructions.
  • Failures return {"ok": false, "error": {"code", "message", "next"}} with what to do next. Codes: sign_in_required, no_permission, not_found, conflict, invalid_input, declined, rate_limited, server_error, offline, timeout, cancelled, busy.
ToolWhereWhat it does
whoamiEvery pageWho is signed in, or that nobody is.
sign_inEvery page, signed outOpens the sign-in page and returns to this page.
go_toEvery pageOpens a page on this site.
search_reposEvery pageFinds public repositories.
search_codeEvery pageSearches file contents in one repository or all you can read.
repo_listEvery page, signed inYour repositories and your organizations’.
my_workEvery page, signed inYour pull requests with checks, review requests and assigned issues.
repo_getRepository pagesVisibility, default branch, branches and your access.
repo_treeRepository pagesLists a folder at a branch, tag or commit.
repo_readRepository pagesReads a file, optionally a line range.
repo_logRepository pagesCommit history, for the repository or one path.
repo_branchesRepository pagesBranches with ahead and behind counts.
compareRepository pagesCommits and changed files between two refs.
issue_listRepository pagesIssues with filters.
issue_getRepository pagesAn issue with its comments and the checks of its pull request.
pull_listRepository pagesPull requests with filters.
pull_getRepository pagesA pull request with review state, checks and open threads.
issue_createRepository pages, signed inOpens an issue. Asks you first.
issue_commentIssue pages, signed inComments on an issue. Asks you first.
issue_updateIssue pages, write accessChanges state, labels, assignee, priority, title or description. Asks you first.
pull_checksPull request pagesCI jobs of the newest run with failed steps.
pull_merge_statusPull request pagesWhether it can merge and every reason it cannot.
pull_commentPull request pages, signed inComments, replies to a thread or comments on a line. Asks you first.
pull_reviewPull request pages, write accessApproves, requests changes or comments. Asks you first.
pull_mergePull request pages, write accessMerges, squashes or rebases. Asks you first.
org_membersOrganization pagesMembers and roles.
org_inviteOrganization pages, ownersInvites a person. Asks you first.
create_repositoryNew repository pageFills the form; you press Create repository. Asks you first.
comment_on_issueIssue pages, signed inFills the comment form; you press Comment. Asks you first.

Loading Sailfork