Plandex vs Aider: workflow, models and total cost

Choose Aider if you want a third-party OpenAI-compatible endpoint: its documentation makes that two environment variables and an openai/ prefix, with no plan, account or license gate. Choose Plandex only if you will self-host it — its own support matrix restricts custom providers to self-hosted installs, and its last code push was October 3, 2025. The two projects are in very different states, and that decides more here than any feature list.

Both are free terminal coding agents with no charge for the software, so the comparison is not a price comparison. It is a comparison of how much work each one makes you do to reach the endpoint you have already chosen, how each one treats your working tree, and what you inherit by picking a project whose vendor domains stopped resolving.

What you are actually choosing between

One distinction that table depends on. The company and the cloud service wound down; the repository did not. The maintainer's wind-down post — dated October 3, 2025, and readable only in the Internet Archive now that the domain is dark — says he accepted an engineering position elsewhere, that new trials and signups were disabled and subscriptions paused, that Plandex Cloud would keep running for existing customers until it shut down on November 7, 2025, and that the open source project would remain on GitHub. It does not promise further releases, and it does not announce that the code is abandoned; on GitHub the repository is still live, MIT and unarchived. The DNS result is an observation, not a verdict: the .ai registry record for plandex.ai is paid up to September 2027 and was last updated on August 12, 2026, about six weeks before this check, while the live delegation points at nameservers that refuse queries. The domain could come back.

Aider's own numbers need a caveat in the other direction. Its release notes list 2026-era model support — Claude 4.5/4.6 aliases, Gemini 3 preview models, GPT-5.1 through 5.4 settings, a new /ok command — under a "main branch" heading, meaning merged but not necessarily released. Work landed on main in April and May 2026, after the February 12, 2026 PyPI upload, so a plain pip install aider-chat does not include it. That is an inference from the dates, not a diff of the package. A third-party fork, cecli-dev/cecli, ships far more often, but its releases and features are not Aider's and should not be read as such.

Two search collisions to clear before you read anything else on this pairing. A different company in accounting software also trades as Aider, at a different domain, selling period-close automation for accountants and bookkeepers with its own published pricing; those prices describe that product and have nothing to do with the coding CLI, which is why no monthly figure for Aider appears anywhere on this page. And the Plandex copies that surface in results — Pendia/Plandex-AI-Coding-Agent and mattias50/plandexV2 — are zero-star forks last pushed in May and March 2025 whose GitHub descriptions copy or lightly reword the original's. The only official repository is plandex-ai/plandex.

Which one to pick

Pick Aider when your endpoint is a third-party OpenAI-compatible API, when you work inside git and want every edit committed as it happens, when you want an install path that still receives packages, and when you want the setup to be a shell export rather than a JSON schema. The migration cost is minutes plus one metadata file.

Pick Plandex when the review model matters more than the cadence: you want AI changes held in a sandbox and applied only after you have read the diff, you want a preset autonomy ladder rather than per-message mode switching, and you are willing to run Docker and Postgres to self-host. Accept in advance that you are adopting a July 2025 release whose vendor domains do not resolve, and that the documented quick install is one of the things that breaks because of it.

Pick neither when you want an actively released agent with a first-class custom-endpoint path and a maintained plugin surface. Aider alternatives covers that field, and OpenCode, Claude Code and Cline are three exits with published Kunavo configuration.

Workflow: review-then-apply against apply-then-undo

This is the largest structural difference and it is not a matter of taste. Plandex's reviewing-changes documentation states that changes are not applied directly to your project files by default and instead accumulate in a version-controlled sandbox so you can review them first, through plandex diff, a browser view at plandex diff --ui, and plandex reject for per-file rejection before plandex apply. Separately, its version-control documentation says the plan itself is versioned: context changes, prompts, responses, builds, rejections, applies and model-setting changes each create a version, browsable with plandex log and reversible with plandex rewind. That history is Plandex's own store, not your git history.

Aider inverts it. Its git page says it asks to create a git repo if you launch it in a directory without one, that whenever it edits a file it commits those changes with a descriptive commit message, and that it first commits any preexisting uncommitted changes so your edits stay separate from its. The recovery is /undo, which undoes and discards the last change. Whether Plandex needs a git repo at all was not confirmed here.

Neither project documents an embedding index, and Kunavo serves no embedding model either, so nothing in either documented setup calls for one. Both describe a tree-sitter route instead: Plandex's README advertises tree-sitter project maps, and Aider's supported-languages page says its repository map is built from each language's tree-sitter grammar tags.

Permissions: a ladder against a mode switch

Plandex gates behavior with five presets set by plandex set-auto or launch flags. Aider has no equivalent matrix: its chat modes are code, ask, architect and help, switched per message or stickily, and shell execution is the explicit /run (alias !) and /test commands its in-chat command list documents, rather than a permission level.

Read from Plandex's autonomy documentation on September 21, 2026. The practical reading: the sandbox advantage above only holds below Full, because auto-apply is what Full turns on.

Bringing your own endpoint: the sharpest difference

Aider's OpenAI-compatible page publishes the complete configuration. There is no tier to clear: no aider.chat page checked here puts a plan, an account or a hosted service between you and the endpoint.

export OPENAI_API_BASE=<endpoint>
export OPENAI_API_KEY=<key>

aider --model openai/<model-name>

The real boundary is metadata rather than permission. Aider's warnings page says that for a model it has never heard of it will use an unlimited context window and assume the model is free, so a custom id reports untrustworthy costs until you write the file its advanced model settings page documents: a .aider.model.metadata.json in your home directory, the git root or the launch directory, keyed by the fully qualified provider/model name and carrying input_cost_per_token, output_cost_per_token, litellm_provider and mode. Aider pricing works through that file and the edit-format default that sits behind it.

Plandex publishes a support matrix instead, and it is the single most decision-relevant paragraph in either project's documentation: self-hosted supports custom models, providers and model packs in full; Cloud with BYO API Keys supports custom models and packs but says models can only use built-in providers; Cloud with Integrated Models supports custom packs of built-in models only. On that matrix, neither paid Cloud tier reached a custom endpoint. And the environment-variable reference lists no general base-URL override for a provider — AZURE_API_BASE is Azure-specific and OLLAMA_BASE_URL is a server-side setting for reaching Ollama outside the container — so the JSON file opened by plandex models custom is the only documented route to an arbitrary OpenAI-compatible endpoint.

{
  "providers": [
    {
      "name": "my-provider",
      "baseUrl": "https://api.example.com/v1",
      "apiKeyEnvVar": "MY_PROVIDER_API_KEY"
    }
  ]
}

That block is the easy third of it. Each custom model then declares modelId, publisher, description, defaultMaxConvoTokens, maxTokens, maxOutputTokens, reservedOutputTokens and a preferredOutputFormat of either xml or tool-call-json, and each model is then slotted into a pack by role — planner, summarizer and builder are required, with architect, coder, wholeFileBuilder, names, commitMessages and autoContinue optional. The generated template's $schema points at https://plandex.ai/schemas/models-input.schema.json, which does not resolve; the documentation describes that schema as what gives editors autocomplete, validation and inline documentation, and the schemas are checked into the repository under app/cli/schema/json-schemas/, so a local path is the workaround. Treat that as an editor-tooling concern — whether the CLI ever fetches the remote schema at runtime was not checked.

Two more Plandex details worth knowing before you adopt it. Its documentation and its code disagree about the cloud API host: environment-variables.md says PLANDEX_API_HOST defaults to https://api.plandex.ai while clients.go hardcodes https://api-v2.plandex.ai. Both are on the domain that does not resolve, so the discrepancy changes nothing practical, but trust the code. And for built-in providers, Plandex's provider documentation reads keys from environment variables at CLI launch, gives direct providers precedence over OpenRouter when both are set, and claims direct providers cost about 5% less than OpenRouter — that is Plandex's claim from July 2025, not a measurement, and our OpenRouter comparison is the place to check it against current terms.

What each one actually bills you

Neither project takes money. Aider is Apache-2.0 with no plan, tier, seat or hosted service on any aider.chat page checked here, and Plandex is MIT with a public server image and a self-hosting quickstart that documents no license step. The entire bill in both cases is model tokens at your provider's rate, plus, for Plandex, whatever running Docker and Postgres costs you. Plandex Cloud is the one price anchor floating around search results, and it is not something anyone can buy: its Cloud page still lists $45 a month for Integrated Models and $30 a month for BYO API Keys under a notice saying Cloud is winding down as of 10/3/2025 and is no longer accepting new users, and the post that notice links says the service shut down on November 7, 2025.

Because the software is free on both sides, the useful cost question is which model you run and how many roles bill. Both tools fan a task across more than one model — Plandex's pack requires a planner, a summarizer and a builder, and Aider's options reference describes --weak-model as the model used for commit messages and chat-history summarization — so a single headline rate understates both. The figures below are illustrative token arithmetic, not measured task costs and not a bill ceiling: assume one session sending 250,000 uncached input tokens and receiving 15,000 output tokens, at live Kunavo catalog rates per million tokens.

The side roles are where the two tools reward the same habit. Assume the summarizer and commit-message traffic is another 40,000 input and 2,500 output tokens: that costs $0.063 if it runs on Claude Sonnet 4.6 and $0.021 on Claude Haiku 4.5. Plandex lets you name a different model per role in the pack; Aider exposes --weak-model for the same purpose. Scale by your own sessions before treating any of this as a budget, and check the charge your provider account actually recorded — neither tool gives you a dependable spend readout on a custom endpoint, as the FAQ below explains.

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. The minimum top-up is $10 in prepaid credit, which is a funding minimum rather than a task fee or a subscription — see billing details. For the method behind shrinking these numbers, see AI cost optimization.

Migration cost, side by side

One Python detail for the Aider column, because the two official sources describe different packages: the install page mentions Python 3.8–3.13 for the bootstrap installer, while the aider-chat 0.86.2 metadata on PyPI declares >=3.10,<3.13. The package metadata is what pip enforces for the tool itself.

Setting either one up against Kunavo

Kunavo publishes a configuration reference for Aider and none for Plandex. Neither client has been runtime-tested against Kunavo's endpoint, and a published setup page is a configuration reference rather than a compatibility result. For Aider, the Aider integration guide has the two variables and the openai/ prefix. For Plandex, there is no Kunavo setup page to follow: you would write the provider and model blocks above by hand against a self-hosted install, using the field list from Plandex's own documentation. The OpenAI-compatible API guide covers the shared mechanism both routes rely on. Keep a working route available while you try either, run one bounded task, then read what your account recorded. Create a Kunavo account when you are ready to fund a key.

FAQ

Checked September 21, 2026 for this page: the GitHub repository metadata, contributor counts and release tags for both projects, the PyPI record for aider-chat, DNS resolution for plandex.ai and aider.chat on Google Public DNS and Cloudflare DNS, the .ai WHOIS record for plandex.ai, the Docker Hub record for plandexai/plandex-server, the Internet Archive capture of the Plandex wind-down post, and the Plandex and Aider documentation pages linked above. Neither tool was installed or run, so every behavior claim here is read from documentation or source rather than observed. Kunavo token rates come from the live catalog and every dollar figure is illustrative token arithmetic.