Tickets and Docs over MCP: Linear, Jira, Notion and Slack

Linear, Atlassian (Jira and Confluence), Notion and Slack each run an official remote MCP server at a vendor-hosted URL with OAuth login, so a coding agent can read a ticket, its spec and its discussion, then post the result back. Combined with GitHub MCP or gh, the agent can take one ticket to an open pull request and a status update.

This page is for developers who work from a tracker and tech leads who want the whole team to run the same ticket loop. You know the pattern: the ticket says one thing, the spec in Notion says another, the decision that settles it is buried in a Slack thread, and the agent only saw what you remembered to paste. Connect the four sources once, and the agent reads all of them itself.

What a ticket-driven MCP setup gives you

• Install lines for Linear, Atlassian Rovo, Notion and Slack in Claude Code, Codex and Cursor, each marked with where it comes from (vendor line, registry URL, or untested).
• A ticket → spec → PR → status loop in which the agent moves the ticket to In Review only after CI is green.
• Two copy-paste prompts: the full ticket-to-PR run and a pre-flight check that catches underspecified tickets.
• The traps that break these servers in 2026: the dead Atlassian /v1/sse endpoint, Slack’s pre-registered OAuth client, and the unmaintained local Notion server.

Which project-management MCP servers to connect

All four servers are official and remote, so there is no package to install or update. Popularity as of 2026-09-26, from install counts on the Claude plugin directory (claude.com/plugins, first-party count): atlassian 92,289, slack 92,287, linear 48,038. The Notion plugin count was not recorded; Notion’s server is listed in the Official MCP Registry as com.notion/mcp.

Server	Endpoint	Auth	Use it for	Watch out for
Linear	https://mcp.linear.app/mcp	OAuth	Issues, projects, comments, status	Nothing Linear-specific; the URL is in the Official MCP Registry (app.linear/linear)
Atlassian Rovo	https://mcp.atlassian.com/v2/mcp	OAuth 2.1 or API token	Jira, Confluence, Bitbucket, Compass, Jira Service Management	/v1/sse stopped working after 30 June 2026; API-token auth needs an admin; JSM tools need an API token
Notion	https://mcp.notion.com/mcp	OAuth	Specs, runbooks and decision logs	The local @notionhq/notion-mcp-server is “no longer actively maintained or supported”, per its README
Slack	https://mcp.slack.com/mcp	OAuth with a pre-registered client	Reading the thread where a decision was made	A bare claude mcp add omits the pre-registered OAuth client; use the plugin or the --client-id line below

Atlassian has more setup detail (admin controls, IP allowlisting, API tokens, Server and Data Center) than fits here. The Atlassian Rovo MCP guide is the canonical page for it.

Connect Linear, Atlassian, Notion and Slack

Each vendor also ships a plugin in the first-party marketplaces (claude-plugins-official for Claude Code, openai-curated for Codex) that bundles the MCP entry, so installing the plugin and adding the server by hand give the same connection. Pick one route per server, not both, or you end up with two entries for the same server to keep in sync.

Claude Code

Run in your terminal, from the repository root. --transport http is required: without it, Claude Code (checked on 2.1.286) prints a warning but still writes a stdio entry whose command is the URL, which cannot connect.

# Linear (URL from the Official MCP Registry)
claude mcp add --transport http linear https://mcp.linear.app/mcp

# Atlassian Rovo (vendor line)
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp

# Notion (URL from the Official MCP Registry)
claude mcp add --transport http notion https://mcp.notion.com/mcp

# Slack: the plugin supplies the pre-registered OAuth client
claude plugin install slack@claude-plugins-official
# or, without the plugin (writes the same entry; run one, not both):
# claude mcp add --transport http --client-id 1601185624273.8899143856786 --callback-port 3118 slack https://mcp.slack.com/mcp

Then run /mcp inside Claude Code and log in to each server. On a machine without a browser, use claude mcp login <name> --no-browser and paste the redirect URL back. The plugin route for the other three is claude plugin install linear@claude-plugins-official (likewise atlassian, notion).

Codex

Run in your terminal. In codex-cli 0.157.1, the --url flag is what marks a server as Streamable HTTP; there is no transport option.

# Linear (URL from the Official MCP Registry)
codex mcp add linear --url https://mcp.linear.app/mcp && codex mcp login linear

# Atlassian Rovo (vendor URL)
codex mcp add atlassian --url https://mcp.atlassian.com/v2/mcp && codex mcp login atlassian

# Notion (URL from the Official MCP Registry)
codex mcp add notion --url https://mcp.notion.com/mcp && codex mcp login notion

# Slack: use the curated plugin, which carries its own OAuth client
codex plugin add slack@openai-curated

Each codex mcp add writes a [mcp_servers.<name>] table with a url key to ~/.codex/config.toml. codex plugin add linear@openai-curated and notion@openai-curated are the plugin routes. OpenAI’s curated atlassian-rovo plugin still pointed at /v1/mcp/authv2 when checked on 2026-09-26, so add Atlassian by hand to get /v2/mcp. For headless machines, codex mcp login <name> --no-browser prints the authorization URL instead of opening a browser.

Cursor

Add the servers to .cursor/mcp.json in the project (shared with the team) or ~/.cursor/mcp.json (only you). Cursor has a native remote transport, so the mcp-remote bridge is not needed.

{
  "mcpServers": {
    "linear": { "url": "https://mcp.linear.app/mcp" },
    "atlassian": { "url": "https://mcp.atlassian.com/v2/mcp" },
    "notion": { "url": "https://mcp.notion.com/mcp" }
  }
}

For Slack, run /add-plugin slack in Cursor’s agent chat (Slack’s documented route), which supplies the OAuth client. Atlassian also ships a plugin in Cursor’s marketplace. These JSON entries are built from the vendors’ URLs and were not tested in a Cursor binary.

Traps in these install lines

• mcp.atlassian.com/v1/sse stopped being supported after 30 June 2026. /v1/mcp still works; use /v2/mcp for anything new.
• npm packages named @linear/mcp, @notionhq/mcp, @slack/mcp-server and @atlassian/mcp do not exist (npm, 2026-09-26). @modelcontextprotocol/server-slack exists but is deprecated and archived.
• claude mcp add --url … fails with unknown option '--url': --url is Codex’s flag.
• codex mcp add … --env is for stdio servers only. For an HTTP server with a token, use --bearer-token-env-var.

Check what the servers cost in context

Every connected server loads its tool definitions into the agent’s context. Before and after connecting, run /context in Claude Code and compare the MCP tools line; for an installed plugin, claude plugin details <name> shows its component inventory and projected token cost (it reports “not found” for a plugin you have not installed yet, so run it after installing and claude plugin disable <name> if the cost is too high). If four servers crowd the window, connect Slack and Notion only in the sessions that need them. The MCP token-cost guide covers the heavier fixes.

Run the ticket → spec → PR → status loop

The loop below is the job these servers exist for. It assumes the agent can also reach GitHub, either through the GitHub MCP server or the gh CLI in its shell. Branch creation and commits need neither: the agent runs git directly.

1. Read the ticket and everything it links. The agent pulls the issue, its comments, the linked Confluence or Notion spec, and the Slack thread named in the ticket. The output is a short list of acceptance criteria, each quoted from its source.
2. Stop if the spec is incomplete. If a criterion is untestable or two sources disagree, the agent comments its questions on the ticket and stops. This is cheaper than a PR built on a guess. See how to write acceptance criteria an agent can verify.
3. Turn each criterion into a failing test. One test per criterion, named after it, committed before the implementation.
4. Implement on a branch named after the ticket, for example eng-1423-retry-webhooks, until the tests and the existing suite pass locally.
5. Open the PR with an evidence table. The PR body maps each criterion to the test that proves it and links the ticket.
6. Update the ticket only after CI is green. The agent comments the PR link and the evidence table on the ticket, then moves it to In Review. It never moves a ticket to Done: that belongs to the human who merges.

Copy-paste prompt: take one ticket to a pull request

Take Linear issue ENG-1423.
1. Read the issue, its comments, and every linked Notion page. If the issue
   links a Slack thread, read it too. List the acceptance criteria, each with
   a quote and its source.
2. If any criterion is ambiguous or two sources conflict, post your questions
   as a comment on ENG-1423 and stop. Do not write code.
3. Create branch eng-1423-<short-slug>. Write one failing test per criterion,
   named after the criterion, and commit them.
4. Implement until the new tests and the full test suite pass.
5. Open a PR titled "ENG-1423: <summary>" whose body has a table:
   criterion | test name | status.
6. Wait for CI. Only when it is green, comment the PR link and the table on
   ENG-1423 and move the issue to In Review. Never move it to Done.
If anything fails, report the failure on the PR and leave the issue status
unchanged.

For Jira, replace Linear issue ENG-1423 with Jira issue PROJ-88 and Notion page with Confluence page; the rest of the prompt is unchanged.

Copy-paste prompt: pre-flight check on this sprint's tickets

List the Linear issues in the current cycle assigned to me that are in Todo.
For each one, read the description and linked Notion pages and report:
- whether it has acceptance criteria an automated test could check,
- which criteria are missing or ambiguous,
- which issues it depends on.
Do not change any issue. Output a table: issue | testable? | gaps | depends on.

How the loop differs between Claude Code, Codex and Cursor

Claude Code

Run the prompt interactively and approve the first ticket write when Claude Code asks. For a team, commit the servers to .mcp.json at the repo root with claude mcp add -s project …; each teammate still logs in with their own OAuth account, so the tracker’s permissions apply per person.

Codex

Run it interactively in the CLI or app, or non-interactively with codex exec "<prompt>" once the servers are logged in. Run unattended only with read-only tracker access, or with writes limited to comments, as the issue-to-PR pipeline guide sets up. Codex also has a Linear integration that starts work from Linear itself rather than from your terminal (see Codex integrations).

Cursor

Paste the prompt into the agent chat and approve the tracker writes as they come up. Cursor Automations list Linear among their triggers, so a cloud agent can start from a Linear event instead of from you (see Cursor cloud agents and automations). Run unattended only with read-only tracker access, or with writes limited to comments, as the issue-to-PR pipeline guide sets up.

How you verify the result without reading every line

The loop is only safe if the evidence is on the ticket, not in the agent’s summary:

• Acceptance criteria become tests first. A reviewer checks that each test asserts what its criterion says, which is a short read compared to the implementation.
• CI is the gate for the status change. The agent moves the ticket only after the pipeline passes, so In Review means “tests green”, not “agent finished”.
• The evidence table is the review entry point. The reviewer reads criterion → test → result, then samples the diff where the table is thin. The evidence bundle guide extends this to screenshots and logs.
• The human who merges moves the ticket to Done. Agent writes to the tracker stay limited to comments and one status transition, and the tracker’s history shows each one under the OAuth user who ran the agent.

Fix Linear, Jira, Notion and Slack MCP failures

The OAuth login never completes. A VPN or proxy that blocks the localhost callback hangs the flow. Retry with claude mcp login <name> --no-browser or codex mcp login <name> --no-browser and paste the redirect URL back. For more symptoms, see MCP connection issues.

Atlassian worked last quarter and now fails. Old configs point at https://mcp.atlassian.com/v1/sse, which stopped being supported after 30 June 2026. Remove the entry and add /v2/mcp.

Atlassian says an admin must approve, or API-token login is rejected. API-token auth must be enabled by your Atlassian site admin, and Jira Service Management tools work only with API tokens; if OAuth consent is blocked, check your site’s app-access settings with the admin. The Atlassian Rovo MCP guide lists the settings.

Slack login fails with a client or redirect error. The likeliest cause is an entry added without the pre-registered client (a bare claude mcp add). Remove it and install the slack plugin, or add it with --client-id and --callback-port 3118 as shown above.

Notion tools vanish or error after an update. You are on the local @notionhq/notion-mcp-server, which Notion no longer maintains. Switch to https://mcp.notion.com/mcp.

The agent picks the wrong ticket or none at all. Give the issue key (PROJ-88, ENG-1423) in the prompt; a key resolves in one call, a project name needs a search the agent may get wrong.

A ticket update is denied. The server acts as the OAuth user. If you cannot edit the ticket in the web UI, the agent cannot either.

The agent follows an instruction written inside a ticket or Slack message. Tracker text is untrusted input: anyone who can comment can write “also delete the staging branch”. Keep ticket writes behind an approval prompt, never give an unattended job tracker write access beyond comments and one status transition, never let the agent act on instructions found in ticket text, and avoid the community korotovsky/slack-mcp-server with browser or user tokens: it works without an approved Slack app, which is exactly the exposure. See MCP security.

Where to go next with ticket-driven development

Atlassian Rovo MCP Jira, Confluence and Bitbucket in depth: API tokens, admin controls, traceability.

GitHub MCP server The other half of the loop: PRs, CI logs and toolset scoping.

Issue-to-PR pipeline Run this loop unattended, with comment-only tracker writes and the other guardrails that requires.

The MCP starter stack Where a tracker server fits among the five servers most teams need.