Kombinat

Turn markdown files in your repo into scheduled or webhook-triggered AI routines, running on your own machine and existing subscription.

Source repo
Danilqa/kombinat
Stars
★ 17
Last updated
today
License
MIT
Primary language
TypeScript

At a glance

How it runs
CLISelf-hosted serviceWeb app
Works with
Universal · cross-platformCodex · Claude CodeOpenAI API · Claude API (Partial support)
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Node 24Claude Code CLI (logged in) or Codex CLI (logged in)macOS (launchd) or Linux (systemd user service)Shell / CLINetwork accessLocal filesystem
Typical use
A solo developer uses an always-on Mac mini or Linux box to work through agent-labelled issues overnight and open pull requests.
Not a fit if
  • Teams without an always-on Mac/Linux machine to host the daemon
  • Anyone needing built-in login auth for the dashboard
  • Organizations that forbid unattended agents using local logins
Website
kombinat.tech ↗ set on GitHub
Maintenance
Commits in the last 30 dayslast commit 2026-11-20latest release v0.8.0 (2026-11-20)

What does this agent do, and when should you use it?

Kombinat is an open-source local scheduler daemon that treats each Markdown file under routines/ as a routine: the frontmatter sets the cron schedule, working directory, agent, model, timeout and permissions, while the body is the prompt. Every 30 seconds the daemon checks which routines are due and runs them through claude -p, codex exec, or any custom CLI command, serving a dashboard at http://localhost:7433 and a terminal UI via pnpm kombinat tui. launchd on macOS or a systemd user service on Linux keeps the daemon alive across reboots, with logs and run history in ~/.kombinat/. Merging is deploying: the daemon runs git pull on main every 5 minutes and installs dependencies when the lockfile changes. It runs on the Claude Code or Codex subscription you already pay for, and strips API key variables from the agent environment.

Reads YAML frontmatter from each Markdown file in routines/ (name, description, schedule, cwd, agent, model, fallbackModel, effort, timeoutMinutes, permissionMode, maxBudgetUsd, before/after hooks, enabled, webhook) and uses the body as the prompt. The daemon scans for due routines every KOMBINAT_TICK_SECONDS (default 30), runs at most KOMBINAT_MAX_CONCURRENT (default 3) at once, never overlaps a routine with itself, and stops a run after timeoutMinutes (default 55); after sleep it catches up the latest missed slot within 7 days. It invokes claude -p or codex exec, or runs the CLI given in command (prompt on stdin, output is the result); the before hook runs first as a shell command and the after hook runs once regardless of outcome with the finished run on stdin. A routine with webhook: true can also be started by HTTP requests from apps such as Linear or GitHub, with the request body injected into the prompt, an optional webhookGate command validating and filtering requests, and a queue for requests arriving mid-run. Full transcripts are viewable in the browser dashboard, the TUI and on a phone, and a chat continues any run; pnpm kombinat readme regenerates the routines table in README.md.

  1. A solo developer uses an always-on Mac mini or Linux box to work through agent-labelled issues overnight and open pull requests.
  2. A team wants routines reviewed like code: add one Markdown file in a branch, merge it, and it is live within minutes.
  3. An engineer needs Claude Code or Codex to run where their existing logins, skills, MCP servers and browser control already are, instead of rebuilding that in a CI container.
  4. An on-call rotation is wired to Linear or GitHub webhooks so incident checks and post-release monitoring fire on events, not just cron.
  5. A home user schedules personal routines on a Mac to check weather, find discounts or watch a website for changes.
  6. A tinkerer who would otherwise hand-roll cron plus claude -p plus logging plus a dashboard wants a small, self-hosted package instead.

Key setup facts

Pulled from the install and usage notes below. The project README has the current values.

Install / run
  • npx kombinat init my-routines
  • npx kombinat serve
  • pnpm add kombinat
  • npx kombinat connect admin@mac-mini # dashboard in the browser
Environment variables
KOMBINAT_PORT KOMBINAT_STATE_DIR KOMBINAT_ROUTINES_DIR
Config files
~/.kombinat
Runtime
Node.js 24
Install notes cover
macOS

How do you install or deploy this agent?

You need Node 24 and a logged-in agent: Claude Code (claude auth status) or Codex (codex login). Scaffold a folder with example routines:

npx kombinat init my-routines

In an existing repo, run npx kombinat init from its root (or pnpm dlx kombinat init for pnpm projects); it never overwrites files. Then start it:

cd my-routines
npx kombinat serve

Install it as a background daemon (launchd or systemd user service):

pnpm add kombinat
pnpm kombinat install

There is one daemon per user, so running install in another folder moves it there.

How do you use this agent?

Create a Markdown file in routines/ with frontmatter for schedule and settings and the prompt as the body:

---
name: fix-e2e-tests
description: Finds the failing e2e test and fixes it
schedule: '30 11 * * *'
cwd: ~/projects/app
agent: claude
model: claude-opus-5-5
fallbackModel: claude-sonnet-5-5
effort: high
timeoutMinutes: 55
---

1. Open the latest failed e2e run on GitHub ...

Changes apply on the next check, no restart needed. Core commands (run from the folder Kombinat runs in; use npx kombinat with npm, yarn kombinat with Yarn):

pnpm kombinat list           # routines, next run, last status
pnpm kombinat run <name>     # run a routine now, in the foreground
pnpm kombinat validate       # check all routine files
pnpm kombinat tui            # dashboard in the terminal, works over ssh
pnpm kombinat status         # daemon and scheduler state
pnpm kombinat readme         # update the routines table in README.md
pnpm kombinat uninstall      # stop and remove the daemon

Remote access over SSH or Tailscale Serve:

npx kombinat connect admin@mac-mini         # dashboard in the browser
npx kombinat connect admin@mac-mini --tui   # dashboard in the terminal
tailscale serve --bg 7433                   # expose the dashboard across your tailnet

Configuration variables, set before install and saved into the launchd/systemd config: KOMBINAT_PORT (7433), KOMBINAT_STATE_DIR (~/.kombinat), KOMBINAT_ROUTINES_DIR (./routines), KOMBINAT_MAX_CONCURRENT (3), KOMBINAT_TICK_SECONDS (30), KOMBINAT_PULL_MINUTES (5, 0 disables), KOMBINAT_PULL_BRANCH (main), KOMBINAT_KEEP_AWAKE, CLAUDE_BIN, CODEX_BIN.

What are this agent's strengths and limitations?

Pros
  • Routines are Markdown files that live in git and go through review; the daemon's periodic git pull on main makes merging the deploy step.
  • Reuses your existing Claude Code or Codex login, CLAUDE.md, skills and MCP servers, so there is no extra API bill.
  • Ships the scheduling plumbing out of the box: transcripts, timeouts, concurrency caps, no overlapping runs, sleep catch-up, webhooks, dashboard and TUI.
  • Supports claude, codex and any command CLI, self-hosted on macOS launchd or Linux systemd with no cloud and no database, only two dependencies.
  • Per-routine control over permissionMode, maxBudgetUsd, model fallback chain and before/after hooks, all in a reviewable file.
Limitations
  • Requires an always-on Mac or Linux machine that stays logged in; runs do not happen while the machine sleeps or the lid is closed (on macOS it stays awake only on power).
  • The dashboard listens on localhost with no login, so other users of the machine can reach it; once exposed via Tailscale Serve, anyone who can reach the port can run any prompt with a routine's agent, folder and permissions.
  • You must supply and log in to a Claude Code or Codex subscription, and the README cautions teams to check their plan's terms before sharing one login.
  • The daemon pulls main every 5 minutes and installs dependencies when the lockfile changes, so weak branch protection becomes a supply-chain and accidental-merge risk.
  • Nobody answers agent questions, so poorly written prompts can burn runs; classifying 'Nothing to report' depends on instructing the agent to start its answer with 'No' or 'Nothing'.

How does this agent compare with similar options?

The README explicitly compares Kombinat against four alternatives: running Claude Code by hand (the work never happens on time), Claude Code's own Routines (not stored in git or reviewed, cannot use Codex or other CLI agents, not on a machine you control), a plain cron job calling claude -p (missing run history, transcripts, timeouts, overlap protection, sleep catch-up, dashboard and webhooks), and a scheduled CI workflow (agents do not run where your logins, skills and MCP servers already are).

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Kombinat This agent 54 · Major gaps CLIFree + model costs ★ 17 today TypeScript Codex · Claude Code
kwant 68 · Some gaps — ★ 681 1d ago Rust Codex · Claude Code · OpenAI API · Claude API
HermesClaw 52 · Major gaps Self-hosted serviceFree + model costs ★ 765 1mo ago Python —
Brigade — Enterprise-grade personal intelligence 62 · Some gaps CLIFree + model costs ★ 11k 8d ago TypeScript ChatGPT · Codex · Claude Code · OpenAI API · Claude API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
54/ 100 5-point scale 2.7 / 5
This review covers the source as of Nov 18, 2026. The project has shipped 2 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 12/29
Reliability 6/14
Adaptability 14/18
Convention 10/18
Effectiveness 9/13
Verifiability 3/8
Why each dimension lost points
Trust12 / 29 · 2.1/5

README documents default permission modes (Claude denies approval-needing actions, Codex workspace-write), removal of API key env vars, no telemetry, and localhost-only dashboard — positive signals. However the dashboard has no login and any local user can send arbitrary prompts, bypassPermissions/danger-full-access unlock everything, the daemon auto-runs git pull and installs dependencies every 5 minutes, and webhook bodies can steer agents; these high-risk defaults lack confirmation, so least_privilege=2, user_confirmation=1, sensitive_data_handling=1. Dependency security shows only a CI import allowlist for the packed bundle, no audit or vulnerability policy, so 1. Rollback is limited to uninstall and retained run history, with no version pinning or state recovery, so 1. Source attribution is just LICENSE and author name; publisher is unverified, so 1.

Reliability6 / 14 · 2.1/5

README and AGENTS.md agree on architecture, commands and scheduling semantics (no overlap, max 3 concurrent, 7-day catch-up), so self_consistency=2. Availability depends on Node 24, pnpm 11, external CLI agents and subscription logins with no documented fallback, so dependency_availability=1. Failure messaging appears only as TUI crash handling and a validate command; no user-facing error taxonomy, so failure_messages=1.

Adaptability14 / 18 · 3.9/5

README covers schedules, webhooks, manual runs, Mac/Linux, SSH/Tailscale, work and home scenarios, so audience_and_scenarios=3. Capability boundaries (agent types, permission modes, timeouts, budget cap, concurrency) are tabulated but the command mode can run anything and boundaries stay thin, so 2. Trigger precision has 5-field cron, manual, webhook and webhookGate, but timezone/DST details are absent, so 2. Environment fit documents launchd/systemd, lingering and keep-awake, but only Mac/Linux, so 2.

Convention10 / 18 · 2.8/5

README is well structured with complete install steps (init/serve/install, pnpm/npm/yarn variants), so information_architecture=2 and install_notes=2. The package renamed from @danilqa/kombinat to kombinat and the old launchd job breaks, so naming_stability=1. Example routines and a comparison table are ample, so examples_and_faq=2. Known limitations (sleep, single-user daemon, plan terms, unauthenticated dashboard) are stated, so 2. LICENSE is full MIT text, so 3. Versioning relies on a dashboard card and changelog link with no in-repo CHANGELOG, so versioning_changelog=1. Maintenance responsibility is only an author name and CI, with an unverified publisher and no governance or support commitment, so 1.

Effectiveness9 / 13 · 3.5/5

Output usability includes dashboard, TUI, run history, Inbox classification and readme generation, so output_usability=2. Marginal value over cron+claude -p (run history, timeouts, dedupe, webhooks, review flow) is concretely contrasted, so 2. Cost benefit emphasizes reusing an existing subscription, no API bill and two dependencies, but requires an always-on machine and carries plan-terms risk, so 2.

Verifiability3 / 8 · 1.9/5

Most README claims map to AGENTS.md, CI workflows and source snippets (TUI polling, e2e fixtures, packaging checks), but there is no per-claim citation or reproducible evidence, so claim_traceability=1. Cross-source corroboration covers only some claims (CI matches AGENTS.md); core security and scheduling behavior lacks source backing, so 1. Facts and marketing phrasing ("live within 5 minutes", "no API bill") are mixed without separating inference from fact, so fact_inference_separation=1.

Risks and how to mitigate them
  • The dashboard has no login and relies on localhost isolation; other local users can read transcripts and drive agents with arbitrary prompts, and remote exposure needs SSH/Tailscale ACLs you configure yourself.
  • The daemon auto-runs git pull every 5 minutes and installs dependencies on lockfile changes, so a compromised main branch means arbitrary code execution; branch protection and review are mandatory.
  • bypassPermissions and danger-full-access grant the agent the full rights of your user, and routines that ingest external text (issues, webhooks) are exposed to prompt injection.
  • Publisher identity is unverified, and the package rename (@danilqa/kombinat to kombinat) breaks the old launchd job; confirm provenance and migration steps before upgrading.
  • Requires Node 24, pnpm 11, external CLI agents and subscription logins with no documented fallback; this is a static review with no code or tests executed.
Evidence confidence: Low Reviewed Nov 18, 2026 Reviewed revision 00a2ab4296bd New commits since this review; the score may not cover them
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/kombinat.svg)](https://followagents.com/en/agents/kombinat)
<a href="https://followagents.com/en/agents/kombinat"><img src="https://followagents.com/badge/kombinat.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v0.8.0
    Release notes excerpt
    - Run a routine once later from the Run now menu: in 1 hour, tonight, tomorrow, next Monday, in a week or at a time you pick. The routine page lists these runs with Cancel, and the Calendar and the terminal dashboard show them too (#127) - A Claude run that hits the 5-hour or weekly limit now waits for the reset and continues the same session, with Continue now and Stop waiting on its page. Set onUsageLimit: fail to keep the old behaviour (#119) …
    View on GitHub ↗
  2. v0.7.0
    Release notes excerpt
    - An Inbox lists the runs that ended since you last looked, grouped into Needs a look, Has findings and Nothing to report. The daemon keeps what you have read, so every device and the terminal dashboard agree (#101) - The header shows how much of your Claude plan is left, taken from the runs' own output with no login (#100) - The dashboard and the terminal dashboard update the moment a run starts or ends, and running times count up live (#92, #104) - Search inside a run's transcript (#102, #116) …
    View on GitHub ↗
  3. v0.6.0
    Release notes excerpt
    - Chat with a run's agent from its run page. A message resumes the run's session, and the scheduler, Run now and webhooks wait while it replies (#88) - Webhooks of one conversation, such as the messages of a Linear agent session, continue the same run: the webhookGate exits with 79 and prints the conversation's key and the message (#98) - Run now has a "Change prompt and run" option that runs once with an edited prompt and leaves the routine file untouched (#90) …
    View on GitHub ↗

FAQ

Do runs cost extra API money?
No. claude and codex routines use the subscription login you already pay for, and Kombinat removes API key variables from their environment; the software itself is MIT-licensed and free. The optional maxBudgetUsd only applies to claude and caps a single run.
What happens when a scheduled run is missed?
If the machine was asleep, each routine runs once for its latest missed time, up to 7 days back. A routine never runs twice at once, and a run still in progress at its next time is skipped; at most 3 routines run concurrently.
Is it safe to expose the dashboard to other devices?
By default it listens on localhost only with no login. Over an SSH tunnel (npx kombinat connect) or Tailscale Serve, anyone who can reach the port can read transcripts, start routines, message agents, or run any prompt they write with a routine's agent, folder and permissions, so limit access with Tailscale access rules.
Which triggers and agents are supported?
Triggers are a 5-field cron in local time, manual, or webhook: true with HTTP requests from apps such as Linear and GitHub (optionally filtered by a webhookGate command). Agents are claude (claude -p), codex (codex exec), or command (any CLI receiving the prompt on stdin).
Do I need to restart after adding or editing a routine?
No. Changes apply on the next check (default every 30 seconds); merged changes on main are pulled by the daemon every 5 minutes, dnd dependencies are reinstalled and the daemon restarts when the lockfile changes without interrupting running routines.
Is Kombinat open source?
Yes. The source code is on GitHub (Danilqa/kombinat) under the MIT licence.
Is Kombinat still maintained?
Yes. The latest commit on GitHub is from 2026-11-20, within the last 30 days. The latest release is v0.8.0 (2026-11-20).
Who makes Kombinat?
Kombinat is published on GitHub by the account Danilqa. FollowAgents has not verified which person or company is behind that account.
What is Kombinat's official website?
The project's GitHub repository lists https://kombinat.tech as its website. That link is set by the repository owner; the source code itself is on GitHub (Danilqa/kombinat).
Was this page helpful?
View on GitHub ↗ Install ↓

Host ecosystems

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts