Conversation
Phase 3, first slice. Everything here is content that belonged to Portkey the company rather than to the gateway product. Welcome page rewritten: the YouTube embed, the star callout, the edge-worker locations, the ISO/SOC2 line, the managed-hosting paragraph, the 25M-requests and Status Page block, the pricing accordion and "where can I reach you" are all gone. The OpenAI examples no longer route through the Portkey SDK. Open-source repo marketing removed — the `## Open Source` section, the "give us a star" callouts, the GitHub contribute cards and the PR links. Operational references stay: the Helm chart name `portkey-ai/gateway` and the self-hosting deployment instructions are things a reader actually needs. Model counts normalised to 3,000+ across 55 files; the 250+/200+/1600+ variants disagreed with each other. Bare provider counts (50+, 200+) are left alone — they are a different number and I do not have the real one. Integrations: the Ecosystem group is gone, `ecosystem.mdx` becomes `overview.mdx` titled Overview under an Integrations group, "Preferred Partners" becomes "Integrations", and the become-a-partner scheduling iframe and the `partner.mdx` contributing-guide redirect are removed. Marketing links swept: 13 `portkey.ai/pricing` plan links unwrapped, 13 `portkey.ai/docs` links repointed at their `/aigw/` equivalents, 10 that had no equivalent unwrapped, plus the blog, status and trust portal links. `security-portkey.mdx` keeps its structure but no longer asserts Portkey's compliance posture or links Portkey's legal pages — those need Palo Alto's. Also fixed: 29 "the AI Gateway's gateway" stutters left by the Phase 2 rename, the orphaned Prompt Library section on the features page, and the 3 navigation group icons. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 3, item 16. The two pages described the same system from different angles: `architecture.mdx` had the infrastructure detail (Helm, storage backends, auth methods, resource requirements, encryption, rationale) but described a generic hybrid deployment; `scm-architecture.mdx` had the accurate two-plane model, component roles, identity flow, request lifecycle and storage ownership but no deployment detail. One page now, titled Architecture, built on the SCM two-plane model as the spine with the infrastructure and security sections folded in. The mermaid diagram is redrawn to show both planes and the SCM API Gateway boundary rather than a generic "Control Plane" box. Dropped along the way: the two portkey.wiki demo-booking CTAs, the sample log/metric file cards pointing into this repo's stripped image tree, and two deployment-guide cards whose routes no longer exist. The Next Steps cards now point at the deployment guides that do. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 3, item 4. Where a request already carries an integration slug, `x-portkey-provider: @openai-prod` plus `"model": "gpt-4o"` becomes `"model": "@openai-prod/gpt-4o"` and the header goes away. 60 code blocks converted across 34 files, in cURL, the OpenAI SDKs and LangChain — the `default_headers` / `defaultHeaders` argument is dropped entirely where the provider header was its only entry. Three cases keep the header, deliberately: - **Bare provider names** (83 blocks). `x-portkey-provider: openai` is the direct-auth path where `Authorization` carries the provider's own key. There is no slug to fold into the model name. - **Endpoints with no model parameter** (106 blocks). Files, Batches and fine-tuning select the provider by header because the request body has nowhere else to put it. The two batch blocks the first pass converted are restored — the model there sits inside `requests[].params`, which is a provider-native body, not the gateway's. - **`headers.mdx`**, which documents the header itself. It now carries a note saying the model-prefix form is the one to reach for first, and when the header is still the answer. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 3, item 12. The old page was three cards and a video that pointed readers at the Portkey SDK and an api-reference route that does not exist in this version. It never actually set anything up. The new page mirrors the SCM setup flow end to end, in Strata Cloud Manager's own navigation vocabulary: Admin Settings → Plugins for the Prisma AIRS key, Workspace Control, Integrations → Add LLM Integration, Catalogs → Create Guardrail, Profiles & Policies → Config, Security Keys → Create New, then Observability → Logs and AI Runtime → AI Sessions on the Prisma AIRS side. It uses OpenAI rather than Vertex AI, since the point is the shortest path to a working request, and ends with a runnable call in cURL, OpenAI Python and OpenAI Node — authenticating with `Authorization: Bearer`, selecting the provider through `model: "@openai-prod/gpt-4o"`, and inheriting the guardrail from the config attached to the API key. Identifiers are placeholders (`pg-xxxxxx`, `pc-xxxxxx`). The setup guide's real ones are internal. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The 433 blocks that imported createHeaders / PORTKEY_GATEWAY_URL from portkey-ai now use the OpenAI SDK on its own: the gateway base URL is inlined, the gateway key moves to api_key/apiKey, and residual headers become a literal default_headers / defaultHeaders map. Where the provider was an @slug and the block had a model to carry it, the slug folds into the model parameter; otherwise it stays as x-portkey-provider. Object-valued config and metadata headers are serialised with json.dumps / JSON.stringify, which createHeaders used to do implicitly. Verified with a per-language syntax check of every fenced block against HEAD: no new Python or JavaScript parse failures. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deletes the 867 fenced blocks that construct a Portkey SDK client and the 148 orphaned blocks that called methods on one, leaving REST, OpenAI Python and OpenAI Node as the example set. Tab, Tabs and CodeGroup wrappers left holding nothing are removed with their contents; no page is left without a runnable example. These examples come back in Phase 6. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Picks up the 89 blocks the first converter skipped — those with a comment in front of a key, or an object-valued metadata / config header — plus the last 21 helper references, handled individually: literal header maps for the Bedrock Files, Strands, LangChain JS and Configs examples, the model parameter for the Agent Builder provider-switching snippets, and reworded prose where createHeaders was only mentioned. No portkey-ai helper remains in the docs. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The pages no longer construct a Portkey SDK client, so the dependency is dead weight. Lines where portkey-ai was the only package now install the OpenAI SDK instead. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Drop the 58 "SDK Reference" cards that pointed at the Portkey SDK docs (they were also broken links inside the aigw version). - Rename the remaining "AI Gateway SDK" headings and prose to the OpenAI SDK / REST wording the examples now use. - Delete the emptied Python/NodeJS tabs left behind by the SDK removal, plus the "Coming Soon!" placeholder tabs on aws-bedrock. - Rewrite the tracing examples in traces.mdx to REST + OpenAI Python + OpenAI NodeJS, with the trace/span IDs sent as headers. - Rebuild llama-agents.mdx: the examples had no gateway API key and still used the old config-with-api_key form. - api-reference introduction now documents two integration paths instead of three. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Comments inside code fences now say "AI Gateway" / "Strata Cloud Manager" instead of "Portkey" / "Portkey App" / "Portkey dashboard", and sample variables named after the vendor (portkey, portkeyClient, portkeyConfig, portkey_handler, the C# PortkeyAzureClient helpers) are renamed to gateway-neutral names. Functional identifiers are untouched: PORTKEY_API_KEY, x-portkey-* headers, the aigw.portkey.ai host, package names, the promptfoo portkeyConfig key, and everything under self-hosting/ and changelog/. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Rewrites prose references to the retired SDK surface (prompts completions, rerank, feedback), repairs the garbled "Production-ready" headings on four agent pages, retargets six anchors whose headings were renamed earlier in Phase 3, and de-brands sample resource names in the library and MCP-server setup steps. Fixes the $@YOUR_PROVIDER_SLUG typo in the batches cURL example. Functional identifiers (repo URLs, config keys, JWT claims, OTel attributes, the @portkey-mcp prefix, portkey_options) are left as-is. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reorders 115 CodeGroup and Tabs blocks across 82 integration pages so the cURL/REST tab leads, with the OpenAI SDK examples following. Pure reordering: the set of lines in every touched file is unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Drops the x-portkey-provider header from 56 chat/embedding/image/audio examples and moves the slug into the model value (model: "@openai/gpt-4o"), so Authorization is the only header those requests need. Left in place where the header is load-bearing: requests with no model param (files, batches, fine-tuning), custom-host and ad-hoc provider credential examples, the gateway-native "portkey" batching sentinel, and blocks that set sibling x-portkey-provider-* headers. Also corrects the @openai prefixes that would have landed on Bedrock models in bedrock/embeddings.mdx, and a gpt-4bo typo in authentication.mdx. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
We cannot point docs.json at the Palo Alto Networks GitHub organisation, so aigw/airs-openapi.yaml is a committed mirror of the spec repo's main rather than a remote fetch. The mirror is flattened: overlays/docs-prose.yaml is applied at sync time, so the docs build consumes one self-contained file while structure and prose stay split upstream where that split earns its keep. Mirrored from openapi@84b8bb9 -- 181 operations, 118 paths, 41 tags. aigw/scripts/sync-openapi.sh is the only supported way to refresh it. It refuses to run against a branch other than main or a dirty tree, so a mirror is always reproducible from a named commit. aigw/scripts/wire_openapi_nav.py generates the navigation from the spec repo's docs-navigation.json. Inference's 17 tags go to the Gateway APIs tab; the 24 control-plane tags go to Admin APIs. It is idempotent -- it replaces the generated groups rather than appending, keeping the hand-written API Reference group first. Prompt endpoints need no exclusion here: Prompts, Prompt Partials, Labels and Collections are already in the spec repo's drops.yaml, and its build fails if one reappears. mint validate is green and resolves the local mirror. Not verified locally, and worth knowing: group-level OpenAPI generation produces no pages under mint dev 4.2.893 on Node v22.23.0. The preview logs a recurring undici TypeError and renders only the 9 hand-written pages. This is not a fault in this configuration -- the known-good remote spec generates nothing through the same group-level shape either. Confirm on a Mintlify preview deployment before trusting the generated URLs. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Supersedes the mirror decision in dc0389d9. Every generated group now points at raw.githubusercontent.com/PaloAltoNetworks/openapi on main, so there is no committed copy to keep in sync and no sync script to remember to run. Removes aigw/airs-openapi.yaml and aigw/scripts/sync-openapi.sh. This costs nothing today. The public document is structure-only -- the spec repo keeps reader-facing prose in overlays/docs-prose.yaml -- but that overlay's actions currently only blank inherited prose rather than supply new prose. The bare spec and a flattened one publish an identical 68 non-empty description and summary strings, so nothing is lost by reading the bare one. When the overlay is populated against the KB, the fix is for the spec repo to publish a resolved document; docs should not go back to mirroring. Adds aigw/scripts/check_nav_matches_remote.py for the coupling this introduces. Navigation is generated from the spec repo's local docs-navigation.json while Mintlify fetches from remote main, so a tag dropped upstream would leave an empty group and a tag added upstream would go unrendered. The check fails on either. Currently 40 tags, in sync. Picks up two upstream changes in passing: a self-hosted second base URL (84b8bb9) and the removal of Workspaces and SCIM workspace mappings (d155f8c), which is why Admin APIs now carries 23 tags rather than 24. mint validate green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs(aigw): Phase 3 — de-Portkey the content surface
Links under aigw/ that pointed at /api-reference/... were resolving against the Latest Portkey-branded version, so nothing flagged them as broken while a Prisma AIRS reader silently left the Prisma AIRS docs. Three classes, handled individually rather than by bulk prefix: - Targets that exist under aigw/ -- repointed (inference-api/headers, inference-api/config-object, inference-api/open-api-specification, integrations/libraries/microsoft-semantic-kernel). - Targets that will never exist -- /api-reference/sdk (no SDK reference is carried into aigw) and the Workspaces / Users / Log Exports Admin API groups (dropped from the spec). Cards removed, prose unlinked with the text kept. - Targets never carried into aigw -- gateway-for-other-apis, anthropic-transform, a Messages API reference. Unlinked, text kept. Also fixes open-api-specification.mdx, whose frontmatter still redirected to https://github.com/Portkey-AI/openapi instead of the PaloAltoNetworks repo. Roughly 90 links remain. Those all target generated operation pages and are blocked on x-mint.href being set upstream in the spec. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 5, first item. Two classes of contact ask, both removed on your call: Sales CTAs (13 files) — "Book a Demo" and "Schedule a Demo" cards, the "Interested? Schedule a Call Below" iframe on enterprise-offering, the "schedule a quick call"/"consultation" links on rate-limits, budget-limits and enforce-budget-and-rate-limit, and two "Reach out to our team" blocks. Support trailers (33 files) — the "If you face any issues with X, reach out to support@portkey.ai" paragraph and its variants. Where the trailer was the only content under a "## Get Support" or "## Support" heading, the now-empty heading goes with it; 29 such headings were removed. Consequence worth naming: those pages no longer state an escalation path. support@portkey.ai survives elsewhere (it is its own RESIDUE item), but the per-integration pointer is gone by design, not by oversight. aigw/scripts/strip_ctas.py does the repeating shapes and reports anything it does not recognise rather than guessing; the six it flagged were hand-edited. It is line-based on purpose — an earlier whole-file blank-line collapse reflowed 119 files that had no CTA in them at all. mint validate green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 5, second item. 22 callouts across 21 files stating that a feature is gated behind a commercial tier, usually paired with an ask to contact us. Three distinctions drove what stayed: Plan tier vs deployment shape. "Enterprise self-hosted plans only" is two claims wearing one coat. The tier goes; "available on self-hosted deployments" is a real constraint a reader needs and stays — budget-policies, metadata, otel/complete-logs, mcp-gateway/advanced-configuration. Plan tier vs version requirement. secret-references and budget-policies gated on a tier *and* a minimum gateway version. Only the tier was removed. Ours vs theirs. Cursor, Figma, Atlassian and Tavily document their own vendors' plan tiers. Those are facts about a third party, not our pricing, and are untouched. Also removed the "Plan restriction" cause from the permissions troubleshooting page: with tiers undocumented it is no longer a diagnosis a reader can act on. Descriptive "Enterprise Features" headings and cards are not gating callouts and were left alone. mint validate green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 5, fourth item. The ticker was a single top-level "banner" key in docs.json, not an aigw-scoped element, so there was no way to drop it from Prisma AIRS alone. Removed globally on your call. This is a deliberate exception to "nothing in the Latest version is modified" — the banner is now gone from Latest and Virtual Keys (Deprecated) as well, and both linked to portkey.sh campaign URLs that outlive the webinar anyway. mint validate green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 5, third item. 13 human-readable mentions across 10 files. The item reads larger than it is. aigw carries 1,384 occurrences of some "portkey api key" spelling, but 1,243 are the PORTKEY_API_KEY environment variable, 79 the x-portkey-api-key header and 62 the portkey_api_key config key — all functional identifiers on the never-rename list. Only 13 were prose a reader merely reads. Renamed: Helm values comments, three MCP auth diagrams, the OpenWebUI valve description string and shared-key label, the OpenAI-compatible setup step, and the LiveKit placeholder. Where a line held both, the identifier was left and only the sentence around it changed — "PORTKEY_API_KEY=<PORTKEY_API_KEY> # Replace <PORTKEY_API_KEY> with Gateway API Key". Guarded by a before/after count over six identifier patterns; all six are unchanged. Also caught in passing: those same Helm comments said the key comes "from Portkey website" while already pointing at the Strata Cloud Manager URL. Now they say Strata Cloud Manager. mint validate green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ames Phase 5 items 5 and 6. The Setup Guide's scope had crept into plugin enablement, guardrail creation and config creation, none of which belong in a first-request walkthrough and all of which were wrong in the details. Cut those three steps, renumber the rest, and point at the pages that own that content. Rename eight pages whose filenames had drifted from their titles or carried typos, including the requested simple-setup.mdx. Navigation and inbound links repointed. No redirects: these are new URLs with no inbound traffic (PLAN assumption 9). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six headings had a full sentence glued onto the title, so the sentence appeared in the sidebar and the anchor link. Split each into a heading plus the paragraph it was swallowing. No wording changed. Long headings that are genuinely one title — "Example: ...", "MCP Gateway: ...", linked headings — were left alone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
|
This was referenced Sep 15, 2026
Devansh0012
approved these changes
Sep 15, 2026
Contributor
Author
|
Superseded by #1080, which targets Closing rather than merging on purpose. This branch's Verified against both pushed heads: this branch has 40 tag keys and 0 operation entries; #1080 has 0 and 187, and is a strict superset of this branch's commits. 🤖 Generated with Claude Code |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
One PR, replacing the stack. Supersedes #1077 and #1078, both now closed.
Why this contains three phases and not one. #1075 merged to main at 06:00 with phases 1 and 2. Phase 3 (#1076) then merged into the umbrella branch at 11:03 — after the umbrella had already landed. So phases 3, 4 and 5 never reached main. This branch is the phase-5 tip merged with current main, carrying all three.
323 files, +5,384 / −27,806.
mint validategreen.Phase 3 — branded content, SDK removal, REST-first
Portkey SDK examples are gone everywhere. Every page offers REST, OpenAI Python and OpenAI Node, with REST leading in all 115 CodeGroup/Tabs blocks. Code comments and sample identifiers de-branded inside fences; functional identifiers untouched.
Chat, embedding, image and audio examples no longer send
x-portkey-provider— the provider slug rides inmodelas@openai/gpt-4o. The header stays where it is load-bearing: no-model endpoints (files, batches, fine-tuning), custom-host and ad-hoc credential examples, the gateway-nativeportkeybatching sentinel, and blocks with siblingx-portkey-provider-*headers.Removed: page icons, the welcome-page YouTube link, the GitHub star ask, edge worker locations, the ISO/SOC/GDPR/HIPAA compliance line, the no-storage and managed-hosting paragraph, the 25M-requests claim, pricing questions.
Phase 4 — API reference generation
Every
aigwnavigation group reads the spec straight from the PaloAltoNetworks/openapi raw URL. No mirror, no sync script. 40 tags across six generated groups;aigw/scripts/wire_openapi_nav.pyrebuilds navigation idempotently, andcheck_nav_matches_remote.pyfails on spec/nav drift.api-referencelinks that resolved to the Latest version are repointed where a target exists, and unlinked with their text kept where one never will.Phase 5 — callouts, naming, page hygiene
cc27c7c2f808ad615090818fdocs.json0b5b2b58cd88027dc153a29aWhat Phase 5 deliberately kept
aigw/, 1,371 are the env var, the HTTP header or the config key. A before/after count over six identifier patterns proves none moved.Example: …,MCP Gateway: …, linked headings.Three things to confirm
/introduction/make-your-first-request, but that page has exactly three headings — Get your Portkey API Key, Integrate Portkey, Next Steps — and covers none of the three cut steps. Pointing at it would have been a dead end, so the page points at/aigw/integrations/guardrails/palo-alto-panw-prismaand/aigw/product/ai-gateway/configsinstead.Blast radius outside
aigw/The webinar banner was the global
bannerkey indocs.json, not an.mdx— removing it also takes it off Latest and Virtual Keys. Agreed in advance as an explicit exception to the no-Latest-edits rule.No redirects were added for the eight Phase 5 renames, per PLAN assumption 9: these are new URLs with no inbound links.
🤖 Generated with Claude Code