jira-mini-mcp

jira-mini-mcp

A Jira Cloud MCP server for coding agents: 6 read tools, 9 with writes enabled.

One job. One tool. Done right.

General-purpose Atlassian MCP servers expose dozens to hundreds of tools. Every one costs context before the agent does any useful work, and every near-duplicate makes the agent's choice less certain. This server gives a coding agent the Jira context it needs for a ticket, the three ways to answer back, and nothing else.

• 6-9 tools, not 98 - every one earns its place in context, and the descriptions are tested so an agent picks the right one; see why so few and how it's compared with the alternatives.
• A write surface of exactly three tools - comment, transition, update. Set READ_ONLY_MODE=true and they never register, not even as a disabled entry the agent can see.
• Compact, predictable output - stable JSON schemas, Markdown for rich text, no null spam, no self URLs, emails, or avatars; see what the tools return.
• Roughly 40% less output on the wire - each result is sent once, as JSON in content, rather than again as structuredContent; set STRUCTURED_OUTPUT=true if your host needs the typed copy. See Cheaper output.
• Nothing is silently cut short - exact totals on comments and changelog, cursor paging on search, and limit=0 to fetch the rest. A failed request is an error, never an empty list.
• Errors an agent can act on - a wrong transition name lists every valid transition and where it leads, so there is no separate discovery tool. Errors never contain your Jira URL, credentials, or raw response bodies.
• API token or OAuth - one minute with an API token, or a browser login through your own OAuth 2.0 app with automatic refresh; see Configure.
• On PyPI - uvx jira-mini-mcp or pip install jira-mini-mcp, no repo clone or git URL required.

[!TIP]

Set READ_ONLY_MODE=true and only the six 🟢 read tools register - the

three 🔴 write tools are withheld entirely, see Configure.

• Install
• Configure
• Writing to Jira
• Why so few tools
• Compared with the alternatives
• Contributing and security

More detail lives in docs/: configuration, OAuth setup, and what the tools return, with examples.

Install

uvx jira-mini-mcp

or

pip install jira-mini-mcp

Pin a version when you want a fixed surface: uvx jira-mini-mcp==1.1.0.

Running an unreleased commit straight from GitHub also works:

uvx --from git+https://github.com/proprock/jira-mini-mcp jira-mini-mcp

Requires Python 3.12+ and uv (or pip).

Configure

With an API token, three values:

Add READ_ONLY_MODE=true to withhold the write tools. Every setting, including STRUCTURED_OUTPUT, is in configuration.md.

[!NOTE]

Prefer OAuth to a stored token? Set JIRA_AUTH_METHOD=oauth, register a

free OAuth 2.0 (3LO) app, and run jira-mini-mcp login once to authorize in

your browser. The server then refreshes its token by itself. Step by step:

oauth.md.

Configuration is validated at startup, and an error names the missing setting without printing its value or your Jira URL. Keep the token in the host's own configuration and never commit it. The server acts with your account's permissions: an account that cannot transition an issue still cannot, whatever this server exposes.

Claude Code

claude mcp add --env JIRA_BASE_URL=https://example.atlassian.net --env JIRA_EMAIL=you@example.com --env JIRA_API_TOKEN=your-token --transport stdio jira-mini -- uvx jira-mini-mcp

Put at least one other option between the last --env and the server name, as above - the CLI otherwise reads the name as another KEY=value pair.

Claude Desktop

In claude_desktop_config.json:

{

"mcpServers": { "jira-mini": { "command": "uvx", "args": ["jira-mini-mcp"], "env": { "JIRA_BASE_URL": "https://example.atlassian.net", "JIRA_EMAIL": "you@example.com", "JIRA_API_TOKEN": "your-token" } } } }

Codex CLI

codex mcp add jira-mini --env JIRA_BASE_URL=https://example.atlassian.net --env JIRA_EMAIL=you@example.com --env JIRA_API_TOKEN=your-token -- uvx jira-mini-mcp

Any other stdio host

Command uvx, argument jira-mini-mcp, and the three environment variables. Add READ_ONLY_MODE=true to withhold the write tools. OAuth host examples are in oauth.md.

Writing to Jira

Three tools, chosen so an agent can close the loop on a ticket it worked:

add_comment(issue_key, body)

transition_issue(issue_key, to, comment=None) update_issue(issue_key, fields)

Issue creation, links, attachment upload, worklogs, and deletion are out of scope. Creation needs per-project, per-type required-field discovery and is a feature in its own right; a link, or a request for one, fits in a comment.

Three things are worth knowing before an agent writes:

• `transition_issue` takes a name, not an id. A transition name or the name of the status to reach, matched ignoring case. They differ in real workflows - a transition called In Progress can produce a status called In Development, and two differently named transitions can reach one status - so prefer the transition name. When nothing matches, the error lists every available transition and where it leads. To see the options first, ask get_issue(fields=["transitions"]). That is why there is no separate get_transitions tool.
• `update_issue` replaces `labels` and `components` wholesale. There is no add or remove verb, so read the issue first if you mean to add one value. It takes the same values get_issue returns: assignee as an account id, an email, a display name, or the literal "me", description as Markdown, customfield_* as raw Jira JSON. It refuses status and comment, naming the tool that does each.
• Markdown is converted, not guessed at. Headings, lists, fenced code, inline marks, and links become Jira rich text; anything outside that set stays literal rather than being reinterpreted.

Each write tool is annotated readOnlyHint=false with honest destructiveHint and idempotentHint values, which is what READ_ONLY_MODE filters on.

Why so few tools

A tool definition is a name, a description, an input schema, and often an output contract. Depending on the client, all of it enters the model's context before any work happens. A large toolset therefore spends context on capabilities the current task will never use, and raises the chance of picking the wrong tool, confusing similar ones, or passing bad parameters.

Six to nine compact schemas stay affordable for a whole session, leaving the context budget for source code, issue descriptions, stack traces, and reasoning. The design follows Anthropic's guidance for agent systems: keep toolsets small, role-scoped, and clearly differentiated.

This is a claim, so the repository tests it. The offline half runs with the suite and checks that every tool is described substantially, that no two descriptions are near-duplicates, and that every parameter whose behavior cannot be guessed from its name is explained in prose. The other half puts the real tool definitions in front of a real model and scores which one it picks - see evals/README.md.

The same principle shapes the responses. get_issue does not dump hundreds of comments, the full changelog, attachment contents, or every custom field; large resources are fetched only when asked for.

Compared with the alternatives

The official server is the better choice when you need breadth across Atlassian products, OAuth without registering an app of your own, Jira Service Management, or organization-level controls such as permission groups, IP allowlisting, and audit logs. It is Atlassian's own product, it tracks their APIs, and nothing here competes with that.

`mcp-atlassian` is the better choice when you need Confl