AX Unified API Index · audited October 2026
We scored 14 unified API platforms on the seven things an AI agent needs to integrate on its own: readable docs, a hosted MCP server, tool discovery, a machine-readable spec, structured errors, reliable webhooks and self-serve access. Every score links to the dated public source it rests on.
AX score, out of 5
9 of 14 fully verified
Unweighted average of 7 dimensions, each scored 1 to 5 from a dated, linked public source. Only platforms with every dimension verified are ranked; 90 of 98 scores are verified so far.
What AX measures
Developer experience asks whether a person can integrate. Agent experience asks whether software can, with no one to read between the lines. Each dimension is scored on a published 1 to 5 rubric.
Can an agent read the docs without scraping HTML?
1 = absent. 3 = present but partial. 5 = complete llms.txt + llms-full.txt covering every endpoint.
llms.txt and llms-full.txt, a markdown twin of every docs page and Accept: text/markdown negotiation. See it · Evidence
Is there an MCP server, and does someone else host it?
1 = none. 2 = repo without releases. 3 = self-hosted Docker. 5 = hosted endpoint.
A hosted MCP server at mcp.apideck.dev covering every Unified API. Nothing to clone or run. See it · Evidence
Can an agent enumerate every action from one URL?
1 = scrape required. 5 = single endpoint enumerates all tools.
One endpoint lists, searches, describes and runs every tool, plus a public MCP Explorer. See it · Evidence
Is there a public, complete OpenAPI spec an agent can load?
1 = none. 3 = via Postman only. 5 = public URL, GitHub-published, drives the SDKs.
Public OpenAPI specs for every Unified API, published on GitHub and used to generate the SDKs. See it · Evidence
Can an agent branch on an error instead of guessing?
1 = generic 4xx. 3 = structured JSON. 5 = structured + custom codes + metadata.
A stable type_name on every error, the downstream status and body, and a retriable flag for timeouts. See it · Evidence
Can an agent rely on events arriving, and recover when they do not?
1 = polling only. 3 = native webhooks. 5 = native + virtual + replay + dead-letter.
Native and virtual webhooks, 10 delivery attempts over about 28 hours, auto-disable and delivery logs. See it · Evidence
Can an agent get credentials without a sales call?
1 = none. 2 = sales-gated. 4 = free trial. 5 = self-serve sandbox.
A 30-day free trial with API keys from signup, no sales call. See it · Evidence
Apideck AX score
4.6/ 5
#2 of 9 fully verified
Next: StackOne at 4.7
Every score
Sort by any column to see who leads on what. Select a cell for the evidence behind it. A dash means no dated, linked source has been recorded yet; it is not a low score.
Select a cell to read how it was scored.
Your agent, your weights
A coding agent leans on the spec and the docs. A runtime assistant lives in MCP. Pick a profile or set the weights yourself and watch the ranking move.
Every dimension counts the same. This is the published index.
Weighted AX score
Apideck ranks #2 at 4.6
Not ranked until every dimension is verified: Chift (6/7), Membrane (6/7), Codat (5/7), Kombo (5/7), Rutter (5/7).
Head to head
Averages hide trade-offs. Pick any platform to compare its shape with Apideck's, dimension by dimension.
Apideck
4.6
StackOne
4.7
Apideck leads on 1, ties on 4, trails on 2
What an agent sees
The scores describe this path. Step through what an agent does with Apideck, from reading the docs to handling a timed-out write.
$ curl https://developers.apideck.com/llms.txt
# Apideck
> Apideck is a unified API platform. Integrate once to connect to
> 200+ services across CRM, Accounting, HRIS, ATS, File Storage,
> E-commerce, Issue Tracking, and more.
- [Full index](https://developers.apideck.com/llms-full.txt)
- [Markdown index](https://developers.apideck.com/md/index.json)
- [OpenAPI specs](https://github.com/apideck-libraries/openapi-specs)
...Where we lose points
An index is only useful if it is honest about Apideck too. These are the dimensions where Apideck's verified score is below 5, and what a 5 would take.
llms-full.txt lists every API operation from one URL, and the MCP Explorer page says each operation maps to one MCP tool. On the MCP server itself, dynamic-mode list_tools returns at most 50 tools per call and static mode currently registers only four workflow tools, so a single tools/list does not enumerate every tool. Scored 4 rather than 5 for that gap. Evidence is vendor documentation, not a runtime test.
For a 5: One tools/list call that enumerates every tool. Today list_tools returns at most 50 per call.
Evidence recordThe webhooks guide documents native webhooks, Virtual Webhooks that poll connectors without native support, HMAC-SHA256 signatures, idempotency keys and a 10-attempt retry schedule over about 28 hours. Failed events can be redelivered only by contacting support, and there is no self-serve replay or dead-letter queue, so it falls short of level 5. Evidence is vendor documentation, not a runtime test.
For a 5: Self-serve replay of a failed event. Today support redelivers it.
Evidence recordGet started and the pricing page offer a self-serve 30-day free trial with full access and no sales call. Dedicated sandboxes for systems such as Workday and NetSuite are listed as an Enterprise feature. This meets the free-trial level (4).
For a 5: A permanent self-serve sandbox.
Evidence recordMethod
Audited October 2026 against each vendor's own public docs. Every score is re-checked at least every 90 days.
It measures how well an AI agent can integrate with an API platform on its own, without a person reading the docs for it. Each platform is scored 1 to 5 on 7 dimensions: llms.txt, MCP server, tool discovery, OpenAPI quality, error responses, webhook setup and sandbox access. The AX score is the average of the verified dimensions.
The ranking only includes platforms whose 7 dimensions all have dated, linked evidence. 9 of 14 platforms qualify so far; the index shows the current order.
DX covers everything a developer touches, including search, code examples, changelogs and connector breadth. AX keeps the dimensions an agent depends on when no person is in the loop: machine-readable docs and specs, a hosted MCP server, tool enumeration, structured errors, reliable webhooks and self-serve credentials.
Each score rests on a dated record linking the vendor’s own public documentation, pricing or trust pages, scored against the rubric published on the methodology page. Where no source has been recorded, the dimension is Not verified: it is left out of the average, not scored low, and the platform is not ranked until every dimension is verified.
Every score is re-audited at least every 90 days, and sooner when a vendor ships a relevant change such as a hosted MCP server.
The same rubric applies to Apideck. Its verified score is below 5 on tool discovery, webhook setup and sandbox. The index lists what a 5 would take on each, with the evidence behind the current score.