AI Systems Atlas

AI Systems Atlas is a curated directory of operational AI systems, provider-independent model releases and the labs that develop them, the specifications that connect systems, managed inference services, runtimes you host yourself, unscored agent packs a host agent installs, and unscored AI robots recorded for what their makers document. One Directory surface discovers systems, the complete models.dev source catalog, inference services, and local runtimes while preserving the distinction between attributed source metadata and Atlas-reviewed records. Models and Labs are specialist sibling views, and Specifications remains a separate artifact view.

Browse the published directory at peacefulcoexistance.com. Every reviewed record is also published as JSON with no key and no rate limit — the API page lists each file, and llms.txt says the same thing for an agent.

The repository contains the canonical editorial catalog, its validation and refresh automation, and a dependency-free static web interface.

Start here

uv sync --locked
npm ci
npx playwright install chromium
uv run ruff check scripts tests
uv run python scripts/validate_directory.py
uv run python -m unittest discover -s tests -v
node --test tests/test_web.js
npm run lint:js
npm run test:e2e
uv run python scripts/serve_web.py 8765

Open http://127.0.0.1:8765.

Sizing a change? Measure it rather than reading a figure out of a document, which is stale within days:

uv run python scripts/measure_engineering.py
uv run python scripts/build_web_payload.py --counts

Choose your path

Read only what your task needs:

Goal	Read next
Understand direction and sequencing	ROADMAP.md
Understand families, roles, and scores	docs/TAXONOMY.md
Add or understand a protocol, convention, or format	docs/SPECIFICATIONS.md
Add or understand a managed model-inference service	docs/INFERENCE_SERVICES.md
Add or understand a self-operated inference runtime	docs/LOCAL_RUNTIMES.md
Understand coverage and choose a research batch	docs/COVERAGE.md
Add, remove, classify, or rescore a project	docs/CURATION.md
Understand JSON fields and timestamp semantics	docs/DATA_MODEL.md
Run refreshes or review candidate/license incidents	docs/OPERATIONS.md,

AI coding agents should begin with AGENTS.md, which provides the same routing in an execution-oriented form.

Repository map

directory/projects.json         reviewed projects and editorial scores
directory/taxonomy.json         enum definitions, license catalog, score profiles
directory/license-evidence.json reviewed source paths and immutable blob evidence
directory/specifications.json  reviewed unscored protocols, conventions, and formats
directory/inference-services.json reviewed and separately scored managed inference services
directory/local-runtimes.json  reviewed and separately scored self-operated inference runtimes
directory/models.json         reviewed and separately scored model releases
directory/models-dev.json     complete commit-pinned models.dev source snapshot
directory/packs.json          reviewed unscored agent packs a host agent installs
directory/labs.json           reviewed unscored labs that develop the reviewed model releases
directory/robots.json         reviewed unscored AI robots, recorded for what their makers document
directory/model-candidates.json unpublished text-output model review queue
directory/openrouter-model-leads.json unpublished OpenRouter cross-check leads for models.dev gaps
directory/openrouter-model-dispositions.json OpenRouter terms review and lead decisions
directory/exclusions.json       reviewed family/role boundary decisions
directory/candidates.json       durable provisional discovery queue
directory/license-review.json   unresolved license-evidence review queue
directory/discovery-sources.json allowlisted official discovery feeds
directory/hn-signals.json       unpublished Hacker News attention-source signal queue
scripts/                        refresh, synchronization, validation, and measurement
tests/                          Python invariants, web logic tests, and browser E2E tests
web/                            static directory UI and published data copies
docs/                           task-focused policy, model, operations, and ADRs

Trust model

• Operational relevance determines inclusion; source model and licenses are reviewed traits and filters.
• Authoritative license and terms sources are human-reviewed; GitHub metadata is only a drift signal.
• A detected license mismatch marks evidence for review without hiding the project or rewriting human conclusions.
• Automated model-source records are visibly attributed and receive no editorial score or review date; workflow candidates remain unpublished.
• OpenRouter's public model list is an unpublished cross-check for models.dev gaps: its leads are pointers, never evidence, and nothing is fetched before a recorded terms review.
• Editorial verification dates are separate from live GitHub metadata dates.
• Memory, agent, and assistant scores are never ranked across families.
• Specifications are classified by type, integration scope, and status; they are never operationally scored.
• Inference services are classified by service boundary, delivery, model sources, and API style, then scored only on stable operational-service traits; their scores are not comparable to system-family scores.
• Labs are unscored records of the organizations that develop reviewed releases. A lab joins its releases, services, runtimes, specifications, and packs by the names the catalog already uses for it, and never carries a licence, score, or rank.
• Local runtimes are classified by runtime type, accelerators, model formats, serving modes, and deployment surfaces, then scored only on documented execution capability; their scores exclude throughput, latency, and benchmark rank and are not comparable to any other profile.

The weekly workflow refreshes live metadata, preserves candidate and license-review queues, validates the complete catalog, runs tests, and commits only verified data changes.

License

The repository carries two licences, because the code and the catalog are different kinds of work.

• Software — scripts/, web/*.js, web/*.css, tests/ — is Apache-2.0.
• The curated catalog — reviewed directory/*.json data and its synchronized web/*.json copies — is CC BY 4.0. Share it and adapt it, including commercially; credit the Atlas and link the licence. The derived models-dev.json snapshot remains attributed to models.dev under its MIT License, preserved in third_party/models.dev-LICENSE.txt.

Project names, descriptions, and marks remain the property of their respective projects and are used for factual identification and commentary. Linked licence and terms documents stay under the terms of the projects that published them.