Factory Droid alternatives: choose by workflow and API control

Pick a Factory Droid alternative by the boundary you actually hit, not by feature count. Three boundaries send most people looking: Factory's plan fee survives your own API key, custom models reach only the Droid CLI and the desktop app, and every request needs headroom in three independent rolling windows. Each of those points at a different replacement, and one of them is not a reason to leave at all.

Droid is not in trouble, so this is a fit decision. The npm package droid was at 0.223.0 on September 21, 2026, published September 19, with six releases in six days, and the public Factory-AI/factory repository is not archived. The domain move confuses people: factory.com is the canonical marketing host, factory.ai answers a redirect to it, and documentation stayed on docs.factory.ai.

Six reasons people leave, each from Factory's own documentation

Read from Factory's individual-plans page, its BYOK reference and its models list, all checked September 21, 2026. One pairing is misread constantly and is worth stating carefully: the BYOK page's sentence about hosted web and mobile is about custom-model reach, while the FAQ answer on factory.com/pricing — "All products; IDE, Web, CLI, Slack etc are included in all plans" — is about plan access. Both are published and both are true. Factory names the CLI and desktop app as included and hosted web and mobile as excluded; Slack appears in neither sentence, so this page does not claim it either way. The full plan table lives on Factory Droid pricing.

Which alternative answers which exit

Repository state read from the GitHub API and the licence from each project's own LICENSE file on September 21, 2026. GitHub's detector reports Crush as NOASSERTION, which is a tooling artefact rather than a claim that it is unlicensed — its LICENSE.md is explicit. Two clients frequently suggested as destinations are not recommended here: Continue's releases stopped at June 19, 2026 after the Cursor acquisition, and Aider's last commit is May 22, 2026 with its last PyPI release dated February 12, 2026. Neither repository is archived, which is why the dates rather than the word "dead" are the honest description. Aider alternatives covers that project's own successors.

Say the trade out loud before you move. Leaving takes Droid Core — the free open-weight model pool with its own separate rate limits that catches you after Standard Usage runs out — along with Factory Router (shown in some surfaces as Auto Model), Missions, and the hosted web, mobile and Slack delegation surfaces. Factory Router is easy to misread as a bring-your-own-gateway feature; it routes across Factory's own billed models, which is the opposite. None of that comes with you.

What migrates, what gets rewritten, what stays behind

The MCP rewrite is the step that bites, because the two files look similar enough to copy and then fail quietly. The key, the type vocabulary and the shape of command all differ:

// Droid — ~/.factory/mcp.json or .factory/mcp.json
{
  "mcpServers": {
    "my-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "my-mcp-command"],
      "env": { "MY_ENV_VAR": "value" }
    }
  }
}

// opencode — opencode.json. Different key, different type words,
// and "command" is an ARRAY that swallows the args.
{
  "mcp": {
    "my-server": {
      "type": "local",
      "command": ["npx", "-y", "my-mcp-command"],
      "enabled": true,
      "environment": { "MY_ENV_VAR": "value" }
    }
  }
}

Kunavo serves no embedding model, so if your destination client indexes a repository with embeddings, that step needs a separate provider — it is not part of a gateway key. Sources: Droid's MCP page, its instruction-file page, Claude Code's MCP docs, opencode's MCP docs and Codex's config reference, checked September 21, 2026. Codex's documentation moved: developers.openai.com/codex answers a redirect to learn.chatgpt.com.

A worked estimate for the endpoint side

Factory's own figures cannot produce a comparable number, because the Standard Credit has no published dollar value. So this prices only the part a move changes — the tokens your new client sends — and leaves the plan fee you drop as a published figure you can add yourself. Assume one working day of agent use at 1,200,000 uncached input tokens and 80,000 output tokens, and 20 such days in a month. Those assumptions are invented for illustration, not measured. Rates are live Kunavo catalog prices per million tokens.

This is illustrative token arithmetic, not a measured task cost and not a bill ceiling. Under these assumptions model choice alone moves the modeled month by a factor of 5.0: Claude Haiku 4.5 lands at $12.80 against $64.00 for Claude Sonnet 5 on the identical token shape. Whether that beats a subscription depends entirely on your own volume, and only your own token counts settle it — a heavy month on a metered endpoint can cost more than a flat plan, which is the case the plan is for.

Keep the two questions apart as well. The cheapest listed rate and the lowest cost to finish a task are different claims: a cheaper model that needs three attempts can cost more than a dearer one that needs a single pass. AI cost optimization covers measuring that on your own repository. 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, which is a funding minimum rather than a subscription; see billing details.

Move in five steps, and keep the way back open

Copy the instruction file first — there is nothing to convert. Re-declare your endpoint in the new client's own format. Rewrite or copy the MCP block per the table above. Rebuild any slash commands as whatever the target calls them, and accept that session history stays behind. Then run the same bounded task on both setups for one billing period and read the charge each side recorded.

The rollback is cheap if you plan for it: leave Droid installed and ~/.factory/settings.json untouched while you trial the replacement. Two billing details shape the timing, and Factory states both in the same FAQ answer: "Standard Usage does not roll over month to month. Extra Usage that you purchase does roll over." Extra Usage is also prepaid with a $10 minimum and carries no expiration, so a balance you bought survives the decision (Factory's individual-plans page, September 21, 2026).

If the boundary you hit was the plan fee rather than the client, the cheaper move may be no move at all: Droid takes a custom endpoint through three provider values, and the Factory Droid setup page has the JSON. Factory qualifies that route itself — "Only Anthropic and OpenAI models accessed via their official APIs are fully tested and benchmarked" — and notes that prompt caching on a generic Chat Completions endpoint "varies by provider and cannot be guaranteed". Those are Factory's caveats about custom endpoints in general, and they apply here. If you are moving, start at opencode, OpenHands, Cline, goose, Crush or Codex CLI, and create a Kunavo account when you are ready to fund a key. For the head-to-head against Anthropic's client, see Factory Droid vs Claude Code.

FAQ

Factory's pricing, BYOK, models, MCP and CLI documentation, the npm registry entry for droid, and the GitHub API state of opencode, Cline, goose, OpenHands, Codex CLI, Crush, Aider and Continue were all read on September 21, 2026; Continue's dormancy was confirmed against its default-branch commit list rather than its pushed_at timestamp. No client on this page was runtime tested against Kunavo, including Droid itself, and Kunavo token rates come from the live catalog while every dollar example is illustrative arithmetic.