Give your AI agent full Jira access with just 4 tools.

Most Jira MCPs dump their entire API surface into the model's context. The agent wastes tokens picking between jira_get_issue, jira_fetch_issue, jira_issue_get... and still gets it wrong.

jira-mcp gives the model exactly what it needs:

Four tools that compose naturally: schema to discover, read to find, write to change, user search to resolve people. Less surface area means fewer wrong picks, fewer redundant calls, more context for your actual work.

Your credentials stay on your machine. jira-mcp runs as a local process over stdio — no server, no proxy, nothing between your agent and Atlassian.

Smart by default. jira-mcp hides Jira's complexity so your agent doesn't have to learn it:

• User search — jira_user_search resolves names and emails to account IDs in one call. No more guessing JQL syntax for assignee lookups.
• Required field validation — before creating an issue, jira-mcp checks the project's required fields and returns missing ones by name with allowed values. Your agent gets it right on the first try instead of decoding opaque customfield_10104 errors.
• Issue type validation — if the issue type doesn't exist in the target project, the error lists available types immediately.
• Issue linking by name — jira_write accepts links: [{type: "Blocks", from, to}] and resolves type names to the right link. Discover available types via jira_schema resource=link_types. No raw link IDs, no inward/outward confusion. parent_key sets an Epic (or other parent) in the same call.
• Wiki-markup safety — descriptions and comments expect Markdown and are converted to ADF on the v3 API. Wiki-markup tokens ({code}, {{inline}}, h1., [text|url]) are detected and rejected so a jira_read → jira_write round-trip cannot silently render as literal tokens.

Your agent already speaks Markdown. So does jira-mcp.

Write descriptions and comments in plain Markdown — headings, lists, code blocks, tables, links — and they show up as native Jira content. No ADF schemas to learn, no wiki-markup quirks, no escape sequences. Tables render as tables. Line breaks stay where you put them.

Need legacy wiki-markup for a specific case? Set description_format: "wiki" and the raw string goes through untouched.

mcp-atlassian is a full Atlassian suite — 72 tools, Confluence, OAuth, SSE transport. That's powerful, and it's the right pick if you need all of it.

jira-mcp does one thing: Jira. And it does it with as little friction as possible.

Why that matters for your agent:

• 4 tools, not 72 — less context burned, sharper focus, fewer hallucinated tool calls
• Zero runtime dependencies — single Go binary, no Python, no venv, no pip
• Works out of the box — API token auth, stdio transport, ship it

Use mcp-atlassian if you need Confluence, OAuth, or SSE. Use jira-mcp if you want Jira to just work.

acli is great for humans typing commands in a terminal. jira-mcp is built for AI agents — and that difference matters.

When you give an AI shell access, it can do anything: delete users, change org settings, trigger Rovodev. jira-mcp limits the blast radius to Jira. Structured tool calls instead of shell execution means no injection risk, no accidental admin actions, and a model that stays in its lane.

jira-mcp gives your AI agent:

• Safety by default — no shell injection risk, Jira-only blast radius
• Native Markdown — write comments and descriptions in Markdown, it converts automatically
• Built-in dry run — preview every write before it happens
• Lean context — 4 tools vs. the full CLI surface; your agent stays focused
• One-line read-only mode — just instruct the model to use jira_read only, no extra tokens needed

Use acli when a human is at the keyboard or when you need Admin/Rovodev operations. Use jira-mcp when an AI agent is driving.

Prerequisites: A JIRA Cloud account with an API token.

1. Go to API token management
2. Click "Create API token"
3. Give it a label, click "Create"
4. Copy the token immediately — you can't see it again

Pick one path:

Homebrew

brew tap mmatczuk/jira-mcp https://github.com/mmatczuk/jira-mcp
brew install jira-mcp

claude mcp add-json jira '{
  "command": "jira-mcp",
  "env": {
    "JIRA_URL": "https://yourcompany.atlassian.net",
    "JIRA_EMAIL": "you@company.com",
    "JIRA_API_TOKEN": "your-api-token"
  }
}'

Docker

No install needed. The -e VAR flags (without a value) forward each variable from env into the container:

claude mcp add-json jira '{
  "command": "docker",
  "args": [
    "run", "-i", "--rm",
    "-e", "JIRA_URL",
    "-e", "JIRA_EMAIL",
    "-e", "JIRA_API_TOKEN",
    "mmatczuk/jira-mcp"
  ],
  "env": {
    "JIRA_URL": "https://yourcompany.atlassian.net",
    "JIRA_EMAIL": "you@company.com",
    "JIRA_API_TOKEN": "your-api-token"
  }
}'

Binary

Download the binary for your platform from the releases page and put it on your PATH, then:

claude mcp add-json jira '{
  "command": "jira-mcp",
  "env": {
    "JIRA_URL": "https://yourcompany.atlassian.net",
    "JIRA_EMAIL": "you@company.com",
    "JIRA_API_TOKEN": "your-api-token"
  }
}'

First, confirm Claude Code picked up the server:

claude mcp list

You should see:

Checking MCP server health...

jira: jira-mcp - ✓ Connected

Then open a Claude Code session and ask: "List my Jira projects":

❯ List my Jira projects

⏺ jira - jira_read (MCP)(resource: "projects")
  ⎿  Found 3 project(s)

⏺ 3 projects:

  - ACME — Acme Corp
  - PLAT — Platform
  - OPS — Operations

If you see your projects, you're set. If not, check the server logs for errors and verify your credentials.

Use the same binary and env vars. The server speaks standard MCP over stdio.

MIT