AgentLoopKit MCP Server

AgentLoopKit

A drop-in engineering loop for coding agents.

Your coding agent can write code. AgentLoopKit gives it the repo habits reviewers need: scoped task contracts, safety rules, verification evidence, review summaries, and clean handoffs.

Vibe coding produces code. Agentic engineering produces auditable work.

The screenshots and terminal demo in this README are generated from committed sources in docs/assets/readme/ with Playwright and VHS.

What It Is

AgentLoopKit is a repo-level toolkit for developers using Codex, Claude Code, Cursor, OpenCode, Gemini CLI, GitHub Copilot CLI, and other coding agents.

It does not replace your agent, IDE, or CLI. It gives those tools a repeatable engineering loop inside the repository they are editing.

AgentLoopKit generates:

• AGENTS.md
• AGENTLOOP.md
• agentloop.config.json
• .agentloop/ with task templates, gates, policies, reports, handoffs, and agent instructions

Generated AGENTS.md also includes a specialist roster for product, CLI, template, verification, security, release, docs, compatibility, MCP, and repo-steward work. It helps future agent sessions split work cleanly without adding a background service.

Why It Exists

Coding agents often move fast but leave reviewers with weak evidence: unclear scope, missing tests, broad diffs, and vague summaries. AgentLoopKit gives the agent a repo-local contract:

1. Specify the task.
2. Constrain risk.
3. Plan the change.
4. Implement the smallest useful diff.
5. Verify with configured commands.
6. Review the diff.
7. Handoff with evidence.

Install

Use it with npx from the root of the repository you want to configure:

cd /path/to/your/repo
npx agentloopkit init --dry-run
npx agentloopkit init

init writes files into the current directory. Do not run it from ~ unless you intend to configure your home directory. --dry-run previews the file changes without writing them.

If you want AgentLoopKit for your local agent workflow but do not want to commit the generated harness:

npx agentloopkit init --local-only

--local-only writes the same files, then adds a marked block to this clone's .git/info/exclude for .agentloop/, AGENTS.md, AGENTLOOP.md, and agentloop.config.json. It does not edit .gitignore, global Git config, shell profiles, or files outside the current repo.

Pin the current version when you need repeatable CI or team setup:

npx --yes agentloopkit@0.26.2 version
npx --yes agentloopkit@0.26.2 init

Run the CLI after install:

npx agentloopkit doctor
npx agentloopkit create-task --title "Add settings page" --type feature
npx agentloopkit task list
npx agentloopkit task show .agentloop/tasks/2026-06-09-add-settings-page.md
npx agentloopkit task set .agentloop/tasks/2026-06-09-add-settings-page.md
npx agentloopkit task status .agentloop/tasks/2026-06-09-add-settings-page.md in-progress
npx agentloopkit status
npx agentloopkit next
npx agentloopkit verify
npx agentloopkit handoff
npx agentloopkit check-gates
npx agentloopkit check-gates --strict
npx agentloopkit report
npx agentloopkit badge
npx agentloopkit ci-summary
npx agentloopkit release-notes
npx agentloopkit npm-status
npx agentloopkit mcp-server
npx agentloopkit policy list
npx agentloopkit policy show security
npx agentloopkit policy status
npx agentloopkit task archive .agentloop/tasks/2026-06-09-add-settings-page.md
npx agentloopkit install-agent codex
npx agentloopkit install-agent all
npx agentloopkit completion zsh
npx agentloopkit completion powershell

The terminal demo is generated from committed VHS sources in this repository.

Pinned team usage:

pnpm add -D agentloopkit
pnpm agentloop init

Contributor usage:

git clone https://github.com/abhiyoheswaran1/agentloopkit
cd agentloopkit
pnpm install
pnpm test
pnpm build

CLI Commands

The package exposes two binaries:

agentloop init
agentloopkit init

Shell Completions

AgentLoopKit prints completion scripts to stdout. It does not edit .zshrc, .bashrc, fish config, PowerShell startup files, or other shell profile files.

Inspect a script before installing it:

agentloop completion zsh
agentloop completion bash
agentloop completion fish
agentloop completion powershell
agentloop completion pwsh

Example zsh setup:

mkdir -p ~/.zsh/completions
agentloop completion zsh > ~/.zsh/completions/_agentloop

Generated Files

.agentloop/
  manifest.json
  loops/
  gates/
  handoffs/
  agents/
  policies/
  tasks/
  reports/
  harness/
AGENTS.md
AGENTLOOP.md
agentloop.config.json

init appends to an existing AGENTS.md instead of overwriting it.

The package ships schema/agentloop.config.schema.json for editors and config validation. Generated configs use the GitHub raw schema URL for editor support; the CLI validates config locally and does not fetch that URL at runtime.

See docs/configuration.md for config fields and schema notes.

Fresh init also writes .agentloop/manifest.json so doctor can report which template generation created the local harness. See docs/template-migrations.md for manual upgrade guidance.

Agent Usage

Install instructions for the agent you use:

agentloop install-agent codex
agentloop install-agent claude-code
agentloop install-agent cursor
agentloop install-agent opencode
agentloop install-agent gemini-cli
agentloop install-agent github-copilot-cli

AgentLoopKit only writes safe repo-local Markdown instructions when exact third-party config conventions are uncertain.

Task Contracts

Create a contract before a coding session:

agentloop create-task --type feature --title "Add settings page" \
  --problem-statement "Users cannot manage account preferences" \
  --desired-outcome "Users can update settings from the app" \
  --likely-file src/settings \
  --forbidden-file migrations/ \
  --acceptance "Settings can be saved" \
  --verification "pnpm test" \
  --rollback "Remove the settings route"

Pin the contract when more than one task exists:

agentloop task list
agentloop task show .agentloop/tasks/2026-06-09-add-settings-page.md
agentloop task set .agentloop/tasks/2026-06-09-add-settings-page.md
agentloop task status .agentloop/tasks/2026-06-09-add-settings-page.md in-progress
agentloop task archive .agentloop/tasks/2026-06-09-add-settings-page.md
agentloop task current
agentloop task clear

task list --json gives agents a deterministic list with path, title, status, active, and modifiedAt. Listing tasks does not create or update .agentloop/state.json. create-task --json returns the created task path and Markdown content so agents do not need to parse the human success line. task show --json returns the selected task metadata and Markdown content without changing active state. task status --json updates only the - Status: line. Supported values are proposed, in-progress, blocked, review, and done. Status is not verification evidence; run agentloop verify before claiming completion. task archive --json moves one named Markdown contract into .agentloop/tasks/archive/, refuses to overwrite an existing archive file, and clears the active task pointer when it archives the active task. Archive after verification and handoff, not as a substitute for either.

Each contract records:

• problem statement
• desired outcome
• constraints
• non-goals
• assumptions
• likely files
• files not to touch
• acceptance criteria
• verification commands
• rollback notes

Verification Reports

agentloop verify reads agentloop.config.json, runs configured commands, captures output excerpts, and writes a report:

.agentloop/reports/YYYY-MM-DD-HH-mm-verification-report.md

Pass --task .agentloop/tasks/file.md to include task title, type, status, and path in the report. The path must point to a Markdown task contract inside the configured task directory. Invalid paths are reported as unavailable instead of being read.

It does not hide failures. Failed reports include a short failure summary with each failed command, exit code, and final useful output lines before the full command output. If long logs are truncated, the report keeps the first and last output so the final error stays visible. If no commands are configured, it writes a report saying nothing was verified.

When agentloop verify runs in GitHub Actions, the report records the workflow, event, ref, commit, run URL, and run attempt. Local reports stay quiet. AgentLoopKit does not read .env files or print arbitrary environment variables.

In monorepos, doctor warns on common workspace markers and suggests package-specific verification commands. Add package checks to the task contract when root commands do not cover the touched package, for example pnpm --filter web test or npm --workspace api test. AgentLoopKit records and runs configured commands; it does not infer package graphs or run workspace commands automatically.

doctor also reports potential risk files by category, such as migrations, auth, deployment, lockfiles, and env files. It lists path examples only. It does not read .env contents or claim to scan secrets. See docs/doctor-risk-files.md for category examples, limits, and reviewer actions.

Status

agentloop status gives agents and humans a quick local readout:

• active task contract, using agentloop task set when present
• newest task contract when no active task is pinned
• latest verification report
• working tree state
• configured and missing commands
• next suggested command

Use JSON output in scripts:

agentloop status --json

Use agentloop next when an agent or script only needs the next local command:

agentloop next
agentloop next --json

next uses the same decision rules as status. It does not run verification commands, create task state, call an LLM, make network requests, or read .env contents. When an active in-progress task exists, an older verification report does not count as current evidence for that task.

See docs/status.md for output fields and next-action rules.

Gate Checks

agentloop check-gates checks review evidence without running tests or calling an LLM. It looks for:

• a task contract
• a generated verification report
• a generated handoff summary
• repo harness files
• core safety policies
• git working tree context

Use JSON output in scripts:

agentloop check-gates --json
agentloop check-gates --strict

Warnings keep exit code 0 by default. Use --strict in CI when warning gates should fail the command.

doctor checks setup health. check-gates checks whether the current work session has the evidence reviewers expect.

See docs/check-gates.md for gate statuses and exit-code behavior.

HTML Reports

agentloop report writes a local static HTML page from the current task contract, latest verification report, latest handoff, git status, diff stats, and deterministic review summary:

agentloop report
agentloop report --json
agentloop report --out .agentloop/reports/review.html

The command does not run tests, call an LLM, fetch assets, read .env contents, or send data anywhere. It writes one local file under .agentloop/reports/ by default. Use it after verify and handoff when you want a browser-readable artifact for a PR or CI upload.

See docs/html-reports.md for inputs, output paths, and safety behavior.

Evidence Badges

agentloop badge writes a local SVG badge from existing evidence:

agentloop badge
agentloop badge --source gates
agentloop badge --json

The default badge reads the latest verification report and writes .agentloop/reports/agentloop-verification.svg. Gate badges read local gate status and write .agentloop/reports/agentloop-gates.svg. The command does not run tests, call a badge service, read .env, or upload anything.

See docs/badges.md for badge sources and CI usage.

CI Summaries

agentloop ci-summary reads allowlisted CI provenance fields and existing AgentLoop artifacts:

agentloop ci-summary
agentloop ci-summary --json
agentloop ci-summary --write

In GitHub Actions, GitLab CI, and Buildkite it reports provider, workflow or pipeline, event, ref, commit, and run URL when those allowlisted fields are present. Unsupported CI providers report generic CI when CI=true is present. The command does not call provider APIs, read secrets, upload files, run tests, or dump arbitrary environment variables.

Use --write when CI should upload a small Markdown summary alongside verification reports, HTML reports, badges, and handoffs. Verification consumers still read *-verification-report.md; CI summary artifacts do not replace verification evidence.

See docs/ci-summary.md.

Release Notes

agentloop release-notes drafts local release notes from the repository state:

agentloop release-notes
agentloop release-notes --from v0.19.0 --to HEAD
agentloop release-notes --release-version 0.26.2
agentloop release-notes --json
agentloop release-notes --write

The command reads local package metadata, changelog entries, git history, changed files, working tree status, the active task, the latest verification report, and the latest CI summary when those artifacts exist. It does not create tags, publish packages, call GitHub or npm APIs, read tokens, upload files, or rewrite changelogs.

Use it before creating a GitHub release so the release note draft includes the same evidence reviewers see in AgentLoop handoffs.

See docs/release-notes.md.

agentloop npm-status checks whether npm latest matches the local package version:

agentloop npm-status
agentloop npm-status --json
agentloop npm-status --expect-current
agentloop npm-status --registry-json npm-view.json

Use it after a publish attempt before saying npm has caught up. It runs npm view unless you pass captured registry JSON. It does not publish packages, create tags, read tokens, read .env files, upload files, or change package metadata.

See docs/npm-status.md.

MCP Server

AgentLoopKit can expose existing repo evidence to MCP clients through a read-only stdio server:

npx --yes agentloopkit@0.26.2 mcp-server

The server provides tools for status, next action, task contracts, active task, policies, latest verification report, and handoff summaries. It does not run verification commands, edit files, call external APIs, read .env contents, or upload data.

See docs/mcp.md for client configuration, tool names, and registry notes.

Policies

agentloop policy makes generated repo policy files visible from the CLI:

agentloop policy list
agentloop policy show security
agentloop policy status
agentloop policy list --json
agentloop policy status --json

The command reads Markdown files from .agentloop/policies/. policy status reports current, modified, missing, and extra files by comparing local Markdown with bundled templates. It does not enforce compliance, scan source code, fetch remote policy packs, or mutate policy files.

Local policy files are repo guidance. A modified policy can be intentional; review it like code instead of overwriting it to match the bundled template.

See docs/policies.md.

Security Reviews

Use security-review task contracts when a change touches auth, sessions, permissions, secrets, dependencies, deployment config, migrations, billing, or other sensitive areas.

AgentLoopKit helps reviewers see scope, checks run, checks skipped, risk notes, and rollback instructions. It does not scan code, certify compliance, or prove that code is secure.

See docs/security-review.md and examples/security-review/ for a copyable review task, verification report, and PR summary.

Dependency Upgrades

Use dependency-upgrade task contracts when an agent changes packages or lockfiles. AgentLoopKit records upgrade scope, lockfile impact, verification, risks, and rollback notes. It does not choose package versions, scan advisories, or call package registries.

See docs/dependency-upgrades.md and examples/dependency-upgrade/.

CI Recipes

Use agentloop check-gates --strict as a review-evidence gate in pull request CI. Use agentloop verify, agentloop handoff, agentloop report, agentloop badge, agentloop ci-summary --write, and agentloop release-notes --write in CI when you want evidence artifacts uploaded for reviewers.

CI-generated verification reports include GitHub Actions provenance when available, so reviewers can trace an artifact back to the workflow run that created it.

See docs/github-actions.md, examples/github-actions/, examples/gitlab-ci/, and examples/buildkite/ for copy-pasteable workflows. Pin agentloopkit@0.26.2 or a newer vetted release when reproducibility matters.

Other Install Channels

npm and npx remain the primary install path. AgentLoopKit also ships release assets and maintainer artifacts for:

• GitHub Releases: versioned tarballs for provenance and rollback.
• GitHub Action: a thin wrapper for agentloop commands in CI.
• Docker/GHCR: a minimal image that runs agentloop.
• MCP Registry: read-only server metadata for MCP clients once the matching npm package is published.

See docs/distribution-channels.md for current commands and maintainer release steps.

PR Summaries

agentloop handoff writes a reviewer-ready summary using deterministic inputs:

• git status
• git diff stats
• active task contract, or newest task when no active task is pinned
• latest verification report
• config settings

The summary groups changed files into review areas such as source, tests, docs, CI, config, AgentLoop artifacts, and risk-sensitive paths. It adds review-focus hints from file paths only.

It does not call an LLM.

Use agentloop summarize to preview the same output without writing a handoff file. agentloop summarize --write remains available for scripts.

Safety Principles

AgentLoopKit is intentionally boring:

• no postinstall scripts
• no telemetry
• no cloud backend
• no database
• no API keys
• no hidden network calls
• no reading .env contents
• no destructive actions during install

Env files are reported by path only.

Compared With Plain AGENTS.md

AGENTS.md gives instructions. AgentLoopKit adds process artifacts:

• task contracts
• gates
• policies
• verification reports
• PR summaries
• rollback notes

The result is easier to review because the work produces evidence, not only code.

Compared With Prompt Collections

Prompt collections help an agent respond better. AgentLoopKit gives the repository a standing workflow that any compatible agent can read and follow.

Compared With Superpowers

Superpowers is a broader agent methodology and skills system. AgentLoopKit focuses on repo-level engineering loops, task contracts, verification evidence, and handoff artifacts.

They can be complementary. You can use AgentLoopKit alongside Superpowers.

AgentLoopKit has no affiliation with Superpowers.

Examples

See examples/ for sample generated harnesses:

• examples/nextjs-app
• examples/node-api
• examples/python-service
• examples/docs-only
• examples/empty-repo
• examples/github-actions
• examples/security-review
• examples/dependency-upgrade

See docs/stack-recipes.md for Next.js, React/Vite, Remix, SvelteKit, Node API, Django, FastAPI, Python, docs-only, empty-repo, and monorepo verification recipes.

Roadmap

See ROADMAP.md.

Contributing

See CONTRIBUTING.md for local setup and PR expectations. Maintainers can use docs/contributor-playbook.md for copyable good-first issue examples.

License

MIT.