What a listing does and does not establish, how evidence is handled, and how the protocols differ.
Source-linked research and editorial guidance. Capabilities are not independently tested unless a profile explicitly includes a reproducible test record. Sources checked November 13, 2026.
A proposed directory and research layer for discovering agent companies, products, MCP servers, skills, protocols and task services. This content release contains source-linked records and editorial evaluation guides.
No. The atlas distinguishes organizations, products, servers, skills, protocols and infrastructure. A listing is not a claim that an entity operates autonomously.
Each profile shows what was inspected. Most records summarize published sources. No runtime tests, security audits or payments were performed for this content package.
A company or project can operate several products. Devin is modeled separately from its operator Cognition. Funding belongs to the organization that raised it.
No. A vendor description is attributed evidence of what is offered or claimed. Performance requires a task, environment, version and observable result.
Editorial guidance: define a bounded task, required output, permissions and acceptance rubric. Compare candidates on the same fixture and include retries and review effort in cost.
The research did not establish the value. It does not mean the capability is absent or that the vendor refused to disclose it.
The inspected evidence is uneven. Concise directory records remain separate from expanded profiles, and missing people, funding or pricing are not invented.
This package records observations dated November 13, 2026. Individual source publication dates may be earlier. Recheck volatile facts before relying on them.
Not directly. Seat subscriptions, credits, browser hours, model tokens and accepted outcomes are different units. Review allowances, overages and extra infrastructure charges.
Not necessarily. The published plan may require paid model access, infrastructure, traffic or other services. Check the full workflow bill.
No. A funding round, lifetime funding, revenue and valuation are separate facts. This atlas records the amount and scope actually supported by the source.
Named people and current roles require supporting sources. This release leaves them unknown where those sources were not collected.
A service or package exposing capabilities through Model Context Protocol. Evaluate the exact tools, permissions, deployment and client compatibility.
Sources: modelcontextprotocol.io/docs/2026-07-28/getting-started/intro (opens in a new tab)
No. Directory presence is discovery evidence. Review upstream identity, code or documentation, permissions and behavior before granting access.
Do not assume so. A documentation endpoint may only retrieve reference information. Execution capabilities need separate evidence.
A registry detail page can describe a package hosted elsewhere. Resolve and record the actual upstream repository instead of rewriting the listing URL.
The exact endpoint or package, version, transport, authentication, scopes, tool inventory, side effects, installation source and evidence from any tests.
A skill supplies instructions and resources for doing work. A tool supplies a callable operation. A skill may instruct an agent to use several tools.
Compatibility must be established for the host, dependencies, permissions and task. A portable format alone does not establish successful execution everywhere.
Yes, as different entity types. A pack can contain many skills, but those records must not be counted as separate companies.
A2A describes communication and task coordination among agents. Its Agent Cards advertise identity, skills and interaction requirements.
Sources: a2a-protocol.org/latest/specification (opens in a new tab)
The full name is always required: Agent Client Protocol, OpenAI and Stripe Agentic Commerce Protocol, Virtuals Agent Commerce Protocol, or the historical IBM BeeAI Agent Communication Protocol.
Sources: agentclientprotocol.com/get-started/introduction (opens in a new tab), github.com/agentic-commerce-protocol/agentic-commerce-protocol (opens in a new tab), os.virtuals.io/acp/overview (opens in a new tab), ibm.com/think/topics/agent-communication-protocol (opens in a new tab), lfaidata.foundation/communityblog/2025/08/29/acp-joins-forces… (opens in a new tab)
x402 describes HTTP payments for internet resources. The service endpoint, supported network, asset and payment implementation still need to be checked.
Sources: github.com/x402-foundation/x402 (opens in a new tab), x402.org (opens in a new tab)
The inspected ERC-8004 proposal concerns identity, reputation and validation registries; payments are outside its stated scope.
Sources: eips.ethereum.org/EIPS/eip-8004 (opens in a new tab)
No. The proposed job interface includes an evaluator, but the evaluator may be the client. Record the relationship before describing an evaluation as independent.
Sources: eips.ethereum.org/EIPS/eip-8183 (opens in a new tab)
No. Identity, reputation, capability and accepted delivery are distinct evidence types.
No. Connect payment evidence to a deliverable and a separate acceptance decision.
No. Testnet observations should be labeled separately from production settlement and revenue.
It contains research records for task services and marketplaces. Live availability, terms, provider identity and settlement behavior have not been tested here.
The supplied related links use editorial topic relevance. They are not rankings or verified feature parity. A final comparison should use the same task and constraints.
No tested ranking is supplied. The initial site should use transparent filters and evidence labels rather than invented scores.
A trace records observed execution events. Acceptance evaluates whether the final result satisfies the task. Both can be useful, but they answer different questions.
Editorial policy: show one only when a real assessment identifies its subject, version, date, method, findings and scope. This package supplies no audit badges.
Proposed workflow: submit the profile URL, field to change, corrected value and an official supporting source. A reviewer should retain the previous value and the reason for the change.
Proposed workflow: provide the official URL, entity type, description, documentation, available interfaces and evidence. Submission should enter review rather than publish automatically.
Proposed workflow: verify control of the official domain or another authoritative account. Profile ownership should not let an operator overwrite independent evaluation evidence.
No. The supplied launch controls keep incomplete records out of the initial indexing set. Expanded pages and topic guides still require editorial and technical review.
No. The content and architecture are editorial and engineering recommendations. They do not establish future rankings, indexing or citations.
This atlas content is for discovery and evaluation. A source-linked listing does not constitute an endorsement or a finding of fitness for a particular task.