New: vfx (1080p-default sibling of 4k-vfx).
Updated: anime-soccer, app-sizzle, app-store-screens, baseball-trend,
build-a-brand, explainer, fix-my-look, founder-product-video, gameday,
kiss-cam, language-swap, podcast, stagefight, ugc-ads, voxel-it — these
had drifted from internal main across several unsynced releases
(generate_image_edit migration, identity-subsystem removal, and other
skill-specific fixes).
Generated symbol/mark PNGs often return with a baked-in white background.
Run them through remove_background (mode=logo) before compositing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Update app-sizzle, app-store-screens, build-a-brand, explainer,
founder-product-video, kiss-cam, ugc-ads to match latest pika-claude-plugin.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Clicking https://mcp.pika.me/api/mcp returns 'This MCP server does not
offer a standalone SSE stream. Tool-call notifications are delivered on
the per-request POST response stream.' — the endpoint is correctly
wired for MCP-over-HTTP-POST in the install snippets above, but isn't
useful as a clickable Links destination. Drop the bullet to stop sending
people to a confusing error page.
The install-surface section, badge, and per-client snippets remain
unchanged — those document the actual integration shape.
Public skills/ directory grew to 9 in commit ee60838 "Duplicate Pika
skills" (app-sizzle, app-store-screens, build-a-brand,
founder-product-video added alongside the original podcast, explainer,
ugc-ads, baseball-trend, kiss-cam), but the README, marketplace.json,
and all three plugin manifests still advertised 5 curated skills and
version 1.3.0.
Updates:
- README — version badge 1.3.0 → 1.4.0; install-surface table and §2/§3
headers + body 5 → 9; "Featured workflows" table extended from 5 to 9
rows with input/output/invoke for each new skill
- CHANGELOG — new [1.4.0] — 2026-05-22 section above [1.3.0] documenting
each new skill, the skill-count + version bumps, and a note that the
58-tool MCP surface is unchanged
- .claude-plugin/marketplace.json — version 1.4.0, descriptions, and
skills[] extended from 5 → 9 entries (stable order, originals first)
- .claude-plugin/plugin.json — version, description, keywords
- .codex-plugin/plugin.json — version, description, longDescription
(each new skill gets a one-liner inside the Codex `interface`), keywords
- .cursor-plugin/plugin.json — version, description, keywords, tags
Tool count stays at 58 atomic primitives (tools-manifest.json
unchanged); skill-audit intentionally not ported from internal (stays
internal-only).
* feat: v1.3.0 — Pika Creative Suite repositioning + cross-agent distribution
Brings Pika-Plugins forward to match the cross-agent reality. Same backend
(mcp.pika.me, 58 atomic tools), same 5 curated /pika:* skills, now positioned
and packaged for use beyond Claude Code.
Repositioning
- "Pikafy your Claude" -> "Pika Creative Suite — give any AI agent a face,
a voice, and a full creative studio." README hero, all plugin manifest
descriptions, and the marketplace metadata description aligned to this.
Structure / new files
- .cursor-plugin/plugin.json — native Cursor plugin manifest mirroring
the first-party shape used in cursor/plugins (category, tags, single-
string skills: "./skills/"). Lets Cursor pick up Pika via its own
plugin format.
- .codex-plugin/plugin.json — native Codex plugin manifest with
documented interface block (displayName, defaultPrompt capped at 3,
capabilities, brandColor, etc.).
- .claude-plugin/marketplace.json gains an explicit skills:[] array in
the canonical Anthropic string-array form ["./skills/<name>", ...]
matching anthropics/skills.
Manifest cleanups
- .claude-plugin/plugin.json description rewritten, version 1.2.3 ->
1.3.0.
- .claude-plugin/marketplace.json description rewritten, owner.url
added, plugin version bumped.
- All four manifests now agree on the Pika Creative Suite positioning,
the 58 canonical tool count, and the 5 skill set.
Tool surface
- tools-manifest.json re-vendored from pika-mcp-server with the current
58 canonical primitives plus the 9 deprecation aliases (sora_edit
mode discriminator, pika_* renames, edit_audio_denoise, etc.). 16
net new canonical tools since v1.2.3 — most prominently html_to_pdf,
the connect_* Composio trio, scrape_ads / scrape_social, web_publish,
analyze_clip_highlights.
README
- Full v3 rewrite: hero, 3 independent install surfaces (MCP /
Skills / Plugin) as parallel orthogonal mechanisms, install matrix
with per-client config shapes (Claude type:http vs Cursor bare url
vs Codex --url), Featured-workflows table linking to each SKILL.md,
Capabilities table for the 11 tool families summing to 58.
- 479 lines (Pikafy hero, prose-heavy) -> 220 lines (Pika Creative
Suite hero, table-driven). All claims fact-checked against
vercel-labs/skills source, cursor/plugins first-party manifests,
anthropics/skills canonical marketplace, and openai/codex plugin
docs.
* fix: add repository field to marketplace.json plugin entry
Codex review flagged this — the plugin entry inside marketplace.json
plugins[] lacked the canonical repository field. Aligns with the
top-level plugin.json which already has it.
Two other codex flags triaged as false positives:
- "parrot-staging/animate" in tools-manifest.json is the upstream
pika-mcp-server animation backend endpoint name (vendored verbatim
from pika-mcp-server@staging), not the staging plugin variant.
- "Pikafy your Claude" / "Pika-fy your Claude" in CHANGELOG.md v1.3.0
entry are intentional quoted references documenting what the entry
is RETIRING. Documenting history, not active usage.
New curated skill — viral fake "in-arena Kiss Cam moment" of any two
subjects (any visual style: photoreal humans, 3D toys, illustrated avatars).
Spectator-POV phone shot of the Madison Square Garden Jumbotron with retro
red kiss cam graphic + Knicks vs Bulls scoreboard baked into frame 0
(gpt-image-2), then a 15s kling-v3-omni clip first-frame-locked to the still
so the entire decorative UI stays pixel-static across all 15s — only the
two subjects inside the kiss cam panel animate. Native off-screen
PA-announcer commentary + packed-arena crowd reaction. No names, no chyron.
Engine choice mirrors baseball-trend: Kling-only on video (Seedance's
two-stage face-moderation gate rejects every in-arena reaction shot
because of the crowd faces), gpt-image-2 on still (sharpest LED detail +
strongest reference-likeness lock).
Skill doc follows the public baseball-trend SKILL.md structure
(Prerequisites / Stage 0 / Step 1-3 / Load-bearing phrases / Engine choice
/ Failure cheat sheet / What not to do), with one additive section
(Runtime expectations).
Plugin manifest + marketplace + README + CHANGELOG bumped 1.2.2 → 1.2.3
and "4 curated" → "5 curated".
Co-authored-by: pika-ai-code <272925370+pika-ai-code@users.noreply.github.com>
1.3.0 was bumped for the cb0eb2b camera-position-anchor fix, but that
fix didn't actually work — gpt-image-2 kept rendering the camera in
the wrong position. The real fix landed in ad51ef1 (revert Step 1
to the pre-refactor verbatim prompt), which is a recipe patch, not
a minor feature change.
Renumbering: 1.2.0 → 1.2.2 (skipping 1.2.1 to absorb the cb0eb2b
intermediate version that briefly shipped as 1.3.0). The 1.2.x line
is now the canonical baseball-trend skill release; future fixes stay
in 1.2.x as patches.
- .claude-plugin/plugin.json: 1.3.0 → 1.2.2
- .claude-plugin/marketplace.json: 1.3.0 → 1.2.2
- README.md: version badge + verify-output snippet
- CHANGELOG.md: [1.3.0] section replaced with [1.2.2] capturing the
full fix story (failed scaffold-anchor attempts → verbatim revert)
The cb0eb2b camera-position anchor didn't take — gpt-image-2's
training prior for "fan at MLB game" keeps composing shots with the
field visible behind the subject (camera in the upper deck looking
toward the field, instead of in the foul ground looking back at the
crowd). Two iterations of explicit camera-position instructions
in the bullet-scaffold version still produced inverted shots.
Empirically, the OLD pre-refactor verbatim prompt block produces the
correct composition reliably — the calibrated narrative form weights
constraints in a way that gpt-image-2 follows, where a bullet-scaffold
checklist with the same content does not.
Fix: replace Step 1's "prompt must contain" bullet scaffold with the
verbatim prompt block from the original baseball-trend recipe. Send
the paragraph exactly as written; only substitute ${username} in the
chyron line.
Step 2, call params, nano-banana-pro fallback, engine choice, and
failure cheat sheet are unchanged. Load-bearing phrases section
collapsed five Step-1 bullet anchors into a single "Step 1 prompt is
verbatim" guidance, since the whole paragraph is now the load-bearing
unit. Cheat sheet row updated to point at the new failure mode
(paraphrased prompt instead of verbatim).
Picks up the cb0eb2b camera-position fix as a minor release.
- .claude-plugin/plugin.json: 1.2.0 → 1.3.0
- .claude-plugin/marketplace.json: 1.2.0 → 1.3.0
- README.md: version badge + verify-output snippet
- CHANGELOG.md: new [1.3.0] entry documenting the camera-position lock
Empirical failure: a 2026-05-14 run produced a still with the baseball
field, grass, and infield dirt visible BEHIND the subject — which
means gpt-image-2 placed the camera in the upper deck looking TOWARD
the field, with the subject having turned around to face it. The
correct trend composition is the opposite: camera in the foul-ground
photographers' pit just past home plate, looking BACK at the premium
seats, with the field BEHIND the camera and not visible at all in the
frame (because the crowd is facing the field, which is where the
camera is).
The prior scaffold had "The camera cuts to the audience" but never
specified the camera's physical position. gpt-image-2 has two
plausible interpretations of "behind home plate seats + broadcast
shot" — one with the field visible (wrong), one without (correct) —
and was picking the wrong one ~50% of the time.
Fix: add a verbatim camera-position lock to the Step 1 scaffold:
"Camera is positioned in the foul-ground photographers' pit just past
home plate, low angle, shooting BACK at the premium seats — the
baseball field is BEHIND the camera and NOT visible anywhere in the
frame." Registered in the Load-bearing phrases section so future
edits can't strip it. Failure cheat sheet now also has a dedicated
row for the "field visible behind subject" symptom pointing at this
anchor.
The Pika Claude Code plugin ships with three curated /pika:* skills
and 42 atomic MCP tools at https://mcp.pika.me/api/mcp:
- /pika:podcast — 1-minute two-host podcast video for any URL or topic
- /pika:explainer — ~60–80s explainer video for any URL
- /pika:ugc-ads — 15s creator-style multi-cut UGC ad in 9:16 vertical
Installs on every Claude Code surface — CLI and Desktop (macOS /
Windows app). The 42 atomic MCP tools also work with any other MCP
client (Claude.ai chat connectors, Cursor, Codex, etc.) at the raw
endpoint https://mcp.pika.me/api/mcp.
License: Apache 2.0.
pika-skill-sync