OpenHands alternatives: choose by why you are leaving

Most searches for OpenHands alternatives are aimed at the wrong target: the OpenHands coding agent is alive, and its terminal CLI is the part that stopped. github.com/OpenHands/OpenHands is MIT-licensed, not archived, was pushed on September 21, 2026 and published v1.20.0 on September 17, 2026 (GitHub API, checked September 21, 2026). The separate OpenHands-CLI repository carries a banner near the top of its README reading "This project is no longer actively maintained" and recommends Agent Canvas instead. So the shortlist forks immediately: if you want the agent and its browser surface, staying is a live option; if what you actually used was a terminal agent, you need a different client.

One name check before the list, because it reverses the answer. AI4Bharat/OpenHands is a sign-language recognition toolkit, Apache-2.0, last pushed March 15, 2023, and its GitHub description reads "No longer actively maintained". Articles and AI answers saying OpenHands is abandoned are often quoting that repository. The coding agent is OpenHands/OpenHands, MIT, and it began life as OpenDevin — OpenDevin/OpenDevin still redirects there.

Start from the reason you are leaving

OpenHands' pricing page on September 21, 2026 lists three rows: Open Source (Local) at Free, with Max Daily Conversations "Unlimited" and 1 user; Individual (SaaS) at Free, with Max Daily Conversations 10 and 1 user; and Enterprise at Custom pricing. That page publishes no Enterprise number, and this page does not estimate one. Most exits trace to one of five specific frictions rather than to the agent itself.

Status of every candidate, re-checked rather than remembered

Three names on the usual listicle moved or died during 2026, and an older roundup will send you somewhere that no longer exists. Every row below was read from the GitHub API on September 21, 2026.

A free client is not a zero-cost client

Most rows above carry an OSI licence and no licence fee; Claude Code is proprietary and Crush is source-available under FSL-1.1-MIT, and no subscription price was fetched for any client on this page. None of that is the number you budget anyway. Two lines are: the machine the agent and its sandbox run on, and the tokens. The first is where OpenHands differs structurally from a terminal agent — it wants a backend. Its own VM guide says "Ubuntu 24.04 LTS with 2 vCPU and 4 GB RAM is enough for a single user", and the Modal recipe prices that same shape at about $3.40/day. Those are the host's charges for that documented recipe, not OpenHands' charge and not a figure that transfers to another host.

The token line follows you to every client, which is why it is worth sizing before choosing. Assume one busy agent session that sends 300,000 uncached input tokens and receives 20,000 output tokens — OpenHands' own model documentation warns that it "will issue many prompts to the LLM you configure". Rates are live Kunavo catalog prices per million tokens.

This is illustrative token arithmetic, not a measured task cost and not a bill ceiling: your session shape decides the answer, and a cheap model that needs three attempts can cost more than one that needs a single pass. Cheapest listed rate and lowest cost to finish the task are different questions. Kunavo's catalog amount is a billing floor rather than a cap — 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 tools sit outside this example, and the minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription. See billing details.

What migrates, what does not, and how to roll back

Repository context is the portable half. OpenHands reads AGENTS.md at the repository root, also recognises CLAUDE.md and GEMINI.md, and its skills documentation recommends .agents/skills/<name>/SKILL.md precisely because that location "follows the Agent Skills standard and makes skills portable across compatible agent tools". MCP servers carry over in shape too: OpenHands' MCP settings page supports SSE, Streamable HTTP and stdio — it links stdio and Streamable HTTP to the 2025-06-18 MCP specification and SSE to the older 2024-11-05 one, so do not assume the SSE half is current everywhere — and its configuration lives in the UI or under a [mcp] section of config.toml.

Hooks are less stranded than they look. OpenHands' hooks documentation says the .openhands/hooks.json format "is compatible with Claude Code hooks, so you can reuse hook scripts across both tools" — a claim about those two tools, not about every client on this list. What does not travel is the rest of the OpenHands-specific half: .openhands/setup.sh is OpenHands' own repository-setup contract; conversation history exports as a Markdown or HTML file for reading, and OpenHands' own documentation describes no import path for it; and the openhands/-prefixed LLM key only talks to that vendor's proxy.

Expect the endpoint shape to change as well. OpenHands wants a provider-prefixed Custom Model behind the Advanced toggle in Settings > LLM, and for an OpenAI-compatible server its model-configuration page prescribes openai/<served-model-id> read from the server's GET /v1/models, with a base URL that includes the version path. OpenCode expresses the same key as a provider block instead, so the migration is a rewrite rather than a copy:

{
  "provider": {
    "kunavo": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://api.kunavo.com/v1",
        "apiKey": "{env:KUNAVO_API_KEY}"
      },
      "models": { "claude-sonnet-4-6": { "name": "Claude Sonnet 4.6" } }
    }
  }
}

Kunavo serves no embedding model, no text-to-speech model and no speech-to-text model. That matters for one knob: OpenHands lists LLM_EMBEDDING_MODEL among the settings its UI cannot set, to be supplied as an environment variable or in config.toml — point it at a provider that serves that step, or leave it unset.

A trial that produces an answer in a day: keep OpenHands installed and its profile saved, run one bounded task from the same starting commit in both clients, then compare the accepted diff, the review effort and the usage your provider account actually recorded. Rolling back means keeping ~/.openhands/settings.json and any config.toml untouched while you test — worth doing deliberately, because OpenHands' own installation page warns that upgrading from a CLI version before release 1.0.0 requires redoing the settings setup since the configuration format changed. Note also that OpenHands' LLM settings page says you can save up to 10 LLM profiles per account, so prune before you add a test one.

Set up the next client and check the first charge

Kunavo publishes configuration references for several of the clients above, including OpenHands, OpenCode and Cline. Those are published setup guides, not compatibility tests: none of these clients has been runtime-tested here against Kunavo's endpoint, and every behavioural claim on this page was read from each project's own documentation and repository. Keep a working route available while you switch, run one scoped task, then read the charge your account recorded for it. Create a Kunavo account when you are ready to fund a key.

If your decision is narrower than a full switch, OpenHands vs OpenCode and OpenHands vs Claude Code compare two of these head to head, and OpenAI-compatible API covers the endpoint shape the OpenAI-compatible clients on this list expect — Claude Code is the exception, since it reads an origin-style ANTHROPIC_BASE_URL and speaks the Anthropic shape instead.

FAQ

Repository status, licences and release feeds for all eleven clients read from the GitHub API on September 21, 2026, and release tags move fast enough that Codex and Crush may already be a version ahead; OpenHands plan rows, CLI status, sandbox costs and LLM-profile behaviour read from openhands.dev and docs.openhands.dev the same day; PyPI versions read from pypi.org; the goose donation and the Roo Code shutdown read from their own announcements. Kunavo token rates come from the live catalog, and the dollar example is illustrative token arithmetic rather than a measured task cost.