Context engineering, answered.

Spiderbrain is a context-engineering tool that builds a living context graph of your project and serves it to AI agents over the Model Context Protocol (MCP). Instead of pasting files into a chat, your agent gets persistent memory and structural reasoning: it understands how every file, function and decision connects, and what breaks when one of them changes. It runs as a small desktop app and is free to start.

The product is written as one word, Spiderbrain. It is often searched as two words, Spider Brain, and both refer to the same tool at spiderbrain.ai.

It gives agents memory that persists across sessions, ranks the files that actually matter, and answers what a change will break before you ship it. The map updates itself as you work, and it plugs into Claude, Cursor and any other MCP client.

Yes. Spiderbrain includes a brain-aware agentic chat: you converse with an AI that retrieves your codebase context ranked by importance and shows a blast-radius warning before it makes any edit. It is multi-provider (Anthropic, OpenAI, Google, or local models via Ollama and LM Studio) and bring-your-own-key.

Context engineering is the practice of assembling the right information, memory and tools around an AI model so it can reason reliably. It is widely seen as the successor to prompt engineering: rather than wording a single message well, you engineer the whole context the model works from. Spiderbrain is a context-engineering based persistent cognition tool, built initially as a context graph that also reasons over what it stores.

A context graph is a structured map of how the parts of a project connect and depend on one another, instead of a flat pile of text. Gartner expects most AI-agent systems to use context graphs by 2028. Spiderbrain builds one automatically and goes further: it ranks what matters most and traces what fails when something changes.

RAG and vector databases retrieve chunks of text that look similar to your query. They have no sense of structure, so they cannot tell you what depends on what, or what breaks downstream. Spiderbrain keeps a structural graph and reasons over it, scoring every node by importance and tracing failure across dependencies. Most tools hand your agent a subway map. Spiderbrain hands it the same map with the broken tracks circled in red.

Not quite. A knowledge graph stores facts and relationships you define up front. Spiderbrain builds its graph automatically from your real files, imports and decisions, keeps it current as you work, and scores each node by how much its failure would hurt and how far that failure would spread. It is a context graph that ranks and reasons, not just a store of facts.

Context rot is the well-documented drop in accuracy as you push more text into a model’s context window. Flat dumps and stale embeddings make it worse, because the model spends attention on noise. Spiderbrain works on the input side: it ranks every part of your project by importance and blast radius, so your agent sees the load-bearing context first and the noise last.

Spiderbrain reads JavaScript, TypeScript, Python, Go, Java and C at the structural level, and handles other languages with a lighter parser. Beyond code it ingests Markdown, SQL schemas, PDF, docx, HTML, CSV and chat exports from ChatGPT, Claude and WhatsApp. It builds one graph across all of it, so your docs and decisions sit next to the code they describe.

Spiderbrain is MCP-native, so it connects to any client that speaks the Model Context Protocol. That includes Claude and Claude Code, Cursor, Continue, Cody, Zed, Windsurf, Devin, opencode and Aider. There is no glue code to write: connect the server once and your agent inherits memory and reasoning that stay current.

Spiderbrain connects to any tool that speaks the Model Context Protocol (MCP): Claude and Claude Code, Cursor, Continue, Cody, Zed, Windsurf, Devin, opencode and Aider. It is a hosted MCP server, so there is nothing to run locally. Step by step: (1) in the Spiderbrain desktop app, sign in and build a brain by pointing it at your project; (2) open Settings, then MCP, and copy your connection URL, https://mcp.spiderbrain.ai/mcp, and your access token; (3) in your AI tool, add Spiderbrain as an MCP server, for example {"mcpServers":{"spiderbrain":{"url":"https://mcp.spiderbrain.ai/mcp"}}}, and send your token as a Bearer token (zero-config OAuth is coming; today it is a token connection); (4) reload or restart the tool. Your agent then has Spiderbrain's ranked memory and structural reasoning, staying current as your brain updates. Separately, Spiderbrain's own in-app chat lets you bring your own model key (Anthropic, OpenAI, Google, or local models via Ollama and LM Studio).

Your source code and file contents never leave your device. Spiderbrain parses your project on your machine, then uploads a source-free graph to score it: repo-relative file paths and names, the dependency edges, a SHA-256 fingerprint of each file (never its contents), and privacy-safe git signals (commit timestamps, a fix or revert flag, a salted author-identity hash, and changed paths). Your code, file contents, commit messages and author emails never leave. The scored brain is stored in the EU and served to your AI over MCP. See the Data Governance Policy for the full breakdown.

Parsing your project is native and happens on your device, but an internet connection is required to sign in, score your brain, and connect your AI over hosted MCP. Scoring runs on our EU servers, which is what keeps the engine proprietary, so a scored brain needs a connection. You do not need Node, Python, Java or Docker; it is a small native desktop app.

Spiderbrain is deterministic: the same input produces the same brain every time, with no randomness and no hidden model in the scoring path. The engine ships with a 30-gate test suite of 353 checks, all passing, and has been validated on real codebases. Anything an AI suggests is clearly labelled and never applied on its own, so you always know what is fact and what is a suggestion.

Rebuilding is one click, with no stale embeddings to manage by hand, and drift detection flags files that have changed in ways that no longer match their history. On the Pioneer founding tier, live watch streams file changes into your graph as you work. Re-scoring stays an explicit step on every plan, so your agent always reasons from a picture you chose to publish.

Spiderbrain ships today as a small native desktop app for Windows 10 and 11 (x64) and macOS (Apple Silicon and Intel). Linux is planned. It connects your editor to a hosted MCP server. See the system requirements below for the full list.

Spiderbrain is a small, native desktop app, so the requirements are light and it needs no Node.js, Python, Java or Docker. Minimum: Windows 10 or 11 (64-bit) today (macOS and Linux builds are on the way); a dual-core 64-bit processor, 4 GB of RAM, and about 300 MB of free disk for the app and its brain cache. You also need an internet connection, because sign-in, scoring and the hosted MCP server run online, while parsing your code stays on your device. For large repositories of 10,000 files or more we recommend a quad-core or faster CPU (parsing scales with cores, roughly 4x on four cores), 8 GB or more of RAM, and an SSD; at that size a project parses locally in about half a minute and then scores in the cloud. Git is optional: add it to your PATH to clone remote repositories from inside the app and to enrich a brain with git-history signals, but local-folder projects work without it.

No. Spiderbrain is a single small desktop app with a native Rust parser built in, so there is nothing to install alongside it: no Node.js, no Python, no Java, and no Docker. You download one app, sign in, and point it at a folder. Parsing runs natively on your machine and scoring runs on our servers. The only external tool that helps, and only optionally, is Git, used to clone remote repositories in-app and to read git-history signals.

Every decision is recorded in an append-only log that is signed and hash-chained, so tampering is detectable and you can replay exactly how the brain reached its current state. Nothing is silently overwritten. For teams, that same log makes a shared brain auditable rather than a black box.

Yes. The Local plan is free, with no credit card required. You get one project, the full brain (parsed on your machine, scored in the cloud and served to your AI over MCP), with the on-screen map drawing up to 1,500 nodes. It is free for personal, research and open-source use, and it is the fastest way to see the difference on your own code.

Local is free and covers one project, with the on-screen map drawing up to 1,500 nodes. Pro is $36 a month, and Founding Members lock in $21 a month for as long as they keep the plan; it raises you to five projects, lifts the map to 3,300 nodes, keeps your brains hosted in the cloud and synced across devices, and adds dedicated support tickets through Webby. Founding tiers stack perks on top of Pro (Pioneers, for example, get live watch); see the tiers page for the full matrix. Cortex, at $99 per seat each month, gives a team one deterministic shared brain (map up to 9,000 nodes, plus 3,000 per seat) with assisted merging and on-call support. The full brain is always built, stored in the EU and served to your AI over MCP on every tier; the node count is only how much the local map draws. Enterprise adds licensing, a custom rollout from Perform Digital, priority support and a dedicated account manager.

Spiderbrain Pro is $36 a month. Founding Members lock in $21 a month for as long as they keep the plan. The offer runs until Spiderbrain Cortex reaches general availability, or while founding numbers last: 3,750 across three tiers (Pioneer, Founding Member and Genesis), each with its own perks on top of Pro. Your member number is assigned in signup order and decides your tier; the tiers page has the full matrix and how to avail the benefits in the app.

No. Download the free Local plan and build your first brain in minutes. When you are ready for more projects, cloud-hosted brains or a shared team brain, you can upgrade from inside the app.

Pro lifts you to five projects, raises the on-screen map to 3,300 nodes, keeps your brains hosted in the cloud and synced across devices, and gives you dedicated support tickets through Webby. Founding members activate extra perks in the app on top of Pro, by tier: every tier keeps cloud access forever and adds capacity, the first two tiers add priority support inside the app, and Pioneers get live watch, the graph updating as files change. The tiers page has the full matrix.

Yes, on the Cortex plan. Several people work against one deterministic shared brain, and Spiderbrain merges everyone’s view predictably, surfacing conflicts rather than guessing. Assisted AI merging is available as a labelled beta. For larger rollouts, Enterprise adds licensing and hands-on implementation.

Spiderbrain is built by Perform Digital, a team scaling SaaS since 2019 that builds specialized AI agents for B2B, alongside consulting, training and deep AI research into agent memory and reducing hallucination. The Spiderbrain algorithm is authored and engineered by Abhishek Srivastava, principal at Perform Digital. The engine is proprietary and is not on GitHub today; you can use it free on the Local plan, upgrade to Pro for more projects and cloud-synced hosted brains, and commercial team use is covered by the Cortex and Enterprise plans.

Webby, the assistant in the corner of this page, can answer product questions and, for customers on a paid plan, open a support ticket for you. Support tickets are available to paid members only. You can also email support@spiderbrain.ai any time. For Cortex and Enterprise, you get a direct line to the Perform Digital team.