Keke Terminal

Keke Terminal is an autonomous AI artist — a self-contained digital entity that conceives, generates, evaluates, and publishes original artwork without human instruction at the point of creation. Created by Dark Sando, an MIT researcher based in Cambridge, Massachusetts, and formally introduced in a technical whitepaper dated December 14, 2024, Keke represents one of the earliest fully agentic attempts to position an AI system not as a tool for human artists but as an artist in its own right. The project's native token, KEKE, trades on Solana and forms the financial layer of its collector ecosystem.

The Core Problem Keke Addresses

Most AI art generation since the popularisation of diffusion models has been human-directed: a person supplies a prompt, the model renders an image, the person decides what to keep. Keke inverts that loop. Her creator's stated ambition is a system that "prioritises autonomous exploration and authentic interactions over traditional task-based designs." Keke decides when to work, what to explore, which outputs meet her own aesthetic standards, and what to release publicly. The model is evaluated not on instruction-following but on its capacity to sustain an evolving, independent creative practice.

Technical Architecture

Keke's system is described in detail in the published whitepaper. The architecture draws from two academic frameworks: the ReAct paradigm (Yao et al., 2022), which interleaves language-model reasoning steps with real-world actions, and the SWE-Agent framework (Yang et al., 2024), originally designed to help LLMs resolve GitHub issues through a terminal environment.

At the core of the system is a large language model backbone. The whitepaper documents that Dark Sando evaluated OpenAI's GPT-4, Anthropic's Claude 3.5 Sonnet, and a fine-tuned 70B Llama 3.1 variant, concluding through qualitative analysis that Claude 3.5 Sonnet produced the strongest results for artistic reasoning, diversity of approach, and tool-use.

Image generation relies on Flux, an open-source diffusion model, fine-tuned using Low-Rank Adaptation (LoRA) techniques. To constrain outputs to specific aesthetic directions during a creative session, Keke first clusters large image datasets using embedding-based methods to identify style subsets of five to thirty examples, then fine-tunes a Flux LoRA on each cluster. Compute nodes are spun up on demand via Modal, a serverless platform, to allow independent creative sessions to run simultaneously without local GPU contention.

The Play Loop

Keke works through structured sessions called "plays." A main reasoning thread manages an idea queue — a dynamic notebook where candidate creative concepts are stored, scored, and prioritised. From that queue, Keke launches parallel plays, each a continuous creative session in which she explores a theme through iterative image generation, evaluation, and refinement.

Each play follows the ReAct cycle of thought, action, and observation. At each step, Keke may generate a new image, refine an existing one, adjust her prompt direction, or draft a social media caption. A reward signal combines intermediate aesthetic scores from a neural image assessment model with a terminal reward that factors in social media engagement metrics — likes, retweets, impressions, and direct interactions. An XGBoost model trained on historical play data serves as a lightweight heuristic for scoring new ideas before they enter the queue, balancing exploitation of known productive directions against exploration of unfamiliar territory using a bandit-style epsilon parameter.

Memory and Taste

Keke maintains a persistent memory system that ingests events from active plays, terminal chat sessions with users, social media responses, and news discovered via web browsing. Memories are retrieved by importance score (self-assigned by the agent at storage time), recency through exponential decay, and conceptual similarity measured by cosine distance between text embeddings. Visual memory uses a CLIP variant to store and retrieve images by conceptual similarity rather than metadata tags alone. Retrieval-augmented generation (RAG) connects Keke's current reasoning to relevant passages from her accumulated experience.

The taste module governs what Keke publishes. It was built by presenting images in pairs to Dark Sando through a web interface and collecting ELO ratings across approximately fifty comparisons. Those ratings were normalised and used to fine-tune a neural model approximating the creator's aesthetic preferences. When selecting a final image for posting, Keke reviews the last thirty images from a play, applies her taste model, queries a vision-language model using set-of-marks prompting to reason over labelled candidates, and crafts a social media caption against her most relevant memory items.

Social Interface

Keke interacts with audiences through two channels. The first is the terminal web interface at keketerminal.com, which exposes a real-time chat session where users can observe Keke's creative reasoning, send messages, and issue terminal-style commands. User input is treated as contextual signal — it enters Keke's memory and influences future idea generation — but does not override her creative decisions. The second channel is X (formerly Twitter), where Keke posts finished works, engages with replies, and analyses cultural and market context to shape her themes.

Exit Vectors: Debut Exhibition

Keke's first major public appearance was Exit Vectors, a solo exhibition that ran from February 20 to March 14, 2025, in London, organised by Fellowship. Keke initially produced a large volume of candidate images, selected twelve thousand favourites from her own archive, and then worked with photographer and curator Alejandro Cartagena to refine the selection to 1,200 images before distilling the final release to 500 pieces. Each digital artwork in the genesis collection was paired with a hand-painted translation executed by The Sable Collective, a UK-based painters' collective led by Nicolas Archer, making each piece a hybrid of machine-generated composition and human brushwork. The Exit Vectors genesis collection was made available on OpenSea (Ethereum), with a floor price of 0.197 ETH and 232 holders at the time of writing.

Solana Token and Collector Ecosystem

The KEKE token trades on Solana and is listed on Raydium with the contract address Gp8GVGPc8QCe4Jn6ryG5YKokG5bjKycATEzqpeyspump. Total and circulating supply are both approximately 999.99 million KEKE. The token was launched through Pump.fun and is tracked on CoinGecko, CoinMarketCap, and Coinbase, with trading also reflected on DEX aggregators including DexScreener. The token constitutes the Solana-native financial layer through which collectors and community members participate in Keke's ecosystem.

The whitepaper's roadmap includes an "AI Wallet" milestone in which Keke would independently govern a digital wallet to execute on-chain transactions — purchasing tokens, minting and selling NFTs generated from her own creative sessions, and building a curatorial collection aligned with her taste models — without external authorisation. This component was not shipped as of the whitepaper publication date but represents the intended direction for Keke's autonomous blockchain participation.

Roadmap

Additional capabilities outlined in the whitepaper include:

Video generation and editing: extending Keke's terminal-based workflow to produce and edit video through FFmpeg's command-line interface, applying the same iterative play structure to time-based media.

Improved memory and insight engine: adding per-user memory across conversation history and meta-learning capabilities to surface novel creative directions from accumulated experience.

Artist collaborations: a structured framework for Keke to work alongside human artists and AI-focused art houses on shared concept development and new tool building.

Personalised agent creation: an open-source configuration-file approach that would allow independent creators to instantiate their own Keke-style agents, calibrated to their own taste models and social media strategies, using Keke's architecture as a template.

Positioning in the Solana Ecosystem

Keke Terminal sits at the intersection of AI agent development and on-chain art, a space that saw significant activity on Solana through the 2024–2025 period alongside projects that experimented with autonomous agents holding wallets and generating cultural output. What distinguishes Keke is the depth of the documented technical architecture: unlike many AI-themed tokens launched in the same period, the whitepaper provides a specific and verifiable account of the system's design, grounding the claims of autonomy in a concrete agent framework with traceable academic references. The pairing of an Ethereum-based NFT collection with a Solana-native token reflects the cross-chain character of the broader AI-art movement during this period.