OpenClaw alternative: a self-hosted AI agent you don't have to babysit
You tried OpenClaw, or read enough about it, and decided it is not for you. Fair. This is the whole field on one page: eight alternatives, what each is good at, where each falls short, and a short guide for choosing. Avelina is one of the eight, and we say plainly when it is the wrong pick.
Short answer. People leave OpenClaw for three reasons that recur in third-party coverage: maintenance, security around its skill registry, and token costs. There are three ways out: switch to another open-source runtime (Hermes Agent, ZeroClaw, NanoClaw, Nanobot, PicoClaw, Letta Code), pay a managed OpenClaw host to run the same software for you, or buy a turnkey commercial agent (Avelina AI). Which one fits depends on how much you want to maintain. For the head-to-head with one product, see Avelina vs OpenClaw.
Why people leave OpenClaw
First, fairness: OpenClaw is MIT-licensed, free, and has about 391.5k GitHub stars as of October 10, 2026 (repo). If the three complaints below don't bother you, it is still an excellent choice.
1. Maintenance. It is a self-hosted gateway that needs Node 24.16+ or 26.1+ (README) and ships on calendar-based versions across stable, beta, extended-stable and dev channels (update docs); the latest, v2026.9.9, came out on October 8 (GitHub release). You are the sysadmin. Its early renames were messy too: its creator reportedly said "we messed up the migration" (Hackernoon, secondary source).
2. Security. Researchers flagged 341 malicious skills on ClawHub, the community registry, 335 of them from one coordinated campaign (summary); later audits report higher counts that vary by auditor, and Unit 42 says evasive malicious skills persisted from February to May 2026 (Unit 42). CVE-2026-25253 (CVSS 8.8, affects versions before 2026.1.29) is a one-click RCE per the vendor advisory GHSA-g8p2-7wf7-98mq; NVD record: https://nvd.nist.gov/vuln/detail/CVE-2026-25253.
3. Token costs. Vendor blogs estimate roughly $5-30 a month on cheap models and $500+ on Opus-class defaults (Haimaker), and say long conversation history and multi-step tool calls inflate bills (ClawdHost). These are vendor estimates, not measurements.
The field side by side
Facts come from each project's README or docs and the GitHub API on October 10, 2026. "Not stated" means we did not find it, not that it is absent.
| Option | Self-hosted | Licence | Memory | Persona | Cron / sub-agents | Telegram | Price | Setup |
|---|---|---|---|---|---|---|---|---|
| OpenClaw (reference) | Yes | MIT | Markdown files in workspace | SOUL.md, IDENTITY.md | Both | Yes, one of 20+ channels | Free, you pay the model | Install script, npm, Docker, Nix |
| Hermes Agent | Yes | MIT | Agent-curated, FTS5 search, Honcho | /personality command | Both | Yes | Free, you pay the model | One-line installer |
| ZeroClaw | Yes | MIT OR Apache-2.0 | SQLite plus embeddings | Not stated | Cron yes; sub-agents not stated | Yes | Free | Binary, install script, Docker |
| NanoClaw | Yes | MIT | Per-agent CLAUDE.md | Per-agent CLAUDE.md | Both | Yes, via an add-on skill | Free | Script; needs Docker |
| Nanobot | Yes | MIT | Session history plus Dream long-term memory | Not stated | Both | Yes | Free | pip install |
| PicoClaw | Yes | MIT | JSONL store | Not stated | Both | Yes | Free | Prebuilt binary, make, Docker |
| Letta Code | Yes (plus Letta Cloud) | Apache-2.0 | Memory blocks, git-tracked MemFS | --personality flag | Both | Yes | Open source free; Cloud not checked | npm install |
| Managed OpenClaw host | No, the host runs it | OpenClaw is MIT; host terms not checked | As OpenClaw | As OpenClaw | As OpenClaw | As OpenClaw | Not checked | Host sets it up |
| Avelina AI | Yes, your VPS | Commercial, not open source | Four typed layers in local databases | SOUL.md, biography, emotional state | Both | Yes, native | EUR 280 once, or EUR 480 guided, plus VPS and Claude | Guided, about 30 min, from Telegram |
The alternatives, one by one
Hermes Agent
The closest open-source rival: MIT, 252k stars, version 0.21.6 as of October 8 (repo). It has a learning loop, agent-curated memory, natural-language cron, isolated sub-agents and a Telegram gateway with an allow-list, plus hermes claw migrate to import from OpenClaw. The catch: it is still a runtime you run and update yourself, and the version number is still 0.x.
ZeroClaw
A single Rust binary with 30+ channels, SQLite memory, cron and a default "supervised" autonomy level (repo). Good if you want something compact and fast. One flag: we did not find sub-agents or a persona system in its README. It is dual-licensed MIT OR Apache-2.0 (both LICENSE files are in the repo; GitHub's API shows only Apache-2.0).
NanoClaw
A small OpenClaw alternative that runs each agent in its own Docker container on the Claude Agent SDK; MIT, latest release October 9 (repo). Isolation is the point, and agents can spawn teammates with their own memory. Docker is required, and Telegram is added through a skill rather than built in.
Nanobot
A lightweight Python framework from HKUDS with a web UI, MCP, multi-agent delegation, cron and Telegram (repo). python -m pip install nanobot-ai is a one-line install if you already live in Python. A persona layer was not in the README sections we read.
PicoClaw
Sipeed's Go assistant built for tiny hardware; its README claims under 10 MB of RAM and admits recent builds use 10-20 MB (repo). Telegram, cron and sub-agents are there. The last release we saw dates from July 3, 2026, and persona support was not found.
Letta Code
The memory specialist: memory blocks, MemFS that you can sync to your own GitHub repo, and "dreaming" (repo). Apache-2.0, with Slack, Telegram and Discord channels. Note that the older letta repository keeps the retired V1 API server on an archive branch; current development is in letta-code, so start there. It installs as an npm command-line tool, and Letta Cloud pricing was not checked.
A managed OpenClaw host
Third parties such as MyClaw and ClawdHost run OpenClaw for you. That removes the install and update chore; it does not by itself change the software, so if the host runs stock OpenClaw the skill-registry question comes with it. Where your memory and keys live is the host's decision. We did not check their prices or data terms, so do.
Avelina AI
A commercial, closed-source agent that runs on your own VPS and lives in Telegram, with a personality layer, four typed memory layers, sub-agents, cron and signed releases. It is installed in about 30 minutes through a guided flow, for a one-time EUR 280 (self-install) or EUR 480 (with a one-hour session); you also pay for your own VPS ($5-15 a month) and your own Claude subscription (pricing). Trade-offs: you cannot audit or fork the code, it works with Claude only, and Telegram is its interface rather than one of 30 channels. More in the agent overview.
Which one if...
- You like tinkering and want the biggest ecosystem: stay with OpenClaw, or try Hermes Agent.
- Your hardware is tiny: PicoClaw or ZeroClaw.
- You want agents isolated in containers: NanoClaw.
- You write Python: Nanobot.
- Memory is your research interest: Letta Code.
- You want OpenClaw but not the chores: a managed host, after reading its data terms.
- You don't want to maintain anything, but want your data on your server: Avelina. Read is self-hosting worth it and the VPS guide first.
- You don't care about self-hosting: a cloud agent will do; see self-hosted vs cloud AI.
FAQ
Is there a free OpenClaw alternative?
Yes. Hermes Agent, ZeroClaw, NanoClaw, Nanobot, PicoClaw and Letta Code are all free to download, and their licences are open source. "Free" still excludes your server and your model API or subscription.
Which alternative is easiest to install?
Hermes and OpenClaw both start with a one-line installer, and Nanobot with pip install. Among the options here, Avelina is the one sold with a guided install, about 30 minutes from Telegram. "Easy" should include the updates that come after.
Which has the best memory?
We found no head-to-head benchmark. By design, Letta Code is the most memory-centric (memory blocks, git-tracked MemFS), Hermes curates memory itself, OpenClaw keeps plain Markdown you can read, and Avelina separates facts, events, corrections and inferences into four layers on your own server.
Can I keep my OpenClaw memory files?
Mostly yes. OpenClaw stores memory as Markdown (MEMORY.md, daily notes, USER.md) and persona in SOUL.md, so you can copy them. Hermes has an import command; we did not verify what it carries over. Avelina lists no OpenClaw importer, so you would move key facts and persona by hand.
Is a managed OpenClaw host a real alternative?
For maintenance, yes. For security and token cost, the software is the same, so the same questions apply, and your data sits on the host's servers rather than yours.