The four steps
Read 4,018 public sources
We inspect public documentation, repositories, release records, pricing pages and other cited public material. Word of mouth, private messages and vendor briefings are not sources.
Record 4,152 findings
Each finding keeps its source links, the locator on that page, its check date, and the plan, surface and version it covers. We publish only findings that pass the publication checks for public visibility, review, source ownership, dates and scope.
Typed filter values need their own evidence-bound interpretation. A public finding alone does not establish every possible typed filter. The site reads only the generated public package in
data/; the full research records, notes and review history are published beside it inresearch/.Evidence contains a finding and a locator. A finding may summarize a source and must not be presented as a verbatim quote.
Cross-check 3,241 cross-checked
A second agent rereads the source and checks the answer and its scope. Only then does a value carry the small cross-checked badge. A first-party source can establish a value before that second read; agent cross-checks do not mean a human reviewed every claim.
Date 103 of 513 ADEs researched
Every answer keeps its own check date. In this snapshot, checks run from to , and the package was built on .
Worktrees separate files; they do not establish security isolation. A self-managed execution worker does not establish a self-hosted control plane. A model provider does not establish support for a coding-agent harness. Read the definition beside each filter.
Price filters
Price answers come from each ADE's public pricing page, terms or licence and describe what the ADE itself charges. Model and compute you pay for elsewhere are not counted. "Entirely free" (no paid plan) needs a first-party statement or licence grant that the ADE is free and no paid tier in the checked sources; a missing pricing page proves nothing and stays unknown. "Paid plans from" is the lowest public USD price per month at monthly billing, per user when priced per seat, without currency conversion or annual-price arithmetic. It shows "No paid plan" when the sources establish that no paid ADE tier exists; missing pricing evidence stays unknown. "No paid plan" does not match a paid-plan budget filter. An ADE can charge in several ways at once, so the pricing words can all apply to one product.
Who publishes this
Podium publishes this catalog and sells an agent development environment. Podium chooses the products, criteria, sources and wording, so treat these pages as vendor-authored research. Check important claims against the linked primary sources. The same evidence rules apply to Podium and its competitors, but they do not remove this conflict of interest.
Podium selects the products, rubric, sources and wording, including for comparisons that do not include Podium.
We say "by Podium" and "same rules for everyone". This is vendor-authored research. There are no sponsored positions, affiliate links or vendor calls to action inside results.
Podium's makers have full access to its source code, so Podium's research record cites more inspected source code than most others. Every product uses the same criteria, filters and source rule: a first-party source, meaning the product's own site, documentation, release notes or public repository code, can establish a value. Some of Podium's values rest only on its public source code, as values of other open-source products do. Deeper research can confirm more answers, and only confirmed answers count, so research depth can affect order (see Fit and order). The number of sources behind a single answer does not.
What each label means
- Sourced
- A claim passed the publication gates. A filter value also has a typed mapping that matches its evidence digest, without an accepted independent check of both.
- Agent cross-checked
- On a claim, a different agent checked its evidence, wording, scope and dates. On a filter value, a different agent checked both the public finding and the typed mapping. Lifecycle status copies the verification of its typed notice. This does not imply a hands-on test or human claim review.
- Unknown
- The filter value has no eligible receipt.
The available public evidence does not establish this value. Unknown is never no.
Sourced values appear as normal values. The small cross-checked badge is optional additional verification.
A sourced negative needs an explicit restriction or a bounded test. Silence, a missing page, and missing research do not establish no. Unknown cells link here; they do not count as failed requirements. A narrative finding can exist while its typed filter remains unknown. On the Finder an unknown stays quiet: a faint word with a dashed ring, never a cross.
Fit and order
"Fits" means every must-have has a confirmed yes. "Fits kind of" means a strict majority of must-haves have a confirmed yes. "Fits but shutting down" contains products with a published shutdown or acquisition notice that would otherwise fit in either group; it sits in its own row after the other two. Nice-to-haves affect order within a group and do not disqualify a product.
Within each group, products sort by must-haves met, then nice-to-haves met, then name. With no requirements selected, every researched ADE appears under Fits or, when it has a published notice, Fits but shutting down. There is no overall score, no extra tie-break by coverage and no rule that places Podium higher.
Products below the matching threshold remain in name search and the list of all ADEs. They have not necessarily failed your requirements. Unknown answers, missing evidence and research gaps do not establish a missing capability.
Research depth still affects order, because only a confirmed answer counts. In this snapshot, researched products have between 24 and 48 typed filter values; the median is 43. A product with fewer typed values can rank below others because we have not typed its answers yet, without lacking the capability.
The Parallel local agents preset makes parallel runs a must-have. Mac, agent, subscription, execution, worktree, pull request and review preferences rank the shortlist. No preset hides a product.
The GitHub stars order is optional and off by default. It keeps the groups and must-haves met, then sorts by the star count recorded from the product's own repository, with its observed date. Products without a sourced count are not zero; they follow in the usual fit order. Stars measure repository attention, not users, quality or capability, and closed-source products have none. In this snapshot 13 of 103 researched ADEs have a sourced count.
The phone word accepts a native iPhone app, a native Android app or a phone browser client; the desktop word accepts a Mac, Windows or Linux app. The open-source word accepts an explicitly established Apache-2.0, MIT, GPL-3.0 or AGPL-3.0 main-application license, including the listed or-later variants. Component and hosted-service exceptions remain in each receipt. Other licenses and unknown values stay outside this selection.
Benchmarks from ADEMark
Speed, memory, CPU, token and team-build numbers come from ADEMark, a third-party open benchmark that runs the same tests on every ADE it measures (Mac mini M4, 16 GB, macOS 27 · Claude Opus 5.5, effort high). Its current results cover 15 ADEs, measured up to 12 November 2026. We import its published summary unchanged under CC BY 4.0 and keep a copy at /benchmarks/ademark.json with the mapping from ADEMark's app names to ours. ADEMark's repository has the harness and one row per run.
These are measurements, not documentation findings, so they sit in their own section on each ADE page and in Compare. They never set a typed filter value, never change Best fit and never add up to a score. ADEMark publishes no overall score either. A place such as "3rd of 12" counts only apps with a value for that test; ADEMark treats close medians as ties, and "Did not finish" and "Not measured" are never ranked. The Bare CLI, Claude Code in a terminal, is ADEMark's unranked control and appears only for context.
The Benchmarked by ADEMark word in the Finder only asks whether ADEMark has results for an ADE. An ADE it has not tested is unknown, not a no. We re-import after each ADEMark pass and at least once a month, and check every newly researched ADE against ADEMark's list. ADEMark marks results stale after 60 days; past that, pages say the numbers may be outdated.
Data age
- Price and availability findings are normally scheduled for a recheck within 60 days.
- Capability findings are normally scheduled within 120 days.
- Stable architecture and license findings are normally scheduled within 180 days.
Findings and typed values stay published after nextCheckAt. Show their checked date and age; add a quiet "may be outdated" note only after six calendar months. A rebuild never refreshes evidence.
These dates are research reminders. Going past one never hides a finding or a typed value. Values retain their original checked date, which may differ from the page build date. Only checks more than six calendar months old receive the quiet "may be outdated" note. Contradictions, private evidence, changed review digests and unsupported interpretations still prevent publication.
Shutdowns and status
A published first-party shutdown or acquisition notice is enough to put an otherwise matching product in Fits but shutting down. A cross-check adds the badge. Dates, service scope and any local-app continuation remain in the notice. There is no inclusion toggle. An absent notice means unknown, never active.
The status log contains eligible research and product-change events with published citations. It is not a complete history. Today's finding is not necessarily the exact answer at the date of an older event. A research recheck is not proof of a hands-on test.
Pair pages
A generated pair comparison requires at least 20 public findings on each side and at least 6 filters with a typed answer on both sides. We apply the threshold to every product. Unknown is not a difference and not a match. If corrected evidence removes that coverage, the next build omits the pair page and its sitemap entry. Age alone never removes it.
Static pages are limited to 300 pairs. They go to the pairs with the most filters answered on both sides; pairs with the same count are kept or dropped together, so in this snapshot every static pair has at least 45 shared typed answers (259 pages). Every other pair of researched ADEs opens in Compare with the same receipts.
Open data and code
Everything the site shows is in a public package you can read without the site. Agents should start at /llms.txt.
- /llms.txt
- An index for agents: every file, what it holds, and the rules for citing it.
- /products/index.json
- All 513 names and the typed filters as bitsets, in a fixed row order.
- /products/<id>.json and .md
- One ADE: every finding with its sources, locators, scope and dates, plus any shutdown notice.
- /data/facets.json
- What each filter means, word for word.
- /data/lifecycle-notices.json
- Published shutdown and acquisition notices, with dates and sources.
- /data/status-log.json
- Cited research and product-change events.
- /data/methodology.json
- The rules on this page, as data.
Site code is Apache-2.0. The data package is CC BY 4.0. The code and data live at github.com/madeinorbit/ade-comparison.
What we cannot establish
- Documentation and source inspection do not establish successful hands-on use.
- Cross-checks are done by agents working for the publisher. No human has audited a sample of them yet; a human-audited sample with a published error rate is planned.
- A missing sourced finding is not evidence that a capability is absent.
- Clients, plans, versions and execution modes may have different capabilities. Read the receipt scope.
- Register listings identify products and contain no capability claims. A sourced first-party lifecycle notice can name a listed product before an independent cross-check.
Documentation does not establish reliability, usability, task quality or performance in your environment. We run no benchmarks ourselves; the ADEMark results come from a third party, on one machine and model. We do not estimate your total agent, compute or subscription bill.
Unresearched names are plain listings, without product or capability endorsements. Screenshots show appearance. Their provenance and retrieval dates appear beside them; a fallback image is labeled. Images do not prove that a capability works.
Corrections and missing ADEs
Anyone can report wrong data or suggest a missing ADE. Include the product, what is wrong or missing, a public source URL and the exact sentence that supports the change. A finding on this site may summarize a source, so quote the source itself.
Word of mouth and private messages are not evidence. Vendors get no exception, including Podium.
Both actions open GitHub issue forms. Submissions are public and require a GitHub account. Page and receipt context is prefilled; you supply the supporting source and sentence.
We verify the source and scope, update the research catalog in research/, then regenerate data/ after its publication checks pass. The data package is generated; direct edits to its answers would be overwritten. Submitting an issue does not change the site automatically.
Reuse and privacy
Site code is Apache-2.0. The public data package is CC BY 4.0, with attribution to adecomparison.com by Podium and the linked primary sources. Third-party screenshots retain their owners' rights and provenance. The Radio Canada typefaces are under the SIL Open Font License.
This site uses Google Analytics, loaded after the page has finished loading, to count page views and which filters, ADEs and comparisons people use. We use these counts to decide what to research and publish next. It has no advertising pixels, affiliate links or accounts. Filter and compare state lives in the URL. Copying a prompt uses your browser clipboard only when you press the button.
Rubric version 0.3.