# CLI, Web App or Library? An Interface Census of 93 AI Agents

> Source: The Agents Index — https://theagentsindex.com/blog/ai-agent-interface-census-2026 (sourced research)
> By The Agents Index Editorial · Published 2026-08-12 · Updated 2026-08-19

_We recorded the primary interface of 93 of the 94 AI agents in our index on 20 August 2026. All 14 library-first tools self-host; 4 of 37 web apps do._

**Of the 93 AI agent tools in The Agents Index with a recorded primary interface on 20 August 2026, 37 lead with a web app, 14 with a code library, 12 with a command-line tool, 11 with a visual builder, 10 with an IDE extension, and 9 with an API.** Every listing carries an `Interface` field in its datasheet, set to one of those six values.

The index holds 94 published listings. One of them, [Agentix Labs](/agentix-labs), is a managed-service agency rather than a piece of software, so the question of a primary interface does not apply to it and the field is empty. Every count on this page is out of 93, not 94, and never out of the whole index.

The field sits in the Capabilities group next to Model / LLM and Public API, and it answers a question buyers ask before almost any other. What am I going to be looking at when I use this thing: a terminal, a browser tab, an import statement, or a canvas of boxes and arrows?

## The headline count

| Primary interface | Tools | Share |
|---|---:|---:|
| Web app | 37 | 40% |
| Library | 14 | 15% |
| CLI | 12 | 13% |
| Visual builder | 11 | 12% |
| IDE extension | 10 | 11% |
| API | 9 | 10% |

**The single most common way to touch an AI agent is still a browser tab: 37 of the 93 tools (40%) present a web app first.** Shares are rounded to the nearest point and sum to 101 on this date.

The developer-facing surfaces, meaning library, CLI and IDE extension together, come to 36 of 93. At our previous count that group was exactly level with the web apps, and we said then that we read the tie as coincidence. It has now broken, and it broke in the direction of the browser.

Counting API-first tools as developer-facing too, which is defensible, puts the developer side at 45 of 93 and ahead. Either way the category is usually discussed as though it were uniformly a developer market, and by our count it is about half one.

## What the interface predicts: whether you can run it yourself

The distribution is the least interesting result here. The interesting one is that this field, which describes nothing but ergonomics, sorts almost perfectly against a field that describes infrastructure.

**Of the 93 tools, 41 (44%) offer some path to running the software somewhere you control.** Those 41 are not spread evenly across the six interfaces:

| Primary interface | Offers a self-host path |
|---|---:|
| Library | 14 of 14 (100%) |
| CLI | 9 of 12 (75%) |
| IDE extension | 6 of 10 (60%) |
| Visual builder | 6 of 11 (55%) |
| API | 2 of 9 (22%) |
| Web app | 4 of 37 (11%) |

**All 14 library-first tools in our index can be run on infrastructure you control. 4 of the 37 web-app-first tools can.** Those are the two ends of the index, 89 percentage points apart on a field that is supposed to be about ergonomics.

The library end is close to tautological, and we said so when our deployment census found the same thing from the other direction. A framework you cannot run yourself is a managed service with extra steps.

The web end is the part that is actually informative. A web app is no harder to ship as a self-hosted product than a library is, and plenty of software does exactly that, so 4 of 37 looks like a commercial choice repeated 33 times rather than an engineering constraint.

The practical version, for anyone with a data-residency requirement: the interface is visible from the vendor's homepage in about four seconds, and it will tell you the answer before you open the pricing page. If the product is demoed as a browser tab, the odds we measured are roughly 1 in 9 that a private deployment exists.

## The exceptions, since four of thirty-seven is small enough to name

The four web-app-first tools in our index that offer a self-host path are [OpenHands](/openhands), [CodeRabbit](/coderabbit), [Sim](/sim) and [Vellum](/vellum-ai). Three of those four, OpenHands, Sim and Vellum, are also the only open-source tools among the 37 web-app-first listings, and OpenHands publishes its source [on GitHub](https://github.com/All-Hands-AI/OpenHands).

Vellum is new to both of those lists since our previous count, because its licence and deployment records were corrected on 19 August rather than because the product changed. CodeRabbit remains the one entry that offers a private deployment without publishing its source, which is a combination worth knowing exists.

At the other end of the index, [Claude Agent SDK](/claude-agent-sdk) is the one library-first tool that is not open source.

A wrinkle sits in the CLI bucket. Three of the 12 CLI tools are recorded as cloud-only despite running in your terminal: [Antigravity CLI](/antigravity-cli), [Claude Code](/claude-code) and [Muse Code](/muse-code). Installing locally and running locally are different things. Each of those three executes on your machine and requires the vendor's hosted service to function, so a terminal prompt is not evidence of a private deployment. That is exactly the assumption this pairing of fields is useful for checking.

## Open source, counted against what we could establish

**We could determine open-source status for all 93 tools with a recorded interface, and 33 of the 93 are open source.** The visual-builder gap that this page carried at its previous two counts has been researched and closed, so there is no undetermined row in the table below.

| Primary interface | Open source | Determined |
|---|---:|---:|
| Library | 13 | 14 of 14 |
| CLI | 8 | 12 of 12 |
| Visual builder | 4 | 11 of 11 |
| IDE extension | 3 | 10 of 10 |
| Web app | 3 | 37 of 37 |
| API | 2 | 9 of 9 |

**Open source tracks the interface the same way self-hosting does: 13 of the 14 library-first tools are open source, against 3 of the 37 web-app-first tools.** Every denominator in this table is the full size of its bucket, which was not true when we last counted.

## API-first is not a voice-agent story alone

**There are 9 API-first tools in the index, and 5 of them are voice agents.** The other four are [Browserbase](/browserbase) and [SalesTouch](/salestouch), which are agent tools rather than agents, plus [Devin](/devin) in coding agents and [TinyFish](/tinyfish) in research agents.

Voice is still where API-first concentrates: 5 of the 11 voice agents lead with an API, and no other category is close to that rate. Voice gets bought as a component to embed in something else rather than as a destination product, and the interface field is where that shows up.

What the API bucket does not do is offer a private deployment. Only 2 of the 9 API-first tools do, a rate second only to the web apps, and it has now held across two consecutive counts rather than being the one-off we took it for.

## The category split

| Category | Web | Library | Visual | CLI | IDE | API |
|---|---:|---:|---:|---:|---:|---:|
| Coding agents (25) | 3 | 0 | 1 | 10 | 10 | 1 |
| Agent platforms (15) | 6 | 0 | 8 | 1 | 0 | 0 |
| Agent frameworks (14) | 0 | 13 | 0 | 1 | 0 | 0 |
| Voice agents (11) | 4 | 0 | 2 | 0 | 0 | 5 |
| Research and data agents (11) | 9 | 1 | 0 | 0 | 0 | 1 |
| Support agents (9) | 8 | 0 | 1 | 0 | 0 | 0 |
| Sales and marketing agents (8) | 8 | 0 | 0 | 0 | 0 | 0 |
| Agent tools (2) | 0 | 0 | 0 | 0 | 0 | 2 |
| Agencies (1) | 0 | 0 | 0 | 0 | 0 | 0 |

Sierra and Botsify each belong to two categories and are counted once in each, so the category rows sum to 96 memberships against 94 distinct published listings. The agencies row is empty across all six columns because its one member is the listing with no interface recorded.

**13 of the 14 agent frameworks are library-first, and for the first time the category has an exception.** [AgentsKit](/agentskit) leads with a command-line tool. At our previous count this row read 14 of 14, and the clean sweep was part of what made the framework category legible as a category at all. It is worth watching whether a second framework follows rather than treating one row as a trend. The only library-first tool outside the category is [GPT Researcher](/gpt-researcher), filed under research agents.

**All 8 sales and marketing agents are web apps**, and support agents are 8 of 9. Neither category contains a single tool with a self-host path: 0 of 8 and 0 of 9 respectively, re-derived here on the same date. They are sold as browser products to buyers who were never going to run infrastructure, and two of our fields say so independently.

Coding agents remain the most fragmented category in the index, spread across five interfaces: 10 CLI, 10 IDE, 3 web, 1 visual and 1 API. That is not indecision on our part in classifying them. It reflects a genuine and unresolved split about where a coding agent belongs, whether in the terminal, inside the editor, or in a browser tab away from both. The terminal and the editor have now been exactly level at 10 each across two counts.

## Method, and the judgement call

Counted on 20 August 2026 against the 94 listings published in The Agents Index at that moment. The `Interface` field is a fixed six-value enum (`CLI`, `IDE`, `Web`, `API`, `Visual`, `Library`), populated for 93 of those 94, and it appears on every listing's public datasheet under Capabilities. Any single tool in this report can be checked against its own page.

The one empty value is Agentix Labs, an agency. We have not treated that as a research gap and it is not counted as a "no" anywhere on this page, because a services firm has no primary interface to record. If the index adds more agencies, this denominator will separate further from the published total.

**The judgement call is that the field records one primary interface, and not every surface a tool offers.** Many of these tools have more than one. Claude Code is recorded as CLI although other surfaces exist; Cursor is recorded as IDE although it ships a desktop application. The test we apply: what is the surface the vendor leads with, and the one through which a new user does the main work? Where a tool genuinely leads with two, we record the one its own documentation puts first.

This is a defensible rule rather than an inevitable one, and a reader applying a different rule would get slightly different numbers in the CLI and IDE buckets specifically. It would not move the web-versus-library finding, which is where the margin is widest.

The self-host column counts a tool as offering a path if its `Deployment` field is `Self-host` or `Both`. Our [deployment census](/blog/ai-agent-deployment-census-2026/) counts that same field from the other direction and carries its own measurement date, so its totals describe the corpus as it stood when it was last counted rather than as it stands here. The open-source column is the `Open source` boolean from the same datasheets.

Two things this report does not claim. It is a census of our index rather than a survey of the market: we count what we hold, and there are AI agent tools in the world we have not listed. And the correlation between interface and deployment describes these 93 tools rather than stating a law. It tells you what the odds looked like in August 2026, which is useful for narrowing a shortlist and is not a substitute for checking the vendor.

Corrections and additions: [write to us](/write-for-us/).

---

Tags: ai-agents, developer-tools, self-hosted, open-source, market-analysis, data-report