Pika MCP: Generate Videos in Claude With Pika Agents (2026)

One-sentence definition

Pika MCP is a remote Model Context Protocol server at mcp.pika.me/api/mcp that gives any MCP client OAuth-authenticated access to your Pika Agent — a persistent persona with three identity files (identity, soul, style) that orchestrates fourteen-plus generation models including Pika Video, Sora, Veo 3, Kling, MiniMax, ElevenLabs, and Whisper, billed from a single Pika Wallet.

That sentence is the whole launch. Everything below unpacks one of its words.

Why Pika shipped this

Before Pika MCP, an AI-video workflow looked like a browser-tab tax. Sora for cinematic shots in one tab. Runway Gen-4 for character consistency in another. Pika for fast iteration. ElevenLabs for the voiceover. Suno or MiniMax for music. Whisper for the transcript. Five subscriptions, five API keys, five billing dashboards, five places to remember the same project context.

Pika spent 2025 collapsing that surface into one product. In April 2026 they announced Pika Agents: persistent personas with their own face, voice, and memory, sitting on top of a federated stack of generation models that the agent picks from at request time. The MCP server is the door that lets any AI client walk into that product without a separate Pika UI.

The pitch is exactly that compression. One Pika OAuth. One Pika Wallet. One agent persona that remembers your project across sessions. And every major creative model on the menu, behind a tool list the model itself decides how to use.

Mental model: the four named pieces

Once you know these four pieces, every other tab of the Pika docs is a deeper read of one of them.

┌──────────────┐                  ┌─────────────────────────────┐
│  Claude      │                  │   mcp.pika.me/api/mcp       │
│  Desktop     │ ── streamable ──▶│                             │
│  ChatGPT     │      http        │   ┌─────────────────────┐   │
│  Cursor      │ ◀── tool result ─│   │   Pika Agent        │   │
│              │                  │   │   ─────────────     │   │
└──────────────┘                  │   │   identity   (you)  │   │
                                  │   │   soul    (taste)   │   │
                                  │   │   style  (visual)   │   │
                                  │   └─────────────────────┘   │
                                  │            │                │
                                  │            ▼                │
                                  │   ┌─────────────────────┐   │
                                  │   │ generation models   │   │
                                  │   │ Pika · Sora · Veo 3 │   │
                                  │   │ Kling · MiniMax     │   │
                                  │   │ ElevenLabs · Whisper│   │
                                  │   │ Seedance 2 · …      │   │
                                  │   └─────────────────────┘   │
                                  │            │                │
                                  │            ▼                │
                                  │   Pika Wallet (tokens)      │
                                  └─────────────────────────────┘

• The MCP server — a remote HTTPS endpoint Pika hosts at mcp.pika.me/api/mcp. You don’t deploy it. You add it as a custom connector in your client and authorize.
• The Pika Agent — a persistent persona you create on pika.me or in the iOS app. Three identity files steer it:identity (who the agent is for), soul (its taste and voice), and style (its visual signature). Once Claude reads them, it stops being generic and starts replying in your agent’s voice.
• The generation orchestra — the fourteen-plus models the agent picks from. Inventory in a moment. The point: you don’t pick the model; the agent does, based on what you asked for.
• The Pika Wallet — the shared-token billing layer. Every generation debits the wallet at the model’s own rate. Token packs run $7.99 to $149.99. The connector itself is free; you pay per generation.

One more concept matters when you set up: the identity_persona_read prompt. The first thing you tell Claude after connecting is to read your three identity files. Skip this and the agent acts generic — same outputs you’d get from a raw prompt against pika.art. Run it once per conversation and the outputs change material weight.

Smallest end-to-end example

Five steps to a working setup, then one prompt to prove it works.

Setup in Claude Desktop

1. Open Claude → Customize → Connectors.
2. Click + next to Connector → Add custom Connector.
3. Name it Pika and paste https://mcp.pika.me/api/mcp as the URL. Click Sync, then Connect.
4. Authorize Claude in the Pika OAuth screen. Pick the tools you want exposed (default is all).
5. In a new chat, ask: “Read my identity_persona_read for the identity, soul, and style files.” The agent loads its persona for this session.

Your first real prompt

Once the persona is loaded, the next prompt looks like normal creative direction:

Make me a 6-second product demo for a stainless-steel water bottle.
Gen-Z UGC vibe. Voiceover that emphasizes the temperature-lock feature.
Square aspect ratio. Background music that's punchy but not distracting.

Then suggest a 30-second cut for Instagram Reels and a 15-second cut for TikTok.

What actually happens behind the chat:

• Pika Agent reads style and chooses the right visual model — likely Pika Video for the 6-second iteration speed, possibly Veo 3 if your style file asks for cinematic.
• ElevenLabs (or MiniMax Voice) generates the voiceover in the agent’s saved voice profile.
• MiniMax Music or a similar model lays down the background bed.
• The agent assembles the clip and returns a Pika URL plus the two re-cuts.

Cost: a few hundred tokens out of your Wallet, depending on which video model the agent picked. The orchestration is the point — you didn’t pick the model, you described the outcome.

The 14+ generation models behind one agent

The Pika MCP exposes a single tool surface; underneath, the agent routes to the right model. Here is the lineup Pika lists on its MCP page.

Opinion: the multi-model federation is the under-priced feature. Most analysts compare Pika 2.x to Sora 2 and Runway Gen-4 head-to-head — Pika typically wins on speed and cost, Sora on photorealism, Runway on editor-precision. Pika MCP makes that whole comparison the wrong question. You’re not picking one model; you’re asking an agent to pick whichever is right for the shot, including Sora itself.

Pika Skills and the slash plugins

On top of the raw MCP tool surface, Pika ships pre-built workflows as Pika Skills. Two flavors:

Managed plugins (slash commands)

Three workflow plugins ship in the Pika marketplace and attach as slash commands inside Claude:

• /pika:podcast — build a video podcast with named speakers, B-roll, and chapter markers from a brief.
• /pika:explainer — turn a URL, GitHub repo, or short brief into a narrated explainer with on-screen text and smart cuts.
• /pika:ugc-ads — UGC-style ad variants with creator personas, Gen-Z cadence, and platform-specific cuts.

Open-source skills

The Pika-Labs/Pika-Skills repo is Apache 2.0 and currently exposes one production skill: pikastream-video-meeting — let your Pika Agent join a Google Meet as a real-time animated avatar with your face and voice. Runs on PikaStream 1.0 at 24 FPS / 480p / ~1.5s end-to-end latency on a single H100, billed at roughly $0.20–$0.275 per active minute. The repo’s structure is the standard Claude-Code SKILL.md format, so any agent that reads SKILL.md (Claude Code, OpenClaw, Hermes) can adopt it.

Opinion: install the open-source skill if you do Google Meet calls; install the managed plugins only for the use cases you actually run. The slash commands are worth their weight when you do podcasts or UGC ads weekly. They’re overhead otherwise.

Pika MCP vs calling each model directly

The honest comparison: when does it make sense to skip Pika and call Sora/Runway/ElevenLabs APIs directly?

Opinion: Pika MCP for any creator workflow that runs inside Claude. Direct APIs for production pipelines that batch hundreds of clips per day. The line is when you start writing your own orchestration code. If your agent is doing the orchestrating in chat, you want Pika. If you’re writing Python that calls Sora.create_video() in a queue, the wallet markup isn’t worth it.

What I got wrong

Three assumptions I made testing Pika MCP that turned out to be wrong.

1. I thought Pika MCP was just for video. It is not. The connector exposes voice (ElevenLabs, MiniMax), music, transcription (Whisper), web search, and content analysis alongside video. The video stack is the loudest piece, but the under-marketed value is the agent stitching audio and transcripts into the same flow without leaving Claude.

2. I thought the agent persona was optional. I tried using the connector without running the identity_persona_read prompt and the outputs were generic — same caliber as raw pika.art. Once I ran the persona load, the same prompts produced visibly more on-brand cuts. The persona files aren’t flavor text; they’re instructions the agent threads through every model call. Always load them first.

3. I assumed I’d burn tokens fast on Pika Video. In practice, most agent calls in a chat session touch cheaper models first — image gen, voice, transcript — with full-frame video as the final assembly step. A 30-minute creative session against the 800-token pack got me a usable 6-second hero clip plus three variants and a voiceover. Your mileage will vary by how often you reach for Sora, but I overestimated the burn rate by roughly 3×.

Common mistakes

Who this is for, who it isn’t

Community signal

The Pika Labs team announced the MCP launch on X. The tweet is the cleanest single source for tone:

The persistent-persona thread had been building for weeks. Pika previewed the persona model with AI Selves — agents you “birth, raise, and set loose” — and the audio-driven Pikaformance model that produced hyper-real expressions in roughly six seconds. The PikaStream 1.0 announcement on April 2, 2026 logged 2M+ views on its launch video, per Superhuman’s coverage. The MCP closes the loop — the persona that lives in your phone, your Slack, and your video calls now also lives in Claude.

The contrarian read deserves equal airtime. Efficienist summarized the skepticism in one line: “people already talk to their agents through speech-to-text tools. So you can’t really say this tool is solving that problem. What this adds is a face.” That framing applies to PikaStream more than the MCP itself, but the underlying critique — that personalization isn’t the same as productivity — is fair. Treat the persona as a brand asset that pays back when you ship, not as the value prop on its own.

The verdict

The bigger picture

Three trends converge at this launch.

Persistent agent personas are becoming a primitive. Pika Agents, OpenAI’s memory-enabled GPTs, and Anthropic’s recent persona work are all converging on the same shape: a named, addressable AI character that holds memory across sessions and surfaces. Pika is first to do it for creative content. The MCP server is what turns that persona from a Pika-app feature into a cross-client primitive.

Token wallets are the new platform layer. Every generation model has its own pricing dimension — per-second video, per-character voice, per-image still. Pika’s Wallet abstracts those into one unit. Stripe-for-AI, but at the model layer. Watch for Cloudflare and Hugging Face to push similar wallets in the next year.

Real-time video chat (PikaStream) is where this is heading. Today, your agent answers in text inside Claude. Tomorrow, it joins your Google Meet with your face. The MCP is the table-stakes layer; PikaStream is the presence layer. Whether you find the second compelling or creepy, both run on the same Wallet and the same persona files.

Frequently asked questions

Glossary

All sources

Primary — Pika

• Pika MCP product page — endpoint URL, setup, model list, token packs.
• Pika Agents homepage — the persona product the MCP exposes.
• Introducing Real-Time Video Chat for Agents — PikaStream 1.0 launch (April 2, 2026).
• Pika-Labs/Pika-Skills — the open-source skill repository.

Community and analysis

• Pika on X — MCP launch tweet.
• Pika on X — AI Selves announcement — the persona predecessor.
• Pika on X — Pikaformance audio-driven model.
• Eyerys: Pika Labs launches Agents — April 2026 launch coverage.
• Superhuman: PikaStream video chat with your AI — reach metrics for the launch.
• Sora vs Runway vs Pika 2026 comparison — head-to-head on speed, cost, quality.

Contrarian and risk

• Efficienist on PikaStream — “what this adds is a face” critique.
• Lovart: Pika AI review — honest verdict.

Internal — on MCP.Directory

• What is the Model Context Protocol? A Developer’s Map
• Higgsfield MCP — agentic image and video generation
• Meta Ads MCP and CLI — another official-MCP deep-dive
• Claude Desktop client guide
• Claude Code client guide
• Cursor client guide
• Best MCP servers (2026)
• Remote (hosted) MCP servers
• Browse 3,000+ MCP servers