MartinLoop

Name: MartinLoop
Author: Keesan

Puts budgets, verifier gates and run receipts around coding-agent jobs started from an MCP host

By Keesan (individual) JavaScript Local · stdio Developer tools AI and ML Git and code hosting

GitHub npm Website

43 popularity score ★ 296 ↓ 473/wk

Connect your client

Local (npm)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio Add to Goose

Opens the app, which asks before it adds the server.

Or add it by command or config

Claude CodeTerminal

claude mcp add martinloop -- npx -y @martinloop/mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "martinloop": {
      "command": "npx",
      "args": [
        "-y",
        "@martinloop/mcp"
      ]
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "martinloop": {
      "command": "npx",
      "args": [
        "-y",
        "@martinloop/mcp"
      ]
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "martinloop": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@martinloop/mcp"
      ]
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.martinloop]
command = "npx"
args = ["-y", "@martinloop/mcp"]

Gemini CLITerminal

gemini mcp add --scope user martinloop npx -- -y @martinloop/mcp

What it does

MartinLoop is a control layer for coding agents, and @martinloop/mcp is its MCP server. You give it an objective, a verifier command such as `npm test`, a spend ceiling and an attempt cap. It then hands the work to an agent CLI (Claude Code, Codex or Gemini CLI) and re-runs the verifier after each attempt. It stops when the checks pass, when the budget or attempt limit is reached, or when a policy check blocks the run. Every run ends with a VERIFIED, STOPPED or NEEDS REVIEW outcome and a stored receipt.

The server walks a host through set steps: martin_doctor, martin_plan and martin_preflight, then martin_run. In server.ts, martin_run checks the recorded workflow first and fails with policy_blocked, listing the missing steps, if the host skipped any. Further tools pause, continue or cancel a live run, read saved runs, attempts and verifier evidence, grade a run, check a PR body against the evidence, and draft or open a GitHub pull request from the receipts. Run records are kept as local JSONL files, each signed with HMAC, so later edits to a record can be detected.

In v0.8.10 the default full profile lists 24 tools, while the MCP tool reference names only 21. Missing from the reference are martin_estimate, which gives a read-only cost and route estimate, and two Arcade tools defined in arcade/capabilities.ts. martin_arcade opens a view-only panel in hosts that support MCP Apps, and martin_arcade_status reads the evidence behind that panel. The server also offers 26 martin:// resources and 16 prompts. It talks over stdio by default. Pass --http and it serves Streamable HTTP on 127.0.0.1:3033 at /mcp instead (--host, --port and --path change that), and --tool-profile directory trims the tool list.

What it needs

Runtime	Node.js 20+
Runs on	macOS, Windows, Linux
Claude Code, Codex or Gemini CLI	An agent CLI signed in to its provider, for martin_run. Its model usage is billed by that provider.
GitHub	FreeOnly for martin_create_pr with execute=true, through a signed-in gh CLI.
Also	An agent CLI on PATH: claude, codex or gemini GitHub CLI (gh) for opening pull requests A git workspace with a verifier command such as a test suite

npx needs Node.js 20 or later. The package brings no agent CLI: install and sign in to Claude Code, Codex or Gemini CLI yourself before calling martin_run. Install and sign in to the GitHub CLI (gh) before calling martin_create_pr with execute=true. Runs only finish as VERIFIED when you give a verifier command that your workspace can run.

Limits

Limit A VERIFIED outcome covers only the checks your verifier runs. The README says it does not mean the code is bug-free or safe to merge.
Cost For Codex, cost and token figures are labelled estimated unless the host reports real usage.
Not covered martin_run is refused with policy_blocked until the earlier workflow steps for the same task are on record.
Limit The MCP tool reference leaves out martin_estimate, martin_arcade and martin_arcade_status, which v0.8.10 registers.
Not covered The minimal, diagnostic, full-local and github-review profiles come from `martin-loop mcp print-config`. The server takes only `--tool-profile full|directory`, full by default.
Limit Hosted sync of run evidence needs MARTIN_API_TOKEN and MARTIN_TELEMETRY_ENDPOINT, set through the separate martin-loop CLI.

Who it suits

Developers who already use Claude Code, Codex or Gemini CLI and want hard spend and attempt limits, a required test pass, and an auditable record around each delegated task, started from inside their MCP host. It writes code through the agent it starts and can open pull requests, so it is a poor fit if you only want read-only help. The `directory` tool profile or a host allow-list narrows what it exposes. It does not take the place of CI, code review or the coding agent itself. It is not published by Anthropic, OpenAI or Google.

Tools 24

19 only read, 5 change things. A ? marks an optional parameter.

martin_doctor: Checks the local setup, agent CLIs, workspace and run store before work startsworkingDirectory?: stringrunsDir?: stringengine?: enum
martin_estimate: Estimates cost and route for an objective before anything is spentobjective: stringengine?: enumbudgetUsd?: numberfileScope?: arrayworkingDirectory?: string
martin_plan: Sets out scope, verification and budget for a task before a runobjective: stringworkingDirectory?: stringcontext?: stringpolicyPack?: enumverificationPlan?: arrayallowedPaths?: arraydeniedPaths?: arraymaxUsd?: numbermaxIterations?: integermaxTokens?: integermaxMinutes?: integermaxFilesChanged?: integermaxCommands?: integer
martin_preflight: Checks a proposed run contract (verifier, scope, budget) before executionobjective: stringworkingDirectory?: stringengine?: enummodel?: stringcontext?: stringpolicyPack?: enummaxUsd?: numbermaxIterations?: integermaxTokens?: integermaxMinutes?: integermaxFilesChanged?: integermaxCommands?: integerverificationPlan?: arrayallowedPaths?: arraydeniedPaths?: arrayworkspaceId?: stringprojectId?: string
martin_run changes: Starts a coding run through the chosen agent CLI, held to budget and verifier gatesobjective: stringworkingDirectory?: stringengine?: enummodel?: stringfallbackEngines?: arraymaxUsd?: numbermaxIterations?: integermaxTokens?: integerverifyTimeoutMs?: integerproviderExecutionTimeoutMs?: integerverificationPlan?: arrayallowedPaths?: arraydeniedPaths?: arrayworkspaceId?: stringprojectId?: string
martin_status: Reports the state of the live or most recent runloopJson?: stringfile?: stringloopId?: stringrunsDir?: stringlatest?: boolean
martin_logs: Reads recent events, ledger entries and control receipts for one runfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanlimit?: integer
martin_pause changes: Asks a running loop to pause at the next safe point and writes a control receiptfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanreason?: stringrequestedBy?: string
martin_continue changes: Resumes a paused loop from its saved contextfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanreason?: stringrequestedBy?: string
martin_cancel changes: Stops a running loop and writes a control receiptfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanreason?: stringrequestedBy?: string
martin_inspect: Opens a saved loop recordfile?: stringrunsDir?: string
martin_list_runs: Lists saved run records, with filters by status, model and daterunsDir?: stringlimit?: integerstatus?: stringlifecycleState?: stringadapterId?: stringmodel?: stringupdatedAfter?: string
martin_triage_runs: Ranks saved runs by urgency and missing evidencerunsDir?: stringlimit?: integerstatus?: stringlifecycleState?: stringadapterId?: stringmodel?: stringupdatedAfter?: stringincludeHealthy?: boolean
martin_get_run: Fetches one saved runfile?: stringloopId?: stringrunsDir?: stringlatest?: boolean
martin_get_attempt: Fetches a single attempt within a runfile?: stringloopId?: stringrunsDir?: stringattemptIndex?: integer
martin_get_verification_results: Returns the stored verifier evidence for a runfile?: stringloopId?: stringrunsDir?: string
martin_run_dossier: Builds a compact dossier for a runfile?: stringloopId?: stringrunsDir?: stringlatest?: boolean
martin_dossier: Returns the run dossier in JSON, Markdown or GitHub PR formfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanformat?: enum
martin_eval: Grades a run on verifier coverage, regression risk and reviewabilityfile?: stringloopId?: stringrunsDir?: stringlatest?: boolean
martin_pr_summary: Drafts a pull-request summary from a run's receiptsfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanformat?: enum
martin_create_pr changes: Previews a GitHub pull request, or opens one through gh when execute is truefile?: stringloopId?: stringrunsDir?: stringlatest?: booleanformat?: enumtitle?: stringbase?: stringexecute?: boolean
martin_review_pr: Checks a PR body against the run's evidence and returns a verdict with findingsfile?: stringloopId?: stringrunsDir?: stringlatest?: booleanformat?: enumprBody?: string
martin_arcade: Opens the view-only Arcade panel in hosts that support MCP Apps
martin_arcade_status: Reads the run evidence the Arcade panel shows, for one run or the latestloopId?: string

Resources 27

martin://server/health: Server health and environment readiness
martin://runs/recent: Previews of recent runs
martin://runs/triage: Runs ranked by how much attention they need
martin://runs/latest: Full dossier of the latest run
martin://runs/latest/summary: Short summary of the latest run
martin://runs/latest/proof-card: Small Markdown receipt for the latest run
martin://runs/latest/budget-status: Budget, token and stop-condition snapshot for the latest run
martin://runs/latest/verifier-evidence: Verifier evidence and warnings for the latest run
martin://runs/latest/rollback-evidence: Rollback and artifact evidence for the latest run
martin://verifiers/results: Latest verifier results
martin://policies/current: Local policy packs and the one suggested for this repo
martin://repo/risk-map: Sensitive parts of the repo and the suggested policy pack
martin://agent/next-step: The single next tool, prompt or resource to use
martin://agent/governance-status: Whether work is governed, budget left and runs without receipts
martin://agent/mode-status: Current working mode and how to switch it
martin://agent/memory-summary: Stored preferences, consent signals and budget patterns gathered over time
martin://guides/mcp-usage: Suggested workflow over MCP
martin://guides/agent-start: Short start guide for agents
martin://guides/command-map: Which tool or resource to pick for each job
martin://guides/ide-onboarding: Setup guide for IDE hosts
martin://guides/operating-rules: Rules on when agents must call the tools before working
martin://guides/publish-readiness: Checklist for MCP publish-readiness reviews
martin://runs/{loopId}: One saved loop record
martin://runs/{loopId}/dossier: Dossier for one saved run
martin://runs/{loopId}/attempts/{attemptIndex}: One attempt and its verification event
martin://runs/{loopId}/verification: Verification events for one run
ui://martinloop/arcade/index.html: The Arcade panel page, for hosts that support MCP Apps

Prompts 16

martin_start: Starts a governed workflow with minimal context
martin_preflight: Prepares the preflight payload before a run
martin_triage: Sorts run-store failures and picks what to inspect next
martin_resume: Picks up from an earlier run using its evidence
martin_prove: Builds a receipt from the latest or a chosen run
martin_release_check: Release-readiness review using the server's evidence
martin_governed_coding_kickoff: Frames a coding request before preflight
martin_debug_failed_run: Diagnoses a failed run from its saved attempts and checks
martin_publish_readiness_review: Publish-readiness review of the MCP package
martin_triage_run_store: Ranks saved runs and picks one to debug
safe_bug_fix: Small scoped bug fix through doctor, plan, preflight, run and dossier
write_tests_first: Failing tests first, then a small fix
small_refactor: Keeps a refactor small, path-scoped and checked by the verifier
security_review: Reviews a risky change using risk, scope and verifier evidence
pr_review: PR review checklist built from dossier and eval evidence
release_check: Release check based on run evidence

Popularity

43 popularity score, on our list (the bar is 36)

★ 296 ↓ 473/wk

GitHub stars, whole repository	296	21.4 points
Stars in 30 days	+0	0 points
npm downloads, last week	473	21.4 points

It lives in the folder packages/mcp of keesan12/martin-loop, which holds more than this server. Its stars are the whole repository's, so in the score they never count for more than its own downloads.

Measured 2026-11-07. How the score works.

Where it is published

Code	keesan12/martin-loop/packages/mcp ★ 296, the whole repository
npm	@martinloop/mcp 0.9.0, 2026-11-03

What it can reach

Runs	On your machine
Reaches	Files on your machine, Runs shell commands, Your code, The internet
Changes things	Yes: it can write, send, create or delete
Sign-in	None
Licence	Apache-2.0 (npm)
Last code change	2026-11-03

Runs locally over stdio. With --http it binds to 127.0.0.1:3033 unless --host is given. martin_run starts an agent CLI (Claude Code, Codex or Gemini CLI) in your workspace. That agent edits files and the server runs your verifier commands, so code changes and shell commands happen on your machine. Model calls go out under the agent's own sign-in. The README says policy checks block unsafe verifier commands, risky path changes and secret-like task inputs. Run records are local JSONL files signed with a per-store HMAC key. martin_create_pr calls the gh CLI with your GitHub sign-in when execute=true, and only previews the PR otherwise. Hosted sync of evidence is off unless configured through the martin-loop CLI. The server itself needs no API key.

Over time

GitHub stars, whole repository 296 −113 (−28 %)

414296

9 Oct 20267 Nov 2026

npm downloads a week 473 +40 (+9 %)

1.0k117

18 May 20267 Nov 2026

Before the site measured it, read from npm's daily download counts.

Popularity score 43 −12 (−21 %)

5434

9 Oct 20267 Nov 2026

What is rising across the directory

Sources

Page checked 2026-10-30.

Badge for your README

Shows this server's rank in its category, and stars for a server that runs on your machine. It updates every hour.

MarkdownREADME.md

[![MartinLoop on EveryMCPServer](https://everymcpserver.com/badge/martin-loop.svg)](https://everymcpserver.com/servers/martin-loop)

Connect your client

Add in one click

Or add it by command or config

What it does

What it needs

Limits

Who it suits

Tools 24

Resources 27

Prompts 16

Similar servers

Claudexor

agent-harnesses MCP server

Ruflo

ReevesAgents

autocontext

Quill