mcp agent

SHIP MCP server

Assign Linear, Jira or GitHub issues to an AI coding agent team. Its coding agents plan and build each change, run code review and tests, and open a pull request in your GitHub repository. With this MCP server, your assistant can file missions, follow them, send instructions and read the proof of each run.

SHIP runs an agentic SDLC as a headless AI software factory. You operate the whole mission from your coding agent or AI assistant. Start it, send instructions, read its code review and test reports, and approve the pull request. A Planner, a Builder, a Reviewer and a Tester each do one stage, and your CI is a gate between them. Each role runs the coding agent and the model that you choose: Claude Code, Codex, OpenCode, Kilo Code or the pi coding agent.

Endpoint

https://mcp.letsship.ai/mcp

The transport is Streamable HTTP. The server is stateless, so a client can reconnect at any time and lose nothing.

Sign in

OAuth 2.1. A client that supports MCP authorization finds the sign-in automatically. If a request needs a token and has no valid token, the server answers 401 with a WWW-Authenticate header. The header points to https://mcp.letsship.ai/.well-known/oauth-protected-resource/mcp. SHIP supports dynamic client registration and client ID metadata documents, so you do not have to register a client first. You approve the sign-in in the SHIP console.

Personal access token. A client without OAuth sends Authorization: Bearer pat_.... Create a token at https://letsship.ai/settings/tokens. Bind it to one project, so that it can reach only that project.

Install

Claude Code

claude mcp add --transport http ship https://mcp.letsship.ai/mcp

Claude and ChatGPT. Add a custom connector with the URL https://mcp.letsship.ai/mcp. Sign in when the app asks.

Cursor. Use the button above, or add this to .cursor/mcp.json:

{
  "mcpServers": {
    "ship": {
      "url": "https://mcp.letsship.ai/mcp"
    }
  }
}

VS Code. Use the buttons above, or add this to .vscode/mcp.json:

{
  "servers": {
    "ship": {
      "type": "http",
      "url": "https://mcp.letsship.ai/mcp"
    }
  }
}

Codex. Add this to ~/.codex/config.toml:

[mcp_servers.ship]
url = "https://mcp.letsship.ai/mcp"

Grok. Open grok.com/connectors. Select New Connector, then Custom. Paste the URL https://mcp.letsship.ai/mcp. Then sign in.

xAI API. Pass the URL as a remote MCP tool, with a token in headers:

{
  "type": "mcp",
  "server_url": "https://mcp.letsship.ai/mcp",
  "server_label": "ship",
  "headers": {
    "Authorization": "Bearer pat_your_token_here"
  }
}

A client that supports only stdio connects through mcp-remote:

{
  "mcpServers": {
    "ship": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.letsship.ai/mcp"
      ]
    }
  }
}

With a token. Add the header to any of the HTTP entries above:

{
  "mcpServers": {
    "ship": {
      "url": "https://mcp.letsship.ai/mcp",
      "headers": {
        "Authorization": "Bearer pat_your_token_here"
      }
    }
  }
}

Tools

Each tool declares readOnlyHint, destructiveHint and openWorldHint. With these hints, your client can run a read-only tool immediately. It can also ask you before it runs a tool that removes or replaces something, or opens a payment page. The scopes of the token decide what each tool can do, and a refusal names the scope that the token does not have.

[!IMPORTANT] start_mission, create_mission and run_scheduled_task start a mission, and each mission start costs one credit. Messages to an active mission are free.

Missions

Tool	Kind	What it does
start_mission	reaches outside SHIP	Give an issue that is already in your tracker to the agent team. By default, the mission starts at the Planner. With a plan, it starts at the Builder. With an open pull request, it starts at the Reviewer or the Tester
create_mission	reaches outside SHIP	Create the issue and start the mission in one call, for work that has no issue yet
send_mission_instruction	reaches outside SHIP	Tell an active mission what to change, or record a note for the agent that runs next
inspect_mission	reads	The full record of a mission: its stages, verdicts, provenance and session logs
list_missions	reads	The missions of your projects, with what each mission does now
get_mission	reads	One mission: its stage, its review and test results, its pull request, its agent runs and the run with the proof. With view: "full", also its events and the outcome of each stage
pause_mission	writes	Hold a mission before its next stage. The stage that runs now finishes
resume_mission	reaches outside SHIP	Continue a paused mission. When a checkpoint holds the mission, this approves the result of that stage
stop_mission	removes	Stop a mission and every agent run of it. The pull request stays open
approve_mission_ceiling	reaches outside SHIP	Let a mission continue after it stops at its credit limit. The mission then uses more credits
merge_mission	removes	Merge the pull request of a mission, with the merge method of the repository. On a merge conflict, SHIP starts the Builder to fix the conflict and merges later
decline_mission	removes	Close the pull request of a mission without a merge
set_agents_override	replaces	Change the harness, model, prompt or checkpoint of one role for one mission. The override goes on top of ship.yml and does not change the file
get_agents_override	reads	Read the override that applies to a mission now
clear_agents_override	removes	Remove the override, so the mission uses the project's configuration again
set_mission_auto_merge	reaches outside SHIP	Turn auto-merge on or off for one mission. For that mission, the setting replaces the auto-merge of the project
archive_mission	writes	Hide a mission from the board and from list_missions. Only an owner or an admin can archive a mission
unarchive_mission	writes	Show an archived mission on the board again
set_mission_disposition	writes	Record why a mission ended, for example a test or work that another mission replaced, or clear the value. The delivery statistics use this value

Projects and proof

Tool	Kind	What it does
list_projects	reads	The projects that the token can reach
get_project	reads	The configuration and connections of one project
update_project	reaches outside SHIP	Change the name, status, auto-merge, issue code, default sandbox or connections of a project
list_artifacts	reads	The proof that a run produced, such as test traces and screenshots
get_artifact	reads	One artifact, with a link to open it

Scheduled tasks

Tool	Kind	What it does
list_scheduled_tasks	reads	The missions that a project runs on a schedule
get_scheduled_task	reads	One scheduled task, with its schedule and prompt
create_scheduled_task	writes	Add a mission that runs on a schedule
update_scheduled_task	replaces	Change the schedule or prompt of a scheduled task, or disable it
run_scheduled_task	reaches outside SHIP	Run a scheduled task now, outside its schedule
list_scheduled_task_runs	reads	What past runs did, and what they cost
list_scheduled_task_templates	reads	Templates that you can copy to start a scheduled task
delete_scheduled_task	removes	Delete a scheduled task and its run history. Missions that the task started continue

MCP servers for your agents

Tool	Kind	What it does
list_mcp_servers	reads	The servers that the MCP file of a project declares, each with its binding and status
list_mcp_bindings	reads	Each binding in the projects of your organization, so you can see where each credential is used
bind_mcp_server	replaces	Choose the agents, transport, approved tools and credential for one server
unbind_mcp_server	removes	Remove a binding and delete its credential
refresh_mcp_server_tools	removes	Read the current tool list of the server, so that you can approve tools
start_mcp_server_oauth	reaches outside SHIP	Get a sign-in link for a remote server, for a person to open in their browser
set_mcp_config_path	writes	Set a different MCP file for a project, such as .cursor/mcp.json

Agent runs

Tool	Kind	What it does
list_agent_runs	reads	Agent runs, live and finished, for one mission, one role or one status
get_agent_run	reads	The detail and resource use of one run
get_agent_run_logs	reads	What an agent did inside a run
watch_agent_run	reads	Wait on one run and follow its events, until the run ends or its mission moves to another stage
stop_agent_run	removes	Stop the agent in one run before it finishes

Preview environments

Tool	Kind	What it does
list_environments	reads	Preview environments that SHIP deployed for open pull requests
get_environment	reads	One preview environment, with its URL
destroy_environment	removes	Remove a preview environment early

Credits and notices

Tool	Kind	What it does
get_billing	reads	The credit balance, the credit limit for each mission and the recent use. Read it before you send work
list_billing_ledger	reads	Where the credits went, with the newest entry first
list_credit_offers	reads	The credit packs and the monthly plans, with their prices
start_credit_checkout	reaches outside SHIP	Open a payment page for a credit pack, a monthly plan or more credits on a plan, and return its URL
update_team_subscription	removes	Change the monthly plan from the next renewal, or cancel it at the end of the month that you paid for
request_more_credits	reaches outside SHIP	Ask for more credits. The answer says who has the request
get_notices	reads	What needs attention in the workspace, such as a tracker that you must connect again
dismiss_notice	writes	Close a notice that you can close, such as a low balance

Custom sandboxes

Tool	Kind	What it does
list_sandbox_templates	reads	The sandboxes of your organization, with the active version and the latest build of each
list_sandbox_starters	reads	Sandbox definitions for Python, Java, Go, Rust and Ruby that you can start from
save_sandbox_template	writes	Create a sandbox or change one. A change makes a new version and starts its build
list_sandbox_template_versions	reads	Each saved version of a sandbox, with the active version marked
activate_sandbox_template_version	writes	Make a version the one that missions use from their next run
build_sandbox_template	writes	Build the active version now
get_sandbox_build_log	reads	One page of a build log
list_sandbox_registries	reads	The private package registries that sandboxes install from. A secret is never returned
save_sandbox_registry	replaces	Store or replace the credential for a registry. The credential never enters a sandbox
get_sandbox_capacity	reads	How many agents your missions run at once, against the limit of your organization
delete_sandbox_template	removes	Delete a sandbox. The API refuses a sandbox that an agent or the default of a project uses
delete_sandbox_registry	removes	Remove the credential of a private registry

Platform operators

Tool	Kind	What it does
list_orgs	reads	Each organization on the platform. Only a SHIP operator token can call it

Identity

Tool	Kind	What it does
whoami	reads	The principal behind the token and the scopes of the token
get_setup_progress	reads	What the workspace has set up so far and the next step, as the console's setup checklist shows it
healthz	reads	The status, environment and version of the API

Example: file a fix and answer the review

1. Call list_projects to get the project id.
2. Call create_mission to file the issue and start the mission. Write the acceptance criteria as checkboxes, because the Tester uses them to test the work.
3. The agents plan the change, build it and open a pull request. Then they wait for your CI, review the change, deploy a preview and test it. This takes tens of minutes, and your assistant does not have to wait.
4. Call inspect_mission to read the stage and the verdicts. To disagree with a review, call send_mission_instruction with kind: "follow_up". The Builder then works on the pull request again and follows your instruction.

To send feedback, use send_mission_instruction. Do not call start_mission again, because a restart does not take a message.

Troubleshooting

What you see	What to do
401 and no sign-in window	Your client does not support MCP OAuth. Send a personal access token in the Authorization header.
A tool refuses with a scope name	The token does not have that scope. Create a token that has the scope, or ask an owner for a role that allows it.
create_mission refuses with credits_exhausted	The workspace has no credits left. get_billing shows the balance.
The tool list is empty, or each tool call gets 401	The client did not finish sign-in. Sign in again from the MCP menu of your client, or remove the server and add it again.
whoami names another project	The token is bound to that project. Use a token for the project you want.

Skills for your coding agent

The SHIP skills teach a coding agent to hand work to SHIP from the terminal. To install them, see letsship/skills or run npx skills add https://letsship.ai.

Learn more

• What an agentic SDLC is, from issue to verified pull request
• What an AI software factory is, and how to build one
• SWE-in-a-team, a coding agent benchmark of 260 missions
• Run the pi coding agent headless
• Send a pull request that your coding agent wrote for review and test

Links

• Docs: https://letsship.ai/docs/mcp
• Registry entry: ai.letsship/ship on https://registry.modelcontextprotocol.io
• A2A agent: https://letsship.ai/docs/a2a
• Privacy: https://letsship.ai/privacy - Terms: https://letsship.ai/terms
• Support: https://letsship.ai/support
• Security reports: support [at] letsship.ai

License

The files in this repository are under the MIT license. The SHIP service is under the SHIP terms.