Agent Zero alternatives: choose by execution, memory and maintenance

Agent Zero is an active project, so "alternatives" here is a question of fit rather than rescue: the repository is not archived and shipped v2.12 on September 9, 2026, after v2.11 on August 27 and v2.10 on August 19. The useful move is to name the reason it does not fit, check that reason against Agent Zero's own documentation — two of the four common ones come out weaker or fail it outright — and only then pick from OpenHands Agent Canvas, goose, OpenClaw, OpenManus, AutoGPT or n8n.

First, the entity, because this query returns a crowded set of names. This page is about the Docker-based agent framework at github.com/agent0ai/agent-zero, whose LICENSE file carries MIT terms for "Agent Zero, s.r.o." while GitHub's own license field reports NOASSERTION. It is not the PyPI package called agent-zero, which is an outbound voice-agent framework; not aiming-lab/Agent0, an academic series on self-evolving agents whose repository name reads almost exactly like the agent0ai org; and not the Marvel character, which is why unrelated results sit on the same page. Never run pip install agent-zero expecting this project.

Filter by the reason, not by the shortlist

Four reasons come up repeatedly. Two of them hold, and two mostly dissolve once you read Agent Zero's own docs — which is worth knowing before you spend a weekend migrating.

Six alternatives, and the name each one now answers to

Several of these projects have changed org, name or cadence recently — OpenHands, goose and Open Interpreter all moved — so searching the old name pulls up stale install commands. All metadata below was read on September 21, 2026 and will move.

Two names to handle carefully. Open Interpreter is not the Python tool most tutorials describe: the repo moved to openinterpreter/openinterpreter, the README states it "is a fork of OpenAI's Codex", releases are tagged rust-v0.0.45 (September 20, 2026), and Python users are sent to the community fork endolith/open-interpreter — see Open Interpreter alternatives. And Space Agent, which Agent Zero's README calls "the related, more polished product direction", is not a newer Agent Zero: MIT, not archived, no announced discontinuation, but no release since v0.66 on April 30, 2026 and no commit on its default branch since June 8, 2026. Its BYOK support is unverified here, and on September 21, 2026 space-agent.ai/pricing redirected to a login page.

What actually migrates out, and what does not

This is the part a shortlist hides. No product in this set documents an Agent Zero importer, and the one place that publishes a provider list at all is OpenClaw's migrate command, whose built-in providers are Claude, Codex CLI and Hermes. Even there the list is not closed: the same doc says "Plugins can register additional providers." For every product here, the honest statement is that no Agent Zero importer was found, which is weaker than proving none exists.

Rollback first, migrate second. Agent Zero's VPS guide gives you the whole instance as one path — /opt/a0-instance/, holding .env, memory/ and work_dir/ — and its own backup line is tar -czvf a0-backup-$(date +%Y%m%d).tar.gz /opt/a0-instance/. On the plain docker run path the equivalent is the named volume a0_usr mounted at /a0/usr. Take that archive, keep the container stopped rather than removed, and move one workflow at a time.

On the destination side, four of these routes are documented below — Agent Zero itself, OpenHands, OpenManus and goose — and each has one gate worth knowing. Two more are noted rather than walked through: n8n's OpenAI credential carries a Base URL field its docs describe as useful for other providers offering an OpenAI-compatible API, and whether AutoGPT's hosted Platform accepts an arbitrary custom endpoint is not established on this page. OpenClaw's provider setup is not covered here at all. Agent Zero itself puts the key under External Services → Other OpenAI-compatible API keys and then asks you to select OpenAI Compatible as the provider, with the base URL living in a per-model-slot API URL field — and inside Docker, localhost in that URL means the container, not your host. OpenHands hides three fields behind an Advanced toggle (Custom Model with a provider prefix, Base URL, API Key), saves them as LLM Profiles its settings doc caps at 10 per account, and — its profile doc says — validates the configuration against the backend before it is persisted, though backends that answer 404 to the validation endpoint skip that check. OpenManus wants one [llm] TOML block — and its shipped example still defaults to an Anthropic base URL with a claude-3-7-sonnet-20250219 id, so copying it points you at the wrong place. goose uses environment variables, and its docs note that the endpoint must support tool calling:

# goose reaches any OpenAI-compatible endpoint through its built-in
# "OpenAI" provider. OPENAI_HOST is the root, with no trailing path.
export OPENAI_API_KEY="<your key>"
export OPENAI_HOST="https://api.kunavo.com"
# Appended to the host. Defaults to v1/chat/completions; some proxies
# mount chat/completions with no v1, and a 404 usually means this is wrong.
export OPENAI_BASE_PATH="v1/chat/completions"

Every configuration described here is a reading of that vendor's own documentation. Nothing on this page was runtime tested against any of these clients, so none of it is a compatibility claim, and a published setup guide is not a test either. See the OpenAI-compatible API guide for the general shape, and the goose, OpenHands and Open Interpreter setup pages for the configurations Kunavo publishes.

Same model, same rate: what actually changes is the loop

Switching clients does not reprice tokens. To make the comparison concrete, assume one agent run that accumulates 400,000 uncached input tokens across its turns and returns 30,000 output tokens. That shape is an assumption for illustration, not a measurement of Agent Zero or of any alternative — no loop here was metered. Rates come from the live Kunavo catalog, per million tokens.

The second reading matters more than the ranking. Agent Zero's model-presets guide gives every preset a separate utility model for lighter internal work. On the same assumptions, an extra 60,000 input and 6,000 output tokens of helper traffic costs $0.108 if it runs on Claude Sonnet 4.6 and $0.036 if you name Claude Haiku 4.5 for that slot instead. An alternative that exposes only one model role gives that lever back. Scale these by your own runs per day before treating any of it as a budget.

Kunavo serves no embedding, text-to-speech or speech-to-text model, and that is the one slot no gateway on this page fills: Agent Zero gives every preset a third, embedding model whose shipped default runs locally on CPU, and its bundled Kokoro TTS and Whisper STT plugins keep speech on the Docker bootstrap path as well. Budget that slot as local compute rather than as a line on any provider bill.

Kunavo's catalog amount is a billing floor rather than a ceiling: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tool APIs sit outside this example. The minimum top-up is $10 in prepaid credit with no subscription — a funding minimum, not a task fee. See billing details, or create an account when you have chosen a route.

When staying on Agent Zero is the right call

Stay if the container boundary is doing real work for you, if you are using the first-party plugins the alternatives would make you re-source — the listing on September 21, 2026 includes _browser, _desktop, _office, _orchestrator, _document_query and _time_travel — or if your reason for leaving was the model bill, which a new client will not change. Stay, too, if you are on the OAuth route Agent Zero added to an OpenAI Codex plan: that is a plan-backed path a custom endpoint cannot reach, and the README lists Gemini CLI and Claude Code as coming rather than shipped.

Move if Docker is a genuine blocker, if you need a channel that is not among the three bundled ones, or if your work has drifted from "agent with a computer" toward a repository or a scheduled pipeline — the point at which OpenHands, goose or n8n stop being alternatives and start being the right category. Move one workflow, keep the archive, and read the charge your provider account recorded for a bounded task before moving the rest. Agent Zero vs OpenClaw goes deeper on the two-product comparison, and OpenClaw alternatives, OpenHands vs Claude Code, goose vs Claude Code and n8n AI API cost cover the destinations individually.

FAQ

Checked on September 21, 2026 against primary sources only: the GitHub API for repository, release and commit metadata on every project named here; the registry document for the openclaw npm package; Agent Zero's README, installation, VPS-deployment, usage, memory, model-presets, MCP-configuration and contributing-skills docs, its bundled _memory, _email_integration and _whatsapp_integration plugin docs, its live plugins/ listing and its LLM API terms page; OpenHands's README, VM self-hosting doc, LLM-settings and LLM-profiles docs and pricing page; goose's providers and config-files docs and its April 2026 Foundation-move post; OpenClaw's README, first-run hardware FAQ and migrate doc; OpenManus's example config; AutoGPT's README and pricing page; n8n's pricing page and OpenAI credential doc; the PyPI record for the agent-zero package; and HTTP status checks on five pricing URLs. Kunavo token rates are read from the live catalog. Two items were corrected against earlier internal research and are stated as re-checked here: the bundled WhatsApp plugin, and the openclaw extended-stable tag, which read 2026.7.35 on this date. No client on this page was runtime tested, no loop was metered, and every dollar figure is illustrative token arithmetic rather than a measured task cost or a bill ceiling.