reachjalil/canmyagentuse

Can My Agent Use — AI agent compatibility catalog

Open data from canmyagentuse.com, the independent compatibility catalog for AI agents. It answers one question — can my agent use this? — for exact web, desktop, and CLI agent products, and links every published answer to dated public evidence. Cells without a reviewed source stay unknown.

• —Catalog (HTML): https://canmyagentuse.com
• —Agent entry point: https://canmyagentuse.com/llms.txt
• —Canonical agent prompt: https://canmyagentuse.com/prompt.txt
• —JSON API: https://canmyagentuse.com/api/v1/openapi.json
• —Press kit: https://canmyagentuse.com/press
• —GitHub mirror: https://github.com/reachjalil/canmyagentuse-data

Dataset summary

The default cells config holds 4,557 compatibility cells: one row per atomic capability × exact product surface, across 49 capabilities (MCP client/server roles, skills, plugins, model selection, and more), 93 exact surfaces (web, desktop, CLI) from 60 products by 51 vendors, backed by 959 assertions over 183 dated public sources. Research cutoff: 2026-08-28.

The repository also carries the catalog's curated JSON API exports under data/api/ (features, harnesses, specifications, support matrix, evidence ledger, coverage, conformance tests, changes, contradictions, and the OpenAPI description) and the full frozen snapshot at data/atlas.json.

Fields (cells)

vendor_id, vendor_name, product_id, product_name, surface_id, surface_name, surface_type (web · desktop · cli), version_or_observation, surface_availability, coverage_tier, feature_id, feature_group, feature_name, primary_status (supported · partial · unsupported · unknown), status_detail, assertion_count, evidence_methods, confidence, environment_scopes, qualifiers, source_keys, notes, last_verified_at.

Reading the data honestly

• —Exact surfaces. A row describes one product surface. Never generalize one surface's answer to another surface or to the provider's product family.
• —Unknown is unreviewed, not unsupported. unknown means no public evidence has been reviewed for that cell.
• —Qualifiers survive. Plan, platform, region, configuration, transport, and rollout limits stay attached to the answer.
• —Evidence classes are literal. documented, vendor-attested, listed, tested, reported, inferred, and not-found describe different kinds of support.
• —Coverage is research progress, never usage share or a product score.

Quick start

from datasets import load_dataset

cells = load_dataset("reachjalil/canmyagentuse", "cells", split="train")
mcp_clients = cells.filter(
    lambda row: row["feature_id"] == "mcp.client" and row["primary_status"] == "supported"
)
print(mcp_clients.num_rows, "surfaces with a documented MCP client role")

Point an agent at the catalog

Paste the canonical prompt from https://canmyagentuse.com/prompt.txt (also in this repo as prompt.txt) into any assistant, system prompt, or agent configuration. It routes capability questions through the live catalog — retrieval order, status semantics, and citation rules included. Agent runtimes that load skills can install SKILL.md directly.

Freshness and corrections

This dataset is a point-in-time export (research cutoff 2026-08-28). The live catalog at canmyagentuse.com is the canonical, most current record — cite it, and include the verification dates it reports. Corrections go through the catalog's report route.

License and attribution

CC BY 4.0. Attribution format:

Data: Can My Agent Use (canmyagentuse.com), CC BY 4.0.

Can My Agent Use is an independent publication. Product names and marks belong to their respective owners and are used only to identify the products covered by the catalog. A listing is not an endorsement, and no certification, approval, partnership, or vendor affiliation is implied.

Citation

@misc{canmyagentuse2026,
  title        = {Can My Agent Use — AI agent compatibility data},
  author       = {{Can My Agent Use}},
  year         = {2026},
  url          = {https://canmyagentuse.com},
  note         = {Research cutoff 2026-08-28. CC BY 4.0.}
}