bitbucket-mcp
An MCP server that lets Claude Code read and write Bitbucket Cloud pull requests, comments, diffs, and Bitbucket Pipelines without leaving the terminal.
List and inspect pull requests — list PRs filtered by state with pagination, and fetch a single PR's details, commits, and full activity log (comments, approvals, pushes, status changes)
Create and edit PRs — open PRs or drafts (title, description, source/destination branch, reviewers by UUID, close-source option), update title/description/reviewers/destination, publish drafts, and convert open PRs back to draft
Read changes — fetch the raw diff of a PR or file-level diffstat (files added/modified/removed, lines changed)
Work with comments — list comments, add general or inline (file + line) comments and replies, edit them, and delete them
Manage Bitbucket Pipelines — list pipeline runs, get run details (state, duration, result), list steps and read step stdout/stderr logs
Trigger and stop pipelines — run a pipeline on a branch/tag with variables and selectors (default, custom, pull-requests), and abort a running pipeline
Target repos flexibly — workspace and repo slug auto-detect from the
bitbucket.orggit remote, or can be passed explicitly to any tool
Note: the registered schema (21 tools) differs from the README (27 tools) — the Jenkins build-report, CI health, user-lookup (findUsers/getCurrentUser), comment resolve/reopen, and reviewer add/remove tools described in the README are not present; instead the schema includes Bitbucket Pipelines tools.
Provides tools for interacting with Bitbucket Cloud, enabling pull request management, comments, diffs, pipeline runs, and activity tracking.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@bitbucket-mcp show open pull requests for the current repo"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
bitbucket-mcp
A Bitbucket Cloud MCP server for Claude Code. Lets you interact with pull requests, comments and, optionally, their Jenkins builds without leaving the terminal.
Scope
This server fetches and shapes Bitbucket data. It returns compact JSON records and never renders them for a reader. Deciding which tool answers a question, composing several calls, and formatting the result for a person all belong to the caller (in Claude Code, the bitbucket-fetcher agent and the skills it loads).
Related MCP server: Atlassian Bitbucket MCP Server
Tools (27)
Listing tools return a compact record per PR (id, title, state, draft, author, branches, local timestamps, counts, URL) and fetch every page up to limit (default 25, max 100), prefixed with a line saying how many were returned and whether more exist. When 20 or fewer PRs come back, each also carries build (latest build status: state, name, link, commit, total_statuses) and review (reviewers, approved_by, changes_requested_by, pending). Pass include_details to force this on or off, and verbose: true for the raw API objects.
By default only PRs updated in the last 30 days are returned. Pass updated_within_days to widen the window, or 0 for no cutoff. Filter server-side wherever possible (author, reviewer, title_contains, states) rather than pulling a large list and filtering it afterwards.
Pull Requests
listPullRequests — list PRs in a repo. Filters:
states(one or more of OPEN/MERGED/DECLINED/SUPERSEDED),authorandreviewer(me, a name fragment, or a uuid; an array matches any of several people in one call),title_contains,updated_within_days,sort(default-updated_on),limitlistMyPullRequests — PRs authored by the authenticated user across every repo in the workspace. Same filters minus
author,reviewer,sortgetPullRequest — a single PR: metadata, description, latest build status, reviewers and review state
getPullRequestBuildStatuses — every build status reported on a PR, newest first
getPullRequestDiffstat — per-file change summary (status, path, lines added and removed) with a totals line. Sizes a PR without fetching the diff;
verbose: truefor the raw objectscreatePullRequest — create a PR with title, description, reviewers (names or uuids)
updatePullRequest — edit title, description, destination branch, reviewers.
add_reviewersandremove_reviewersadjust the current list in one call;reviewersreplaces it. People are names or uuidsgetPullRequestCommits — list commits on a PR
createDraftPullRequest — create a draft PR with title, description, reviewers (names or uuids)
publishDraftPullRequest — mark a draft as ready for review
convertToDraft — convert an open PR back to draft
Comments
getPullRequestComments — every non-deleted comment on a PR in chronological order, as compact records (author, local timestamp, body, inline file/line, reply_to). Fetches every page up to
limit;verbose: truefor the raw objectsaddPullRequestComment — add a general or inline comment (file + line)
updatePullRequestComment — edit an existing comment
resolvePullRequestComment — mark a comment thread resolved
reopenPullRequestComment — reopen a resolved thread
deletePullRequestComment — delete a comment
Liking a comment is not offered: the public Bitbucket Cloud API has no endpoint for it. The web UI uses an internal endpoint that is not supported for API tokens.
Activity
getPullRequestActivity — chronological log of PR events (comments, approvals, pushes)
Jenkins (optional, see setup)
getPullRequestBuildReport — explains one PR build. Returns
checked_at_local(when the report was read),outcome(passed,in_progress,aborted,tests_failed,failed), the stage the builddied_in, the stages that failed in flight, were aborted, or never ran as a consequence, the failing tests with the file, example, error line and how many builds each has been failing, and for every failing test itsalignmentwith the files the PR author's own commits changed:direct(the test itself changed),subject(the code it names changed),neighbourhood(the same corner of the codebase changed), ornone. Alignment is measured against the author's commits rather than the PR diffstat so a merge from upstream does not claim its inherited changes as the author's; each failing test also carrieschanged_by, the PR commits that touched it with their authors. When a PR has more than 50 commits the diffstat is used instead andalignment_basissays so. Afailedoutcome carries aclassificationmatched from the console (merge_conflict, checkout_failed, formatting, lint, dependency_audit, quality_gate, coverage, timeout, agent_offline, disk_full) orunrecognised, and always the console lines that named the failure plus a console link, so an unfamiliar failure still comes back with its evidence. Whenever a stage died,failed_stepnames the step inside it that failed, the command it ran, a link to that step's own log, and anexcerptof its output:excerpt_kindisrspec_failures(theFailures:block, or anAn error occurred while loadingblock, through the summary line plus therspec ./spec/...rerun lines),jest_failures,ruby_exception(the exception and its backtrace), ortail(the last 40 non-empty lines) when nothing recognisable is present; long excerpts keep their head and summary and say how many lines were omitted. When the command sends rspec output to a file (--out rspec.txt) the step log holds none of it, so if the build archived that file the excerpt is read from the artifact andexcerpt_sourceisartifact:<path>(otherwisestep_log); an--outfile the build did not archive is named inartifact_missing. The same artifact text fillstests.failedwhen no junit report lists failures, including a spec that raised while loading (its example reads(error while loading the file);failing_for_buildsis absent since only the junit report knows it).excerpt_low_value: truemarks atailexcerpt on a build whosetests.failedalready names the failures. The full console never leaves the server.build_numberselects an older build of the same PR;jobnames the PR's job outright when neither the status URL norJENKINS_JOBSresolves it. Anabortedbuild with nofailed_stepcarriesaborted:by(Jenkins's interruption cause, "Aborted by someone@example.com"),in_stage(the first stage cut short) andstage_log_tail(the last ~20 clean lines of that stage's aborted step, or of the build console when no stage had started). Anin_progressreport carrieslast_completed_build, so the caller can diagnose the most recent finished run instead, andexecution, described belowgetJenkinsBuildReport — explains one build of any Jenkins job by name (
my-app-master, ormy-app/masterfor a branch job inside a folder), for builds that belong to no pull request such as a master or nightly job. Returns the same record asgetPullRequestBuildReportminus the alignment fields:checked_at_local,outcome,died_in, the stage cascade, failing tests with file, example, error line and how many consecutive builds each has failed, console lines,failed_stepwith its command and log excerpt,aborted, classification andbitbucket_incidents, plusbuild.started_localso a dated question gets a dated answer. Defaults to the job's latest build;build_numberselects an older one. Whether a failure is new is answered by fetching the earlier build too and comparing the two reports, which is the caller's jobgetJenkinsConsoleTail — the end of a build's console, for "what is this build doing right now" and "where did it stop". Takes
job(same forms as above),build_number(default: the latest build),kb(default 16, max 256) and returnsbuild_url,size_bytes,last_output_local(the timestamp of the last line written) andlines: the lastkbkilobytes read with two cheap requests (a HEAD oflogText/progressiveTextfor the size, then a GET fromsize − kb*1024), with Jenkins's ConsoleNote blobs stripped, blank lines dropped and timestamps removed. Withstage(a stage or parallel-branch name as the report lists it, matched whole first then as a fragment, latest of that name winning) it reads the current or last step log of that stage instead (execution/node/<step>/log/logText/progressiveText, plain text) and addsstage,stepandstep_command. A step log is the clean per-branch output; the whole console interleaves every branch. Two calls a few minutes apart comparesize_bytesto tell a working build from a hung one
execution on an in-progress build
A pipeline build counts as building from the moment Jenkins accepts it, hours before it gets an agent, so a build's age says nothing about what it is doing. execution says what it is doing, read from the workflow API, the queue, the computer list and the logs:
state:queued(no stage has started, so no agent has been assigned),executing(every unfinished stage has a running step),waiting_for_agent(every unfinished stage sits in the queue for an executor) ormixed(some of each). Unfinished means the stage isIN_PROGRESSor Jenkins reports itSUCCESSwith a negative duration, which is what a live stage looks likestarted_local,waited_for_first_agent(first stage start minus run start; null when no stage has started, in which casequeued_forsays how long it has been queued,queue_reasongives Jenkins's words andqueue_positionits place, oldest first, among the queue items waiting for the same reason, which is the same label),on_executor_for(now minus the first stage start; never the build start),estimated_duration(the previous run's length, Jenkins's ETA baseline; null on a job's first build)stages: every stage and parallel branch withstarted_localandstatusdone(ran_for),executing(running_formeasured from the current step,step_command: the shell step's exact command) orwaiting_for_agent(waiting_formeasured from the queue item,wait_reason: Jenkins's own words). A waiting stage is paired with a queue item of the same build that entered the queue within ~2 s of the stage starting; a step-less stage with no queue item is reported as waiting with await_reasonthat says so (and says when the build holds a spare agent, meaning the stage was just assigned one). Wrapper stages, whose span only encloses other stages, are left out, since their duration is the wait for their first child rather than workagents_held: the nodes whose executors run this build, each with the stage or branch on it, from the computer list;timestampthere is the build start, so time on each agent is not knownlog:size_bytesof the console;growing, which is null unless the caller passedprevious_log_size(thesize_bytesfrom a report minutes earlier) orcheck_liveness: true(the server samples again after 60 s, so the call takes a minute longer), because two reads seconds apart prove nothing;current_step_tail, the last ~20 lines of the most recently started executing step's own log, colour codes and timestamps stripped; and from those timestampslast_output_localandsilent_for, how long that step has written nothingverdict: one plain sentence, e.g. "Executing on 3 agents for 2h 43m after waiting 8h 08m; current stepbundle exec rspec …in Run Feature Tests Batch 2 running 16m, liveness not sampled (call again with previous_log_size or check_liveness), last output 12s ago." or "Waiting for an agent on 2 stages for 1h 39m (executors saturated); nothing is executing, though 1 agent is still held." or "Between stages: no stage has a running step, holding 1 agent, …". It opens with "Possibly hung:" when the console has not grown since the earlier read and the current step has also written nothing for 2 minutes (a console unchanged while the step wrote seconds ago is reported as "too soon to call"), when the current step has been silent for over 10 minutes, or when every executing branch's log reports all tests finished and has been silent 2 minutes since, with or without another branch still waiting for an agent. No estimate is needed; when there is one and the build is past it, the sentence says by how muchlistJenkinsBuilds — the most recent builds of a Jenkins job by name, newest first: number, result (
SUCCESS,UNSTABLE,FAILURE,ABORTED, or null while building),started_local, duration and URL. Finds the build that ran on a given night before callinggetJenkinsBuildReport; a job that also builds on merges has several builds per day, so the caller picks by start time.limitdefaults to 20, max 100retriggerPullRequestBuild — queues a new build of the PR's Jenkins job with a plain POST, no push or commit involved. Checks Jenkins health first and refuses, naming the reason, while Jenkins is quieting down or has no online agents;
force: truequeues it anyway;jobnames the PR's job outright when neither the status URL norJENKINS_JOBSresolves it. Returns the queue item URL and the job URL; the build has not run when the call returnsdiscoverJenkinsPipeline — samples recent completed builds and describes the pipeline as observed: stage names and which run in parallel, outcome counts, how often a test report is published, and the failure signatures seen with example console lines. Given a multibranch folder (
job, default the repo's folder perJENKINS_JOBS, else its slug) it samples the latest completed build of each branch job; given a standalone job such as a master or nightly job it samples that job's lastsample_sizebuilds, andkindsays which it found. Meant for building and refreshing a local manifest of how a given Jenkins behaves, rather than hardcoding one pipeline's shape
CI health
getCiHealth — how CI is looking right now, with a verdict.
assessment.statusisok,degradedordown;assessment.summaryis the one-line answer, opening with the local time it was read, and when things are fine that line is the whole story ("As of 09:14 MDT, situation normal: 3 builds running, nothing queued, all agents online").assessment.findingsis empty when ok; otherwise each names a condition, the evidence and what to do: Jenkins unreachable, a label no online node serves (queued builds stuck), quieting down for a restart, agents offline with reasons, executors saturated with a backlog, a Bitbucket incident,agents_held_while_waiting("N builds hold agents while waiting 30m or longer for more", listing up to five builds with the agents and stages each holds, how many of its stages are queued and for how long: the starvation pattern, where builds keep their agents until every branch has run while their other branches queue for agents the other such builds hold; whether a held agent is working or idle is in that build's report, and cancelling the build nearest to finishing lengthens the wait for everyone), or an unrecognised long wait carrying Jenkins's own queue reasons. Underneath: Bitbucket's status page, and for Jenkins the queue items (task,buildit belongs to as an absolute URL, queued since in local time,waiting_minutes, reason, stuck), every node withexecutorsbusy of total pluscurrent, one{ build, stage }per busy executor from its current executable, and quiet mode. Available without Jenkins; the Jenkins half appears when it is configured. Afailedbuild report also carriesbitbucket_incidents, the status-page incidents that overlapped the build, and any Bitbucket API error of 500 or above is suffixed with what the status page reports at that moment
Users
findUsers — resolve one name or several the way every other tool does. Per name:
resolved, the single member the server picks (a whole name beats a whole word, beats a word that starts with the fragment, beats a substring, sostanis Stanley before it is Tristan), or null when the strongest matches tie; andcandidates, every match with itsmatchstrength. Not a required first step:author,reviewerand the reviewer parameters take names directly and resolve them identicallygetCurrentUser — the user the server is authenticated as
Setup
The Bitbucket tools need only the Bitbucket token. The Jenkins tools are registered only when JENKINS_URL and JENKINS_API_TOKEN are set; without them the server runs with the other 21 tools and nothing else changes. getCiHealth is always present; its Jenkins half appears when Jenkins is configured.
1. Create a Bitbucket API token
Go to Atlassian account > Security > API tokens and create a token with these scopes:
read:repository:bitbucketread:pullrequest:bitbucketread:user:bitbucket(resolvesmeto your account)read:workspace:bitbucket(member lookup forfindUsersand name-basedauthorfilters)write:pullrequest:bitbucket
The server authenticates with your Atlassian account email plus this token (HTTP Basic). See Bitbucket API tokens.
1b. Jenkins (optional)
Create a Jenkins API token from your user page (your name > Security > API Token) and set:
export JENKINS_URL=https://jenkins.example.com
export JENKINS_API_TOKEN=your-jenkins-tokenThe Jenkins username defaults to BITBUCKET_EMAIL; set JENKINS_USER only when Jenkins knows you by a different name.
The PR's Jenkins job is read from the URL in its Bitbucket build status, so the folder can be named anything. When the PR's head commit has no status yet, the multibranch convention <folder>/PR-<id> is tried instead, where the folder is the repo slug unless JENKINS_JOBS says otherwise. List each repository's folder as comma-separated slug=folder pairs; it is one setting for every repository the server sees, and a slug it does not list keeps its own name, so only the repositories whose folder differs strictly need an entry:
export JENKINS_JOBS=zymewire-rails-app=zymewire-rails-app,smartsalesassistant=smart-sales-assistantdiscoverJenkinsPipeline defaults its job the same way; pass it to profile a standalone job. getPullRequestBuildReport and retriggerPullRequestBuild also take an optional job (<folder>/PR-<id>) that bypasses both the status lookup and the fallback. When a build cannot be found, the error says whether the job path itself resolves to nothing or the job exists with no such build, so a wrong folder name is not mistaken for a PR that never built. getJenkinsBuildReport and listJenkinsBuilds take a job name directly, so a master or nightly job that lives outside the multibranch folder is reachable by its own name.
When Jenkins is configured but unreachable (an internal host with the VPN down) or rejects the token, the Jenkins tools return an error naming the cause. The Bitbucket tools are unaffected either way.
2. Build
cd bitbucket-mcp
npm install
npm run build3. Export the token from your shell
Keep the token out of config files. Export it from your shell profile (~/.bashrc, ~/.zshrc):
export BITBUCKET_API_TOKEN=your-api-tokenOpen a new terminal so the export is active before starting Claude Code.
4. Register with Claude Code
Register a user-scope server. The ${BITBUCKET_API_TOKEN} placeholder is stored literally in ~/.claude.json and expanded from your shell environment when the server starts, so the token itself never lands on disk in a config file:
claude mcp add bitbucket --scope user \
-e BITBUCKET_EMAIL=your-bitbucket-email \
-e 'BITBUCKET_API_TOKEN=${BITBUCKET_API_TOKEN}' \
-- node /path/to/bitbucket-mcp/dist/index.jsWith Jenkins, add its variables the same way, before the --:
-e JENKINS_URL=https://jenkins.example.com \
-e 'JENKINS_API_TOKEN=${JENKINS_API_TOKEN}' \
-e JENKINS_JOBS=zymewire-rails-app=zymewire-rails-app,smartsalesassistant=smart-sales-assistant \The server name must come before the first -e, since -e takes every following value until the next flag. There is no edit command; to change the registration, claude mcp remove bitbucket --scope user and add it again. Paste the command as one line if a multi-line paste splits it.
Verify with claude mcp list. It should show bitbucket as connected; a missing-variable warning means the export is not visible to the shell that launched Claude Code.
Note: mcpServers in ~/.claude/settings.json is not read by Claude Code. User-scope servers live in ~/.claude.json via the command above.
Auto-detection
When you're working inside a project that has a bitbucket.org git remote, the workspace and repo slug are auto-detected from git remote get-url origin. You can also pass workspace and repo_slug explicitly to any tool to target a different repo.
Suggested usage
The server returns compact JSON records and never formats output for a human reader. Deciding which tools answer a question, composing calls, and rendering the result are the caller's job. A clean way to set this up in Claude Code:
Read path: a dedicated agent + rendering skill
A read-only subagent (e.g. bitbucket-fetcher) receives the user's question verbatim, calls the appropriate read tools, and renders the result using a rendering skill (e.g. bitbucket-brief) that defines fixed output shapes: a PR list table, a single-PR block, a comment thread, a build report. The agent keeps raw API JSON out of the coordinator's context and returns a finished brief that gets relayed to the user without re-summarizing. Every read goes through it, including CI health; the coordinator never calls a read tool itself.
The agent should have access to the read-only tools only: listPullRequests, listMyPullRequests, getPullRequest, getPullRequestBuildStatuses, getPullRequestBuildReport, getJenkinsBuildReport, getJenkinsConsoleTail, listJenkinsBuilds, getPullRequestDiffstat, getPullRequestCommits, getPullRequestComments, getPullRequestActivity, discoverJenkinsPipeline, getCiHealth, findUsers, getCurrentUser.
Rules that keep the agent honest, learned the hard way:
Pass the user's words, not a resolved target. The coordinator hands over the question as asked and never substitutes a PR number or a person it worked out from the git log, a branch name, or a commit message. A recent-commits block looks like an answer key and is not one; the server resolves names and lists candidates when a name is ambiguous.
Names go straight into the tools.
author,reviewerand the reviewer parameters accept names, and arrays of names, so "Bill's, Ryan's and Tommy's PRs" is one call.findUsersshows what a name resolves to and its candidates; use it after a tie is reported, not before every call.A named build number is a hard constraint. "Build 2 of X" means
build_number: 2on every candidate PR, never the latest build that happens to be on a list record.A question that assumes a failed build expects it found. An
in_progressreport carrieslast_completed_build; diagnose that one and mention the run in progress in passing.A building build is read from
execution, never from its age.execution.verdictis the answer to "what is it doing";statetells queued from executing from waiting for an agent,on_executor_foris the only "running for" there is, andgetJenkinsConsoleTailwith astageshows what a branch is doing right now. A build that is executing with a growing log is working, however old it is.Read
alignmentwithchanged_by.noneplus achanged_bycommit by someone other than the PR author means the test arrived through a merge from upstream, and the brief should say whose commit changed it.Every brief says when it was read. Agents do not feel time pass. Build reports and CI health carry
checked_at_local, and the brief states it, so a reader knows a verdict from an hour ago is not the situation now.CI health is a verdict, not a dashboard.
getCiHealthreturnsassessment.summary; when the status isok, that one line is the answer and nothing else is shown. Findings appear only when something is wrong, each with evidence and advice.Non-PR builds are found by date, then compared by the agent. A master job builds on merges as well as on a schedule, so "last night's build" is the entry in
listJenkinsBuildswhosestarted_localfalls in the nightly window, not the latest build. "Is this the same failure as three nights ago" is twogetJenkinsBuildReportcalls and a comparison of their failing tests and console lines; the server never answers that from the age counter alone.Infrastructure gets named. When a failure classifies as
agent_offline,timeoutorunrecognised, or the report listsbitbucket_incidents, the agent callsgetCiHealthand reports what the services say rather than leaving the user to guess.
Jenkins manifest
Pipelines differ, so the agent keeps its own notes on the one it reads. It stores the output of discoverJenkinsPipeline per repository (e.g. ~/.claude/jenkins/<workspace>-<repo>.json), reads it before interpreting a build report, and refreshes it when the file is missing, a report names a stage the manifest does not list, a classification comes back unrecognised, or the manifest is older than a month. The manifest records what Jenkins shows: stages, parallel groups, outcome counts and failure signatures. A standalone master or nightly job outside the multibranch folder gets its own manifest, sampled by passing its name as job. Team conventions the tool cannot observe, such as the job's name, when the nightly runs, or how its result differs from a PR build's, belong in the agent's instructions, not the manifest.
Write path: the coordinator
Write tools (createPullRequest, updatePullRequest, addPullRequestComment, retriggerPullRequestBuild, etc.) stay with the coordinator or a writing agent, not the read-only fetcher, and run only on the user's explicit go. A PR description skill can compose the text and post it through the MCP server in a single flow. The retrigger is a real build; the agent that diagnosed the failure may suggest it but never runs it.
Permissions
Register the read-only tools in your Claude Code permissions allowlist so they run without approval prompts. Leave every write tool, the retrigger included, on the default approval flow.
API
Targets the Bitbucket Cloud REST API v2.0.
Available Tools
21 toolsaddPullRequestCommentB
Add a comment to a pull request (general or inline on a specific file/line)
Inputs: content (required), pull_request_id (required), inline_to, parent_id, +4 more.
convertToDraftB
Convert an open pull request back to draft
Inputs: pull_request_id (required), repo_slug, workspace.
createDraftPullRequestC
Create a draft pull request
Inputs: title (required), source_branch (required), repo_slug, workspace, +2 more.
createPullRequestB
Create a new pull request
Inputs: title (required), source_branch (required), repo_slug, reviewers, +4 more.
deletePullRequestCommentB
Delete a comment from a pull request
Inputs: comment_id (required), pull_request_id (required), repo_slug, workspace.
getPipelineRunB
Get details of a specific pipeline run (state, duration, result)
Inputs: pipeline_uuid (required), repo_slug, workspace.
getPipelineStepLogsA
Get stdout/stderr logs from a specific pipeline step
Inputs: step_uuid (required), pipeline_uuid (required), repo_slug, workspace.
getPipelineStepsB
List steps within a pipeline run and their results
Inputs: pipeline_uuid (required), repo_slug, workspace.
getPullRequestC
Get details of a specific pull request
Inputs: pull_request_id (required), repo_slug, workspace.
getPullRequestActivityB
Get the activity log for a pull request (comments, approvals, pushes, status changes)
Inputs: pull_request_id (required), page, pagelen, repo_slug, +1 more.
getPullRequestCommentsC
List all comments on a pull request
Inputs: pull_request_id (required), page, pagelen, repo_slug, +1 more.
getPullRequestCommitsC
List commits on a pull request
Inputs: pull_request_id (required), page, pagelen, repo_slug, +1 more.
getPullRequestDiffC
Get the raw diff of a pull request
Inputs: pull_request_id (required), repo_slug, workspace.
getPullRequestDiffStatC
Get file-level change statistics for a pull request (files added/modified/removed, lines changed)
Inputs: pull_request_id (required), page, pagelen, repo_slug, +1 more.
listPipelineRunsC
List recent pipeline runs, optionally filtered by status or branch
Inputs: page, sort, pagelen, repo_slug, +1 more.
listPullRequestsC
List pull requests in a repository, filtered by state
Inputs: page, state, pagelen, repo_slug, +1 more.
publishDraftPullRequestA
Publish a draft pull request (mark it as ready for review)
Inputs: pull_request_id (required), repo_slug, workspace.
runPipelineB
Trigger a new pipeline run on a branch, tag, or custom pipeline
Inputs: ref_name (required), ref_type, repo_slug, variables, +3 more.
stopPipelineA
Stop/abort a running pipeline
Inputs: pipeline_uuid (required), repo_slug, workspace.
updatePullRequestB
Update a pull request (title, description, reviewers, destination branch)
Inputs: pull_request_id (required), title, repo_slug, reviewers, +3 more.
updatePullRequestCommentC
Update an existing comment on a pull request
Inputs: content (required), comment_id (required), pull_request_id (required), repo_slug, +1 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 21 tools
Each tool targets a distinct resource and action, and the pull request and pipeline namespaces are clearly separated. Minor overlap exists between getPullRequestComments and getPullRequestActivity, and between getPullRequestDiff and getPullRequestDiffStat, but the descriptions are clear enough to resolve ambiguity.
Tool names consistently use camelCase and follow a clear verb_noun pattern, with list/get/create/update/delete verbs and PR or pipeline context prefixes. There are minor inconsistencies like addPullRequestComment vs createPullRequest and convertToDraft, but the naming remains predictable overall.
At 21 tools, the server is on the heavy side and covers two fairly broad domains: pull requests and pipelines. The count feels borderline; some tools like getPullRequestDiff and getPullRequestDiffStat could potentially be consolidated, though each still serves a distinct use case.
The pull request lifecycle covers creation, updating, drafts, comments, diffs, commits, and activity, but core Bitbucket actions like merge, decline, and approve are missing. Pipeline coverage is solid with run, stop, steps, and logs, but the PR gaps prevent full workflow completion.
Maintenance
Related MCP Connectors
Plan Salesforce deploys, open pull requests and trigger pipelines from your AI client.
Code intelligence for LLMs. Analyze, search, and retrieve code from any public git repository.
Share context and questions between Claude instances — VS Code, claude.ai web, and mobile.
Multi-brand AI code review for Claude: up to 12 reviewers and a judge, no install.
Related MCP Servers
- AlicenseAqualityDmaintenanceAn integration tool that enables AI assistants like Claude to directly access and interact with Bitbucket repositories, pull requests, and code without requiring copy/paste operations.69,115 npm163ISC
- AlicenseBqualityCmaintenanceEnables AI assistants to interact with Bitbucket Cloud and self-hosted instances for pull request reviews, code search, repository operations, and managing PR comments and approvals.19GPL 3.0
- AlicenseAqualityDmaintenanceConnects Claude Desktop and Claude Code to Bitbucket Cloud, enabling searching PRs, commits, diffs, and file content for AI-assisted bug triage and impact analysis.840 npmISC
- AlicenseNot gradedqualityCmaintenanceEnables pull-request review workflows on Bitbucket Cloud via Claude Code, offering tools to view PR metadata, comments, diffs, and to post comments or approve.6,189 npmMIT