a2atlassian

Agent: "What's the status of PROJ-42? Add a comment with the progress update."
  ↓
a2atlassian → get issue, add comment, transition to In Progress
  ↓
Agent: "Done — PROJ-42 updated and moved to In Progress."

Why a2atlassian?

Existing Atlassian MCP servers (Rovo, sooperset) require Docker, .env files, and mcp-remote bridges. They dump 72 tools into agent context and have known quirks that silently fail. a2atlassian fixes all of that:

• No Docker — pip install a2atlassian and you're done

• Pre-configured connections — define projects in .mcp.json with --register, agent works immediately

• Read-only by default — write access is opt-in per connection

• Connection scoping — --scope limits which projects an agent can see

• Compact output — TSV for lists (30-60% fewer tokens), JSON for single entities

• Dynamic tool loading — MCP clients that support deferred tools (e.g., Claude Code) load tools on demand, keeping context lean

• Error enrichment — bad field names get suggestions, JQL typos get corrections, quirks get auto-fixed

• Secrets stay in env — ${ATLASSIAN_TOKEN} in configs, expanded only at runtime

Scope today: full Jira surface (issues, comments, sprints, boards, worklogs, links, versions, fields, watchers, projects) and Confluence core (pages CRUD, search, metadata-only writes).

Quick Start

# Recommended — installs globally as a CLI tool
uv tool install a2atlassian

# Or with pip
pip install a2atlassian

As an MCP Server (recommended)

Claude Code (with pre-configured connection):

claude mcp add -s user a2atlassian -- uvx --from a2atlassian a2atlassian-mcp \
  --register myproject https://mysite.atlassian.net user@company.com '${ATLASSIAN_TOKEN}'

Claude Code (minimal — agent calls login on demand):

claude mcp add -s user a2atlassian -- uvx --from a2atlassian a2atlassian-mcp

Claude Desktop / Cursor / any MCP client (.mcp.json):

{
  "mcpServers": {
    "a2atlassian": {
      "command": "uvx",
      "args": [
        "--from", "a2atlassian", "a2atlassian-mcp",
        "--register", "myproject", "https://mysite.atlassian.net",
        "user@company.com", "${ATLASSIAN_TOKEN}"
      ],
      "env": {
        "ATLASSIAN_TOKEN": "your-api-token-here"
      }
    }
  }
}

Multiple projects:

{
  "args": [
    "--from", "a2atlassian", "a2atlassian-mcp",
    "--register", "myproject", "https://mysite.atlassian.net", "user@a.com", "${TOKEN_A}",
    "--register", "personal", "https://personal.atlassian.net", "user@b.com", "${TOKEN_B}"
  ]
}

Scoped connections (limit agent to specific saved projects):

{
  "args": ["--from", "a2atlassian", "a2atlassian-mcp", "--scope", "myproject"]
}

--register creates ephemeral in-memory connections (process lifetime, no files written). --scope filters which saved connections are visible. Both limit blast radius.

As a CLI

# Save a connection (validates by calling /myself)
a2atlassian login -c myproject \
  --url https://mysite.atlassian.net \
  --email user@company.com \
  --token "$ATLASSIAN_TOKEN"

# Same, pulling the token from 1Password via `op`
a2atlassian login -c myproject \
  --url https://mysite.atlassian.net \
  --email user@company.com \
  --token "op://Personal/Atlassian/token"

# Enable writes
a2atlassian login -c myproject \
  --url https://mysite.atlassian.net \
  --email user@company.com \
  --token "$ATLASSIAN_TOKEN" \
  --no-read-only

# List / remove connections
a2atlassian connections
a2atlassian logout -c myproject

Tokens accept three forms: literal value, ${ENV_VAR} reference, or op://vault/item/field (resolved via the 1Password CLI at runtime).

MCP Tools

Connection Management

Jira — Read

Jira — Write (requires read-write connection)

Confluence — Read

Confluence — Write (requires read-write connection)

Output Formats

All tools accept a format parameter:

List responses use a compact TSV-style format (header row + tab-separated values) inspired by TOON. This is the same approach a2db uses — column names appear once, then just values. For a 50-issue search result, this typically saves 40-60% of tokens compared to JSON.

TSV example (search results):

# search (23 results, 50ms, truncated: False)
key	summary	assignee	status
PROJ-142	Fix auth timeout	Alice Smith	In Progress
PROJ-141	Add search filters	Bob Jones	To Do

JSON example (single issue):

{
  "data": {"key": "PROJ-142", "fields": {"summary": "Fix auth timeout", ...}},
  "count": 1,
  "truncated": false,
  "time_ms": 85
}

Error Enrichment

When something fails, a2atlassian tells the agent what to do:

Field 'asignee' does not exist
Did you mean: assignee?

Connection 'myproject' is read-only.
Run: a2atlassian login -p myproject --read-only false

Quirks handled automatically:

• Assignee requires display name (not 712020: account IDs) — auto-detected with hint

• Parent field must be plain string — {"key": "PROJ-14"} normalized to "PROJ-14" silently

• Issue type conversion not supported via API — clear Jira UI instructions provided

Security

Read-Only by Default

Every connection starts read-only. Write tools check the connection flag before executing:

Connection 'myproject' is read-only.
Re-run 'a2atlassian login -p myproject --read-only false' to enable writes.

The human operator controls write access — not the agent.

Credential Storage

Connections saved via login go to ~/.config/a2atlassian/connections/ as TOML files:

• File permissions: 0600 (owner read/write only)

• ${ATLASSIAN_TOKEN} syntax — env var references stored literally, expanded at runtime

• No secrets in output — list_connections shows project name, URL, and mode — never tokens

• Ephemeral mode — --register keeps credentials in memory only, never written to disk

Connection Scoping

Use --scope to limit which saved connections a specific MCP instance can access:

# Project config — only myproject visible, even if other connections are saved
uvx --from a2atlassian a2atlassian-mcp --scope myproject

Project-level MCP configs (.claude/mcp.json) override global configs — each repo sees only its own connections.

Rate Limiting

Built-in retry with exponential backoff for Atlassian's rate limits (429) and transient server errors (500). Two retries at 1s and 3s intervals before surfacing the error.

Comparison

Roadmap

Shipped: Jira full surface (v0.3.0) · Confluence core + markdown-to-storage with full CommonMark + GFM fidelity (v0.4.0, v0.5.2) · 1Password op:// token refs (v0.5.1) · metadata-only Confluence writes + preserve-on-omit body semantics (v0.5.2).

Next: confluence_delete_page, Confluence comments + attachments, Confluence integration-test path. Backlog in TODO.md.

Setup by Environment

Local (macOS / Linux)

# Recommended
uv tool install a2atlassian

# Or with pip
pip install a2atlassian

# CLI
a2atlassian login -p myproject --url https://mysite.atlassian.net --email me@co.com --token "$TOKEN"

# Or add as MCP server (see Quick Start)

CI / Automation

uv tool install a2atlassian

# Pre-configured — no login needed
uvx --from a2atlassian a2atlassian-mcp --register ci https://mysite.atlassian.net ci-user@co.com "${CI_ATLASSIAN_TOKEN}"

Development

make bootstrap   # Install deps + pnpm + git hooks
make check       # Lint + test + coverage-diff + security (full gate)
make test        # Tests with coverage
make lint        # agent-harness + jscpd + actionlint (never modifies files)
make fix         # Auto-fix + lint
make similar     # Advisory: report similarly-named functions/classes

Linters: ruff + ty (via agent-harness), yamllint, jscpd (copy-paste detection via pnpm), actionlint (GitHub Actions workflows). Pre-commit hooks run agent-harness fix + lint on every commit. Install pnpm and actionlint via brew install pnpm actionlint.

License

Apache 2.0