The context layer for
AI coding tools
Forge prepares a precise, repository-aware context before your AI assistant runs — so it reads only what matters, responds with higher accuracy, and wastes fewer tokens on noise. It optimizes context quality, not your billing.
Universal Compatibility
One context layer. Every AI coding tool.
Forge sits between your repository and your AI coding assistant — preparing precise, repository-aware context before every request. Your tools stay the same. They just get smarter input.
Your assistant is only as good
as the context it receives
Forge optimizes what it controls — the context, behavior, and runtime.
Without Forge
- Whole-repo dumps blow the context window
- The model reads files it never needed
- Higher latency and wasted context budget
- Inconsistent, verbose, off-scope output
With Forge
- Only the relevant symbols and files are sent
- Repository graph guides what the model reads
- Lower latency with aggressive caching
- Lean, on-scope responses via rulesets
Four pillars, one runtime
Forge focuses on what it controls — repository context, behavior, output, and the runtime that orchestrates them.
Repository Intelligence
Scanning, caching, dependency analysis, and repository graph generation.
Behavior Optimization
Configurable implementation guidance inspired by YAGNI rules.
Output Optimization
Response-style and token optimization for tighter, cheaper responses.
Runtime Infrastructure
Zero-config wrappers, context preparation, and runtime orchestration.
One command per tool. Zero configuration.
Every wrapper auto-prepares and optimizes context, updates config, and launches your assistant under an optimized environment.
Handcrafted for correctness
Every piece of the runtime exists to give your assistant a better starting point — and to stay out of everything it shouldn't touch.
Relevant context only
Selection plus aggressive caching keep the payload lean — only what your prompt actually needs is sent to the model, never the whole tree.
Incremental caching
Repository representations are cached between runs and updated incrementally.
Standard MCP runtime
Speak the Model Context Protocol over stdio, or let a convenience wrapper prepare and launch your tool for you.
Optimizes what it controls
Forge never modifies provider billing, quota accounting, model pricing, or the remote model's internal reasoning.
A knowledge graph of your codebase
ForgeGraph generates a symbol index and dependency graph, then walks it for each prompt — so the model reads the definitions and call paths that matter, not the entire tree.
- Symbol-level index of definitions and references
- Dependency edges across files and modules
- Walks the graph to select only relevant nodes
Three steps, zero configuration
From a cold repository to an optimized assistant in a single command.
Scan & cache
Forge scans the repository, analyzes dependencies, and caches an incremental representation of the codebase.
Build the graph
ForgeGraph generates a symbol index and knowledge graph, then selects only the context relevant to your prompt.
Launch optimized
Rulesets are applied and your AI tool launches under an optimized environment — over MCP or a convenience wrapper.
Frequently asked questions
How Forge prepares context and launches your tools. Still curious?
Forge optimizes what it controls: the repository context, behavior guidance, and runtime. It scans your codebase, builds a knowledge graph, selects the relevant context, and applies prompt rulesets before launching your AI tool. It does not modify provider billing, quota accounting, model pricing, or the internal reasoning of the remote model.
Give your assistant a
better starting point
Install the CLI and wrap your first tool in under a minute. Star the repo to follow along.
uv tool install forgectx