MCP server

Splitkick has a built-in Model Context Protocol server. With it, an agent can do almost everything that you do in the app: find and read sessions, create worktrees, start sessions and send them work, work in a graph, commit, push, open and merge a pull request, run a review. The server has 62 tools. That makes Splitkick a place where agents work together, not only side by side; see Cross-session workflows.

What is new

• 1.5.3: add_artifact puts a file or link that the agent made into the Artifacts tab of a session. See Artifacts.
• 1.5.1: list_instances and remote_tool let an agent use your other Splitkick instances, for example to start a session on another Mac and read its answer. See Work on your other Splitkick instances. Also: create_session takes over the empty "Session 1" of a worktree, get_current_context returns the graph links of the session, get_usage returns every Claude account, list_projects shows the project kind, and the git tools give an error for a plain project.
• 1.4: Graph tools (1.4.1 added list_requests, list_changes and create_graph): agents read graphs, create and change nodes, send you requests and link sessions to nodes. See Graphs below and Graphs and agents.
• 1.2: move_session moves a session to a worktree in another project too. Claude continues its own conversation there; other agents get the history replayed with the next message. A session with queued messages is refused; a running session moves when its turn ends. See Move and transfer sessions.
• 1.1: add_follow_up puts a note on your follow-up list. list_all_sessions, search_sessions and get_current_context return a splitkick:// link for each session, which opens on click in a chat. A session that an agent starts shows "Started by '…'" at the top. A mode given with send_chat_message becomes the session's mode.

Set-up

The server is on by default for every agent in Splitkick, and Splitkick repairs the configuration on start. Turn it off or on in Settings → MCP Servers → Enable Splitkick MCP. In the same place you add third-party MCP servers for your sessions, with overrides per project.

Limits

Two settings in Settings → MCP Servers keep agents from running away. They apply to the tools that create, change or start something; tools that only read have no limit.

• Max recursion depth (default 3, up to 10). A session that you start is level 0; a session that an agent starts is one level deeper than the session that started it. In a session deeper than the limit, these tools fail with an error.
• Spawn rate limit (per minute) (default 20). How often one caller can use each of these tools in a minute.

Archive and delete cannot be undone from the MCP server. The server tells agents to ask you before they archive or delete anything.

Tools

Orientation

Tool	What it does
get_current_context	The calling session, worktree and project, with splitkick:// links. Also the graph links of the session (serves, with graph and node titles), its agentId and its graphPermission (read, write or manage)
get_project_context	Project settings, linked projects, default branch and agent, worktree counts
get_usage	Plan usage for Claude, Codex and Grok, as in Settings → Usage. claude_accounts lists every Claude account you set up, with its label and usage

Projects

Tool	What it does
list_projects	All projects, with their kind: git, plain (a folder without git) or folder (a group in the sidebar)
add_project	Add a git repository on disk as a project, or any folder as a plain project with plain: true
clone_project	Clone a repository and add it
init_project	Make a new git repository and add it

Worktrees

Tool	What it does
list_worktrees, get_worktree	The worktrees of a project, or one worktree
create_worktree	A new worktree, also for a GitHub issue or PR or a Linear issue
create_worktree_from_existing_branch	A worktree for a branch that exists
import_worktree	Add a git worktree that is already on disk
rename_worktree, update_worktree_labels	Name and labels
get_worktree_changes, get_worktree_diff	Changed files and the diff, kept short
archive_worktree, unarchive_worktree, list_archived_worktrees	Archive and restore
delete_worktree, permanently_delete_worktree	Delete for good, with the branch

Sessions

Tool	What it does
list_all_sessions	Every session in every project, with status, labels, pins and stars, linked issue or PR, and a link
search_sessions	Find a session by topic, in names and messages
list_sessions	The sessions of one worktree
read_session_messages	The latest messages of a session; only prompts, without tool calls, or cut to a length
get_session_status	Idle, running, error, … and the latest run
create_session	A new session in a worktree, with the agent of your choice. Optional serves (graph links) and agentId (a saved agent; needs the manage permission). Takes over the empty "Session 1" of the worktree, see below
send_chat_message	Send a message; does not wait for the answer. Optional model (this turn only) and mode
cancel_session_run	Stop the running turn
set_session_model	Model and agent for the next turns
set_session_status	Set a manual status: idle, review, paused, completed or cancelled
add_follow_up	A note on the session's follow-up list; never sent on its own
move_session	Move a session to another worktree, also in another project
archive_session, unarchive_session	Archive and restore

create_session does not leave an unused "Session 1" next to the new session. When the worktree's "Session 1" is its only session that is not archived, was never used, has nothing waiting and was not taken by an earlier create_session, the tool takes it over and returns its id. Otherwise it adds a new session. Worktrees that you create in the app keep their "Session 1".

Git and pull requests

Tool	What it does
create_commit	Stage and commit with a generated message, push if asked
push_worktree	Push the branch
detect_open_pr	The open pull request of the branch, if there is one
create_pull_request	Open a pull request with a generated title and text
merge_pull_request	Merge it with the method the repository allows
run_review	Splitkick's AI review of the branch, with findings

The git tools give an error for a plain project (a project without git).

Issues and alerts

Tool	What it does
list_github_issues, list_github_prs	GitHub issues and pull requests of a project
list_security_issues, list_security_advisories	Dependabot alerts and security advisories
list_linear_issues	Linear issues of a project

Run and show

Tool	What it does
get_run_environments, start_run_environment	The dev servers Splitkick runs for a worktree (from splitkick.json)
open_in_browser	Show a page or file (HTML report, image, PDF) in the worktree's browser pane
add_artifact	Put a file or URL that the agent made (report, screenshot, HTML page, preview URL) into the Artifacts tab of a session. Again with the same target after a change, it adds a new version. See Artifacts

Graphs

Each graph tool needs a graph permission: read, write or manage. Every change needs a reason (why). See Graphs and agents.

Tool	Needs	What it does
list_graphs	read	All graphs: id, title, type, file, node count, open requests
read_graph	read	One graph: its nodes with blocked, progress and active, the sessions on each node, and the open requests
list_requests	read	The requests of a graph, newest first, with your answer once you gave it
list_changes	read	The change log of a graph, newest first: who changed what and why, and which changes you reverted. It marks nothing as seen
create_node	write	Add a node. A state other than proposed needs manage
update_node	write	Change the title, description, properties, relations or attachments of a node
set_node_state	write or manage	Set the state. proposed needs write; committed, done and dropped need manage
send_request	write	Ask you about a node: accept it as done, or pick an option in a decision
set_serves	read or manage	Change which nodes a session is linked to: replace the list, or add and remove. The own session needs no permission. Another session needs manage and a reason; you see the change in Review and can revert it
create_graph	manage	Create a graph file at a path, or inside a project's repo (default .splitkick/graphs/<slug>.yml)

Other Splitkick instances

Tool	What it does
list_instances	The instances you connected: name, id, URL, reachable, version
remote_tool	Run one of the tools above on one of those instances; the result names the instance

Both instances need 1.5.1. Ids and links in a result belong to that instance. Graph changes are refused there, and the other instance cannot send the call on. Details and an example: Work on your other Splitkick instances.

Browser for agents

Splitkick can also give your agents a web browser, as a second MCP server in every CLI you installed. Choose the engine in Settings → MCP Servers → Browser for agents. See Browser for agents.