Continue vs Cline: which to pick in 2026

Cline is the only one of these two still being built: its VS Code extension published 4.1.19 on September 17, 2026 and its npm CLI published 3.0.62 on September 15, 2026, while Continue was acquired by Cursor and its maintainers declared the repository read-only and no longer actively maintained. That does not make this an easy call in one direction, because Continue's final builds still run and its configuration is now purely local. The honest decision is about migration cost, not a feature score — and which side you fall on depends on one question: did you point Continue at your own endpoint, or at Continue's Hub?

If you arrived here searching "cline vs continue" expecting two live products to weigh up, the premise has changed underneath the query. Whatever a review page ranking for these terms still says about Continue's paid plans and its Hub, neither is on sale: continue.dev/pricing and /hub both return 404, and Continue's own homepage says recurring billing has been disabled. Everything below traces to an official source with the date it was checked.

What each project actually is today

Sources, all checked September 19, 2026: continue.dev (its page title is literally "Continue (acquired by Cursor)"), the Continue README, the GitHub API for continuedev/continue and cline/cline, the @continuedev/cli and cline npm registry records, and the Visual Studio Marketplace extension query API.

Two version traps are worth naming, because both are easy to get wrong. First, Continue's README describes "a final 2.0.0 release of the VS Code extension, CLI, and JetBrains plugin" — but npm carries no 2.x of the CLI at all. Across 332 published versions of @continuedev/cli, the newest is 1.5.47. "Continue CLI 2.0.0" is not a version you can install. Second, the VS Code number depends on the channel: GitHub's last non-prerelease tag is v2.0.0-vscode, while the Marketplace currently serves 2.1.0 with its PreRelease property set to true. Name the channel whenever you name the number. And say "the maintainers declared it read-only" rather than "the repo is archived" — the GitHub API reports archived: false, so those are two different claims and only the first one is true.

Who should pick which

Read this by your situation rather than by a score, because the cost of switching is doing most of the work here.

The blunt version: if you already own the endpoint, the shutdown cost you nothing yet and urgency is low. If Continue was also your model catalog, your MCP registry and your login, three things disappeared at once and no amount of pinning an old version brings them back.

Workflow, permissions and surfaces

From Continue's tool-permissions documentation, its config reference, Cline's CLI reference, Cline's configuration guide and the Cline CLI README, checked September 19, 2026. One nuance on the Cline row: the CLI README also describes the interactive TUI as offering live tool approvals, so "auto-approve defaults to true" is the flag default and not a claim that every surface runs unattended. Confirm the behaviour in your own shell before you point it at a repository you care about.

The BYOK boundary, which is where the two really differ

Both clients will talk to a third-party OpenAI-compatible endpoint. The asymmetry is in which protocol each surface will accept a custom base URL for, and it is visible in the source.

Continue treats apiBase as a generic per-model override — its reference says it "can be used to override the default API base that is specified per model". That means both the OpenAI-protocol route and the Anthropic-protocol route are expressible in the same file. The OpenAI-compatible provider page documents the block, and its example carries the /v1 suffix:

models:
  - name: Sonnet via gateway
    provider: openai          # the wire protocol, not the vendor
    model: claude-sonnet-4-6
    apiBase: https://api.kunavo.com/v1
    apiKey: ${{ secrets.KUNAVO_API_KEY }}
    roles: [chat, edit, apply]
    capabilities: [tool_use]  # declare it: autodetection is frozen at June 2026

Two frozen-code details matter here. Continue's reference says it "automatically detects these capabilities for most models" and that capabilities overrides that autodetection — and since the last builds shipped in June 2026, whatever model ids that detection recognises is now fixed. For an id it does not know, declare capabilities: [tool_use] yourself; the reference marks it "required for Agent mode". And the provider page states that Continue itself "uses OpenAI's /responses endpoint for o-series and gpt-5 models" by default, so if your gateway expects chat completions for one of those ids, useResponsesApi: false forces /chat/completions — the same page documents useLegacyCompletionsEndpoint: true for the older completions route.

Cline's OpenAI-Compatible provider asks for Base URL, API Key and Model ID, plus Max Output Tokens, Context Window, Image Support, Computer Use, Input Price and Output Price. In the CLI, the base URL is validated. The auth command on main returns this literal string when you pass one to anything else:

base URL is only supported for OpenAI and OpenAI-compatible providers

So for the Cline CLI, take the OpenAI-compatible route. One caution on the command itself: the validator checks against an internal provider-id constant whose literal string this page did not read, so do not copy a guessed --provider value from anywhere, including here.

# Cline's own help output is the only safe source for the provider id
# string, so read it rather than copying one from a guide.
cline auth --help

# Then either run it interactively...
cline auth

# ...or supply the pieces the validator accepts for an
# OpenAI-compatible endpoint: key, model id and base URL.
#   --baseurl https://api.kunavo.com/v1

A gateway-specific tip that costs nothing and pays back immediately: fill in Cline's Input Price and Output Price fields with your gateway's actual per-million rates. Cline computes its per-task cost readout from those numbers, and left at a default they will describe some other provider's pricing rather than yours.

Two boundaries that are not about protocol. Cline's enterprise remote configuration states that once an admin saves an OpenAI-compatible provider, organization members "won't be able to select other providers or switch to their personal Cline accounts" — so in a managed org, BYOK is an admin decision rather than yours. And Cline's free model promotions run the other way: that page says they are "available to any user with a Cline account" and that free model usage is "not supported through the Cline API", so they are a Cline-account route rather than something your own gateway key reaches.

What the same session costs on either client

Neither client charges for itself. Continue is Apache-2.0 with nobody left to bill you; Cline's pricing page lists its Open Source tier at Free, covering the VS Code extension, the CLI and "Purchase and Model Inference at Cost or BYOK", with Enterprise priced as Custom. So the token bill is the whole comparison, and it is identical between the two for the same model and the same traffic.

The figures below are illustrative token arithmetic, not measured task costs and not a bill ceiling. Assume one session that sends 200,000 uncached input tokens and receives 12,000 output tokens. Rates are live Kunavo catalog prices per million tokens.

What this table deliberately does not do is claim one client sends fewer tokens than the other. That would need a measured run across both, and none was performed here. The client choice changes your ergonomics, your permission defaults and your maintenance risk; the model choice changes the bill. On these assumptions Claude Opus 5 costs about 5× what Claude Haiku 4.5 does for the same traffic, which is a larger lever than anything in the feature table above.

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 multiplied by the applicable markup. Cache charges and external tools sit outside this example. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a subscription. See billing details.

For completeness on Cline's own paid routes, since they change the comparison for some readers: Cline credits are pay-as-you-go with no rate card published anonymously, and ClinePass is $9.99 per month with the footnotes "Standard rate $9.99 per month after promotion period" and "Additional processing fee may apply". ClinePass is a separate product from Cline credits, and it covers open-weight models only — GLM, Kimi, DeepSeek, MiniMax, MiMo and Qwen families — so it is not a route to Claude or GPT and not a substitute for a gateway if those are the families you use. Cline's ClinePass documentation describes usage as measured against three limits — a five-hour rolling window, the calendar week and the calendar month — and the marketing page claims "2-5x the usage compared to standard API rate limits"; neither publishes the underlying quota figures, and this page does not invent them.

The docs are stale; the source is authoritative

This is the single most useful thing a stranded Continue user can be told, and it is re-checkable in a minute. Continue's CLI quickstart still lists "A Continue account, or an Anthropic API key" under requirements, still documents cn login to authenticate in a browser, and still tells headless users to export CONTINUE_API_KEY. None of that can work. The login opens a browser against a service whose API host times out, and the shipped onboarding code on main asks only this:

To get started, enter your Anthropic API key.

There is no "log in with Continue" branch left in that file. The login requirement was removed on purpose — the commit message reads "remove dead login requirement and Login Required UI" — and a later commit replaced Hub slugs in the default config with inlined model definitions, with a source comment explaining that Hub slug resolution has been removed. The in-product deprecation banner said it in the product's own words — Extension configuration is local only as of v2.0.0 — though you will not see it any more: that component carries a hard-coded expiry of September 9, 2026 and renders nothing after it, so the message is now only in the source. The docs site was last modified in July 2026 and now serves from GitHub Pages; the code is the current truth.

Two practical consequences. First, the onboarding prompt validates the key's shape: the CLI's validator accepts it only if it starts with sk-ant-, so a gateway key in any other format is rejected there. The escape the source shows is to pass --config <path> on the first run, which is checked before onboarding. This page did not execute that escape — it is read from the source, not tested. Second, the hub-only flags are dead weight now: --model, --mcp and --agent are each documented in the CLI source as taking a hub slug in owner/package form, and --org takes an organization slug — all resolved from a host that has no DNS record. --prompt and --rule are the exceptions worth knowing: each accepts a file path or literal string content as well as a hub slug, so those two still work. The --config <path> file form works too, because it is a local path.

Setting either one up against a gateway

Kunavo publishes configuration references for both clients. Those are published setup guides, not compatibility tests: neither client has been runtime-tested here against Kunavo's endpoint, and nothing on this page is a benchmark or a measured result. The Continue reference remains accurate for a reason worth stating — the config schema is frozen, so it cannot drift.

Start at the Cline integration guide or the Continue integration guide, and use the Anthropic base URL doc if you are taking the Anthropic-protocol route from Continue or from Cline's IDE extension. Run one bounded task, then read the charge your account actually recorded for it rather than trusting the client's own estimate. Create a Kunavo account when you are ready to fund a key.

Comparing further afield? Cline pricing goes deeper on Cline's three payment routes, best API for Cline weighs the provider options, and Roo Code vs Cline and Kilo Code vs Cline cover the two forks — which are separate products with their own prices, not interchangeable with Cline. If the gateway itself is the open question, OpenRouter alternatives and best LLM gateway are the comparisons to read.

FAQ

Checked September 19, 2026 against primary sources only: continue.dev and its HTTP status codes, DNS resolution for hub.continue.dev and api.continue.dev, the npm registry records for @continuedev/cli and cline, the GitHub API for both repositories, the Visual Studio Marketplace extension query API, the JetBrains Marketplace plugin API, cline.bot/pricing and cline.bot/cline-pass, the Continue and Cline documentation sites, and the projects' own source files on main. Neither client was installed or run, so no claim here is a benchmark, a speed measurement or a measured task cost. Kunavo token rates come from the live catalog, and every dollar example is illustrative token arithmetic.