Issue tracker MCP servers: which trackers your coding agent can actually use

Your agent has finished the change. It knows which files moved and which issue asked for it, and it is one tool call away from putting that issue in review. Whether it can make the call is a procurement question: does your tracker run an issue tracker MCP server, and on what terms.

The honest starting point is that running one stopped being a differentiator some time ago. Atlassian's Rovo MCP Server reached general availability on 4 February 2026 (atlassian.com), and its repository carries a "Status: Generally Available" badge (github.com/atlassian). Linear runs an official remote server at mcp.linear.app/mcp (linear.app/docs/mcp). Notion, Asana and GitHub publish first-party servers of their own. A vendor claiming novelty here is a year behind.

So "does it have one" is the wrong question. Four narrower ones decide whether the thing is usable, and the directory pages answer roughly one.

Who maintains it. A first-party server moves when the product moves. A community wrapper moves when a volunteer has a weekend.

What auth it uses. An OAuth grant and a bearer token fail differently. One is revoked centrally by an admin who can see the consent screen; the other is a string sitting in a config file on a laptop.

Which plan it needs. The column that decides whether you can use the thing at all, and the one nobody publishes on the page you land on.

What the agent may write. Read-only, write, or write plus closing work. Three risk profiles, usually presented as one switch.

Which trackers have an issue tracker MCP server

Every cell comes from the vendor's own documentation, opened in August 2026. Where a vendor does not state something, the cell says so rather than guessing.

The column nobody publishes

Try filling in "which plan" yourself and you will find how thin the ground is. Atlassian's getting-started page for the Rovo MCP Server names the products it reaches — "Jira, Jira Service Management, Confluence, and Bitbucket" — and says nothing about editions (support.atlassian.com). The security page describing it states only that "access is scoped to the user's existing permissions in Atlassian" (support.atlassian.com). The GA announcement does not mention plans either.

Linear is the clearest case of the ambiguity. Its own MCP documentation states no plan requirement at all (linear.app/docs/mcp). Its pricing page carries a row labelled "MCP access" inside the AI and agent workflows comparison (linear.app/pricing), which is where a restriction would live if there were one. We read that row three times and could not get its tier markers to render consistently, so we are not printing a tier for Linear. The cell says what we could actually see.

That is not a complaint about Linear, and it is not evidence of a hidden gate. It is the state of the documentation across this whole table: the question a buyer asks first — can the plan we are on do this — is answered on none of the integration pages above, and only half-answered on a pricing comparison. If you are choosing a tracker on the strength of its agent story, get the plan answer from the vendor in writing before you move anything.

Three kinds of server, and what each one gets wrong

The rows above are not one category. They are three, and knowing which one you are looking at predicts how it will fail.

Local, file-backed servers. Backlog.md keeps tasks as "human-readable Markdown in a project-local backlog folder" and states plainly: "no server, no account, no telemetry" (github.com/MrLesk). Nothing leaves the machine, there is no token to leak, and the task list branches and merges with the code. The failure mode is equally plain: it is invisible to anyone without the repository checked out. Fine for an agent's own working memory, useless as the thing your designer looks at on Tuesday.

Community wrappers over a hosted REST API. PulseMCP lists 91 servers matching a search for "jira" (pulsemcp.com), three of them marked official and the other 88 community-built. sooperset/mcp-atlassian is honest about what it is — the README says "Not an official Atlassian product" — and it carries around 5,700 stars (github.com/sooperset). These wrappers inherit the upstream API's auth model, usually a token as broad as your whole account, and its schema, which changes on the vendor's timetable and nobody's obligation.

First-party servers. Versioned and scoped by the vendor's own permission model. The cost is that you are inside that vendor's plan structure and consent model, and both are theirs to change.

Where the big servers genuinely win

Atlassian's is the strongest auth story here, and it is not close. The repository states that "authentication uses OAuth 2.1 or API tokens, so every action respects the user's existing access controls", and it reaches Jira, Jira Service Management, Confluence, Bitbucket Cloud and Compass from one connection (github.com/atlassian). An OAuth grant with real scopes beats a bearer token in a config file: revocable by an admin who can see it, never a copyable string, consent explicit. If your work is already on an Atlassian Cloud site, the argument ends there.

Linear's wins on a different axis. Its server is attached to a far more finished product, so an agent connected to Linear has more worth asking about. It also ships a read-only endpoint at mcp.linear.app/mcp/readonly alongside the full one (linear.app/docs/mcp). Handing an agent a URL that cannot write is the cheapest safety control on this list, and GitHub is the only other row here that documents one.

And GitHub's costs nothing. GitHub Free includes Issues and Projects on public and private repositories at $0 (github.com/pricing), its remote server is GA, and it goes further than anyone else on scoping the tool surface: toolsets choose which capabilities the model sees at all, and "write tools are skipped if --read-only is set, even if explicitly requested" (github.com/github).

Tokens against grants

Kevta's position on this table is narrow, and worth stating as narrowly as it deserves. The server is not gated: it is on Free and on Pro, and Free is $0 with 2 workspaces, 5 boards per workspace and 10 members per workspace — the numbers are on the pricing page. There is no tier to reach before your agent can reach your board.

The auth is a bearer token rather than an OAuth grant, and the design choice inside that is scope. A token is either user-wide or scoped to a single board, and a board-scoped token cannot touch another board even if the agent is handed another board's ID. Every token carries a fixed 7, 30 or 90-day life, is revocable from settings, is stored only as a one-way hash, and is shown exactly once. Permissions are re-checked against your live access on every request, so losing access to a board loses it for the token in the same instant. Full reference at /docs/api-tokens.

The concession: a bearer token is a copyable string in a plaintext client config, and no expiry policy makes that as good as a scoped OAuth grant. What it buys is a smaller blast radius per token and no org-wide consent screen. That is a trade, not a win.

There are nine tools: list_workspaces, list_projects, describe_board, list_tasks, get_task, create_task, update_task, add_comment, close_task. describe_board is the one worth knowing about — it returns that board's own fields, types and valid options, so an assistant sets a custom column correctly instead of guessing. Every other server here exposes more tools against a product that holds more than Kevta does, and there is no public REST API to fall back on when MCP is not enough.

Which one fits you

If your work is already on an Atlassian Cloud site, use Rovo. Better auth than anything else listed, and the product coverage is not close.

If your work is one repository and everyone involved has a GitHub account, use GitHub's server. Free, official, GA, and the only one that lets you decide which tools the model can see.

If what you want is somewhere for the agent to keep its own state between sessions, you want a local file-backed tool, not a hosted tracker. Backlog.md answers that better and costs nothing.

If you are on Linear, use Linear's server, and settle what your plan includes before anything depends on it.

Kevta fits one case: a small team that needs a board people outside the repository will read, wants the agent connection on the free tier rather than behind a plan, and would rather hand out a token scoped to one board than authorise an OAuth app across everything. If none of those three is true, a row above is the better tool, and we would rather say so than have you find out in week three. Setup is at /docs/mcp.

One thing applies whichever row you pick. Anthropic's MCP documentation puts it bluntly: "Verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection risk" (code.claude.com). A tracker is exactly such a server — ticket bodies are written by customers, contractors and bots, and your agent reads them as instructions. Scope the token accordingly. What else changes once agents do the writing is a longer argument.

If the Kevta case is yours, join the waitlist — an early invite, a founding-member price at launch, and a direct line to shape what we build next.

Disclosure

We build Kevta, which is one of the rows in that table, so read this as a comparison written by an interested party. Kevta is in beta. Everything above about competitors comes from their own documentation, opened in August 2026; where a vendor does not state something, the cell says so rather than carrying a number we inferred. Plans move, so the linked pages are the version that counts if this post has gone stale.