AI & Agents

Your coding agent can run, fix, and review your tests

The Lastest MCP server (@lastest/mcp-server) implements the Model Context Protocol over stdio. Any MCP-compatible agent - Claude Code, Cursor, Cline, Windsurf, or Claude Desktop - gets first-class tools to list, create, run, heal, and review tests in your Lastest instance.

Try Lastest free →Read the docs

What you get

Three things this integration does - and what they mean for your team.

01

stdio transport, Bearer auth

The MCP server speaks stdio and authenticates against /api/v1/* with a Bearer token. The same key works for the VSCode extension. Revoke any time from Settings → Agents & API Access.

02

~20 tools across 7 categories

Repos, tests, functional areas, builds, diffs, jobs, coverage. Each tool returns a structured ToolResponse with status, summary, optional actionRequired, and details - readable by any agent without extra prompting.

03

Drop-in for any MCP client

Same JSON config shape works in Claude Desktop, Cursor (~/.cursor/mcp.json), Cline, Windsurf, and any future MCP-compatible client. One install, every agent.

How it works

Data flow, configuration, and result routing - visualised.

Diagram 1 · Agent ↔ MCP topology

Where the MCP server fits

AGENT PROCESSMCP SERVER (stdio)LASTESTClaude Codecursor · windsurf · clineclaude_desktop_config.jsonapi key@lastest/mcp-servernpx -y/api/v1/*RESTDBSHA256-hashed keystdiobearerHTTPS
Agent process (Claude Code, Cursor, etc.) spawns @lastest/mcp-server over stdio. The MCP server is a thin shim that translates MCP tool calls into authenticated REST calls against /api/v1/* on your Lastest instance.

Diagram 2 · Tool surface

20 tools across 7 categories

MCP server20 toolsReposlist_repos · get_repoTestslist · get · create · update · delete · healAreaslist · create · list_tests_by_areaBuildscreate · get · list · get_runDiffsget · approve · reject · approve_allJobsget_active_jobs · get_jobCoverageget_coverage
Repositories, tests, functional areas, builds & runs, diffs, jobs, coverage. Each tool maps to one REST endpoint and returns a ToolResponse the agent can read directly.

Diagram 3 · Authoring loop

How an agent creates, runs, and approves a test

create_testagent → MCPactionRequiredrun_testqueue buildactionRequiredget_diffon diffactionRequiredapprove_diffor reject_diffeach ToolResponse hints the next step → no human in the chain
Agent: create_test → run_test → on diff, get_diff → approve_diff (or reject_diff). Each step returns a structured response with the next actionRequired so the agent can chain without human turn-taking.

Configure

Drop-in config for MCP Server

Claude Desktop / Cursor / Windsurf config

{
  "mcpServers": {
    "lastest": {
      "command": "npx",
      "args": [
        "-y",
        "@lastest/mcp-server@latest",
        "--url", "https://your-lastest-instance",
        "--api-key", "YOUR_API_KEY"
      ]
    }
  }
}

Related

Frequently asked questions about Lastest + MCP Server

Do I need MCP to use Lastest?
No. MCP is an optional integration for AI agents. The web UI, CLI, GitHub Action, and VSCode extension all work without it.
Which clients are supported?
Any MCP-compatible client. Tested with Claude Code, Claude Desktop, Cursor, Cline, and Windsurf. The same JSON config shape works in all of them.
Where is the API key stored?
In your MCP client config - read by the MCP server at startup and forwarded as a Bearer token. Lastest stores the key SHA256-hashed in the database; only the SHA is checked at request time.
Can the agent approve baselines on its own?
Yes - approve_diff, reject_diff, and approve_all_diffs are exposed. You decide who has approval scope by issuing API keys with the right roles.
How do I see what tools the server exposes?
Run claude mcp list (Claude Code) or use the agent's built-in "list tools" command. The server publishes the current tool surface on startup.

Other integrations

Keep exploring