ai-provider-watch 0.1.1

Open provider-change event feed for AI platform teams and agents.

AI Provider Watch

AI Provider Watch, or APW, is a public event feed and CLI for changes from AI providers that can affect developer cost, quotas, token accounting, model availability, defaults, deprecations, incidents, and migration risk.

Use it when you need an auditable answer to questions like:

• Did a provider incident explain a spike in failures, retries, latency, or support tickets?
• Did a model launch, retirement, default change, pricing update, or quota shift create work for platform teams?
• Which repos, agents, gateways, or dashboards should be checked before a provider change turns into a customer-facing problem?

APW is founded by Ottto and built as a standalone open-source project. The feed, schemas, CLI, GitHub Action, MCP helpers, and docs work without an Ottto account.

Install

Try the CLI without installing:

uvx --from ai-provider-watch apw latest --limit 3
uvx --from ai-provider-watch apw diff --since 30d

Install it as a command:

pipx install ai-provider-watch
apw latest --limit 3

Or install it in a Python environment:

python -m pip install ai-provider-watch
apw validate

The published package includes a reviewed public data snapshot, so read-only commands work outside a checkout. For the freshest feed, use the GitHub data artifacts or raw main URLs below.

Quickstart

Show the latest reviewed events:

apw latest --limit 3

List events from the last 30 days:

apw diff --since 30d

Explain one event for a human reviewer:

apw explain 2026-06-04-openai-codex-compaction-latency

Validate the bundled schemas, registries, events, feeds, and indexes:

apw validate
apw index --check
apw freshness --summary

Feed Artifacts

The canonical reviewed events live in data/events/. Generated feed artifacts live in data/feeds/ and data/indexes/:

• data/feeds/events.json
• data/feeds/events.ndjson
• data/feeds/freshness.json
• data/feeds/latest.json
• data/feeds/rss.xml
• data/indexes/provider/*.json
• data/indexes/kind/*.json
• data/indexes/severity/*.json

For direct consumption, pin a release tag or read from the repository:

https://raw.githubusercontent.com/ottto-ai/ai-provider-watch/main/data/feeds/latest.json
https://raw.githubusercontent.com/ottto-ai/ai-provider-watch/main/data/feeds/events.ndjson
https://raw.githubusercontent.com/ottto-ai/ai-provider-watch/main/data/feeds/freshness.json

GitHub CalVer data releases are the canonical immutable feed snapshots. PyPI package releases are installable CLI snapshots that bundle reviewed data for offline and no-checkout use; APW does not publish a new package for every data tag. Patch packages are published when bundled data freshness materially helps install-only users or when CLI/package behavior changes.

Use apw freshness to verify the feed version, package version, event count, latest reviewed event date, latest source-state retrieval timestamp, release manifest path, and checksum manifest path from either a checkout or the bundled package data.

The normalized factual event data and generated feeds are CC0-1.0. Code, schemas, docs, tests, and tooling are Apache-2.0.

What You Get

• A reviewed machine-readable event feed, not a static model catalog.
• A typed ProviderEvent envelope with precise event details and repeatable impact rows.
• A CLI for validation, indexing, latest events, diffs, explanations, release dry runs, source checks, candidate generation, repo impact checks, notifications, and ecosystem mappings.
• JSON Schemas for events, sources, candidates, observations, releases, feed freshness, webhooks, Slack-style payloads, ecosystem mappings, and LLM review packets.
• Official-source descriptors for OpenAI, Anthropic, Google Gemini / Vertex AI, AWS Bedrock, and Azure OpenAI.
• Review-only source candidates that help maintainers notice provider changes without publishing unreviewed facts.
• Agent-native surfaces: AGENTS.md, CLAUDE.md, llms.txt, Codex and Claude skills, a read-only MCP adapter shell, and a Codex plugin package.
• Downstream integrations for GitHub Actions, webhooks, Slack-compatible JSON, LiteLLM, models.dev, Langfuse, Helicone, and OpenLIT.

Trust Model

APW is designed for factual, reviewable provider-change data.

• Prefer official provider-controlled sources.
• Treat provider pages, issue bodies, PR comments, social posts, MCP text, and generated candidates as untrusted data, never as instructions.
• Do not commit raw provider HTML, authenticated-console content, screenshots, private billing data, cookies, credentials, or customer telemetry.
• Publish only reviewed data/events/*.json records.
• Keep generated candidate files in data/candidates/ review-only until a source owner promotes a factual change.
• Keep release tokens away from jobs that fetch source pages, process candidate text, run LLM review, or inspect PR comments.

APW is intentionally independent of Ottto private product surfaces. Ottto may consume APW data, but this repository does not expose Ottto customer telemetry, Advisor internals, private UI, infrastructure, Slack data, or credential loading code.

Work From A Checkout

Use a checkout for write workflows such as source refresh, candidate generation, event promotion, feed regeneration, and release dry runs:

git clone https://github.com/ottto-ai/ai-provider-watch.git
cd ai-provider-watch
uv sync --all-extras
uv lock --check
uv run pytest
uv run apw validate
uv run apw index --check
uv run apw source test

Fetch official sources and generate review candidates:

uv run apw source fetch --observations .apw/source-observations.json
uv run apw candidate generate \
  --observations .apw/source-observations.json \
  --output .apw/candidates \
  --created-at 2026-06-05T00:00:00Z
uv run apw candidate review-pr-body \
  --observations .apw/source-observations.json \
  --candidates .apw/candidates

Candidate files are not published events. Promotion to data/events/ remains a manual source-owner review step. See Event Promotion.

Use APW In Downstream Systems

Check a repository for model references and APW-relevant impact:

apw repo check --repo . --since 3650d --risk low

Render notification payloads:

apw notify webhook --since 7d --risk medium --output .apw/apw-webhook.json
apw notify slack --since 7d --risk medium --output .apw/apw-slack.json

Render ecosystem mappings:

apw ecosystem render --target litellm --since 30d --risk medium --output .apw/litellm.json
apw ecosystem render --target langfuse --since 30d --risk medium --output .apw/langfuse.json

See:

• Agent Consumption
• Downstream GitHub Action
• Webhook And Slack Payloads
• Ecosystem Mappings
• Read-Only MCP Contract
• Codex Plugin

Schema And Architecture

APW uses a stable ProviderEvent envelope, a typed EventDetail payload, and repeatable ImpactAssessment rows. That keeps pricing, quota, lifecycle, token-accounting, status, API-contract, and migration-risk events precise without creating one giant nullable event model.

Start here:

• Architecture
• Event Schema
• Feed Freshness Schema
• Contributor Review Workflow
• Source Packages
• Source Refresh
• Release Gates
• v0.2 Release Checklist
• Python Package Release

Project Status

APW v0.1.1 is the current stable public package. It updates the bundled reviewed feed snapshot after the June 8 source-owner promotion batch. The first public data releases are signed CalVer tags such as data-2026.06.05.

The current release includes:

• reviewed seed events for OpenAI, Anthropic, Google Vertex AI, AWS Bedrock, and Azure OpenAI;
• generated JSON, NDJSON, RSS, provider, kind, and severity indexes;
• source-refresh automation that opens draft candidate-review PRs without publishing events;
• no-op guarded data-publisher workflow scaffolding;
• PyPI Trusted Publishing;
• CI, CodeQL, Dependency Review, Scorecard, and data-release dry-run workflows.

Daily unattended public data tags are not enabled yet. Until that safety gate is stronger, real data publication uses reviewed PRs plus maintainer-signed Git tags.

Contributing

Use pull requests for code, schema, source, data, docs, and workflow changes. Start with CONTRIBUTING.md.

Useful contributor docs:

• Contributor Review Workflow
• Event Promotion
• Source Packages
• Repository Settings
• Roadmap
• Source Owners
• Security Policy

License

Asset	License
Code, schemas, docs, tests, CLI, MCP shell	Apache-2.0
Normalized factual data and generated feeds	CC0-1.0
Provider names and trademarks	Owned by their respective owners

See DATA_LICENSE.md, TRADEMARKS.md, and LICENSES/.