Plasma AI released a deterministic wiki CLI for coding agents

Plasma AI released Plasma Wiki, an open-source command-line tool that gives AI coding agents a deterministic way to inspect project knowledge stored in Markdown files, avoiding the need to load entire documentation repositories into a model's context window. The tool, published on July 21st and demonstrated on July 31st, is part of a broader system for long-running software projects and is available under an Apache 2.0 license on GitHub and PyPI, requiring Python 3.11 or newer.

Nicholas Diao, Andrew Turner @ https://x.com/ acturner acturner and Colton Berry released Plasma Wiki https://www.plasma.ai/research/wiki , an open-source command-line tool that gives AI coding agents a structured way to inspect project knowledge without pulling an entire documentation repository into a model's context window. https://x.com/plasma ai/status/2083307284667449547?s=46 https://x.com/plasma ai/status/2083307284667449547?s=46 Plasma AI published the tool on July 21st and demonstrated the workflow in an X thread https://x.com/plasma ai/status/2083307284667449547?s=46 on July 31st. The timing matters because Plasma Wiki is part of a broader system the three authors are building for agents that work across long-running software projects, where documentation can outlive any single model session. Turner identifies himself on X as a Plasma AI co-founder. Diao previously co-founded Coursedog, the academic-operations software provider that joined Y Combinator's Winter 2019 batch. Y Combinator lists Coursedog https://www.ycombinator.com/companies/coursedog as acquired and identifies Diao as its co-founder and CTO, as well as founder and CEO of Omega Strategies. Berry arrived with experience building software close to enterprise operations. His public work history https://www.linkedin.com/in/coltonsberry includes Palantir and Saltbox, where he said he was the second engineer and sole dedicated backend developer on a product that was later spun out. That mix of agent infrastructure, enterprise deployment and zero-to-one software work helps explain Plasma AI's focus: turning autonomous agents into systems that can be inspected and controlled while they work. A file-navigation model for project knowledge Plasma Wiki stores knowledge in ordinary Markdown files. Each folder becomes a section, each file becomes an entry, and generated index.md files connect the tree. Developers can edit those files in Obsidian, VS Code or another text editor while agents access them through a predictable set of commands. The core loop has three steps. wiki map prints the documentation tree with word counts, giving an agent a compact view of what exists. wiki search runs text or regular-expression searches. wiki read retrieves a selected page or a specified portion of it. That design borrows the way coding agents already inspect repositories. An agent usually lists directories, searches source files and opens only the relevant code. Plasma AI is applying the same pattern to architecture notes, engineering decisions, conventions and operating procedures. The approach deliberately avoids semantic or vector retrieval. Search results depend on the actual text and a deterministic index, which makes the agent's path through the knowledge base easier to reproduce. Plasma AI says its CLI also maintains links and handles generated sections of index.md files when parallel edits collide. Authored content remains subject to normal merge conflicts, leaving those disputes visible for review. Plasma AI also describes the limits plainly. For a small collection of documents, ordinary files and grep are simpler. The indexed tree becomes useful when a project contains enough accumulated knowledge that an agent needs a map before deciding what to read. The GitHub repository https://github.com/plasma-ai/wiki carries an Apache 2.0 license and had 71 stars and eight forks as of July 31st. The Python package is also available through PyPI https://pypi.org/project/plasma-wiki/ and requires Python 3.11 or newer. Wiki supplies memory for Plasma's larger agent system Plasma Wiki is the memory layer beneath Plasma Fractal https://www.plasma.ai/research/fractal , the group's other open-source project. Fractal organizes agents into a hierarchy of persistent loops. Each node receives its own git worktree, branch, task, memory and spending limits, and can create child nodes for smaller pieces of a larger job. The Fractal repository describes support for Claude Code, Codex, Grok Build, OpenCode and Oh My Pi. Run metadata and costs are stored in a local SQLite database, while operators can inspect and steer the hierarchy through a terminal interface. The Fractal codebase https://github.com/plasma-ai/fractal had 649 stars and 46 forks as of July 31st. Wiki gives those nodes a shared record organized by topic rather than a transcript of prior conversations. That distinction becomes important when several agents are editing the same codebase. Plasma AI says it built Wiki after its agents repeatedly loaded the same material into context and overwrote one another's changes. The product strategy extends beyond two developer utilities. Plasma AI https://www.plasma.ai/ describes its goal as infrastructure for coordinating fleets of agents, including ownership, handoffs, shared knowledge and audit trails. Its site says a hosted version, Fractal Cloud, is planned. Plasma AI is also recruiting for five positions https://www.plasma.ai/careers in New York or Boulder across engineering, product, design and operations. Wiki is the smaller release, but it addresses a basic constraint in autonomous software work: agents cannot reliably follow decisions they cannot find. Plasma AI's bet is that predictable file and command-line primitives will give long-running agent systems a usable institutional memory without hiding that memory inside a proprietary retrieval layer.