Best Self-Hosted Coding Agent Daemons in 2026: Kodama vs Paseo vs Official CLIs
Quick answer: If you want Claude Code or Codex working in the background but refuse to trust a vendor-installed, self-updating persistent process, self-hosted agent daemons are the transparent alternative. Kodama is the most focused option for running Claude Code/Codex sessions as managed background daemons you own; Paseo takes a broader workspace-orchestration angle; the official CLIs remain the most polished but ship with exactly the trust trade-offs (default daemons, silent auto-updates) that pushed many developers to look elsewhere. Choose Kodama for single-developer transparency, Paseo for multi-agent workflows, and official CLIs only if you have audited and accepted their background behavior. ---Why this category exists now
In October 2026, a Hacker News thread about Codex CLI installing a persistent, self-updating background daemon by default โ without clearly telling users โ pulled in hundreds of upvotes and a long tail of "this was on my machine too?" replies. The controversy was not about the feature itself. Running a coding agent in the background is genuinely useful: warm startup, queued tasks, agents that keep working while you sleep. The problem was the delivery model: an opt-out daemon with an auto-update channel means the binary running on your machine tomorrow is not the binary you reviewed today. The same week, projects like Kodama (a self-hosted daemon manager for Claude Code and Codex) were being shown on HN from the opposite direction: identical capability, but the daemon is code you can read, a process you started, and a service you can stop with one command. That is the trust model this roundup compares.What to evaluate in an agent daemon
Before the picks, the four questions worth asking of any tool in this category: 1. Who controls the process? Did you start it, and can you stop it with a documented, one-step command?2. How does it update? Is there an auto-update channel, and can you see a changelog before new code runs?
3. What does it observe? Which directories, credentials, and network calls can the daemon touch while idle?
4. Is it auditable? Open source, logged actions, and a way to replay what the agent actually did. ---
Kodama โ best for transparent single-developer daemons
Kodama is a self-hosted daemon manager built specifically around the Claude Code / Codex workflow. Instead of letting the vendor's CLI decide when a background process exists, you define agent services yourself: which repo, which harness, what triggers a run, and where the logs go. Strengths- The daemon is yours: you install it, you start it, you stop it. No hidden LaunchAgents or systemd units.
- Open source, so the "what is it doing while idle" question has an answer you can read.
- Fits the mental model developers already have from pm2 or systemd: named services, restart policies, log tails.
- Young project; expect sharp edges and breaking changes between releases.
- You own the operations burden โ if the agent hangs at 3 a.m., that is your pager.
Paseo โ best for multi-agent workspace orchestration
Paseo takes a wider swing: rather than managing one vendor's agent, it aims to be a workspace where multiple agents run background tasks across repos, with a UI for steering and reviewing their work. Strengths- Multi-agent by design: parallel tasks, separate worktrees, review-oriented output.
- Self-hosted option keeps credentials and code inside your own infrastructure.
- Broader scope means more moving parts to deploy and maintain than a single-purpose daemon manager.
- Still an early-stage project; documentation and stability lag the ambition.
Official CLIs (Codex CLI, Claude Code) โ polished, but audit first
The vendor CLIs remain the most capable harnesses, and their background features exist for real reasons: daemon pre-warming makes interactive sessions feel instant, and managed auto-updates keep users on supported versions. If you stay on them โ and for many teams that is the right call โ do it with eyes open:- Inventory the background processes they install (
launchctl list | grep -i codexon macOS,systemctl --user list-unitson Linux). - Decide explicitly whether the daemon and auto-update channel are acceptable on a work machine.
- If not, disable the daemon and keep the CLI in on-demand mode โ you lose warm startup, you gain an audit point before every version change.
Comparison table
| Tool | Who controls the daemon | Update model | Multi-agent | Maturity |
|---|---|---|---|---|
| Kodama | You (self-hosted service manager) | Manual / you pin versions | Single-agent sessions | Early |
| Paseo | You (self-hosted control plane) | Manual / you pin versions | Yes | Early |
| Official CLIs | Vendor default (opt-out) | Auto-update by default | Varies | High |
How to choose
- Solo developer, privacy-first: start with Kodama. It solves the exact problem the HN thread surfaced โ background agents without the trust hole โ with the least infrastructure.
- Team running parallel agent tasks: evaluate Paseo, but pilot it in an isolated environment first.
- Staying on official CLIs: that is fine โ just treat the installation like vendor due diligence. Audit the daemons, decide on auto-updates deliberately, and re-check quarterly.