Stop compaction drift

Repo memory without pushing code to a cloud vector database.

CodeFathom is a local-first semantic indexing service that gives your CLI code agents (Claude Code, Codex CLI) persistent codebase memory via MCP tools. Keep conversations focused on intent and decisions, not pasted files.

Local-first privacyMCP-nativeCode-aware retrieval

Start free Pricing

The problem

Compaction drift kills long sessions

CLI code agents have finite context windows. As you paste files and discuss architecture, the window fills. When compaction kicks in, your original intent, discovered constraints, and key decisions get summarized away. That's compaction drift.

Context window fills

You paste files, discuss architecture, explore options. The agent's context fills with raw code and conversation history.

Compaction kicks in

The agent auto-summarizes older context to stay under its limit. Details get compressed into high-level summaries.

Drift happens

Your original intent, discovered constraints, and key decisions get lost. The agent starts making suggestions that contradict earlier conclusions.

The solution

Local semantic index with MCP retrieval

CodeFathom builds a semantic index of your repo locally. When your agent needs code context, it retrieves minimal, relevant slices via MCP tools instead of you pasting entire files.

Index locally

CodeFathom vectorizes your codebase and builds a symbol graph. Everything stays on your machine.

Query via MCP

Your agent calls MCP tools to search semantically, find references, or get architecture summaries.

Stay focused

Conversations stay about intent and decisions. No more pasting files or losing context to compaction.

MCP tools

Agent-facing retrieval surface

CodeFathom exposes five MCP tools that your CLI agent can call to retrieve exactly what it needs from your codebase.

search_semantic

Semantic search

Find code by meaning, not just keywords. Ask for "authentication logic" and get relevant functions.

"Find where we validate JWT tokens"

search_exact

Exact search

Precise text and symbol search when you know exactly what you're looking for.

"Find all uses of UserService"

find_references

Find references

Trace all call sites and usages of a function, class, or variable across the codebase.

"Where is handlePayment called?"

get_related

Get related

Find code entities related to a given symbol via imports, calls, or type relationships.

"What depends on OrderModel?"

describe_arch

Describe architecture

Get a structured summary of a module, directory, or the entire codebase architecture.

CodeFathom

No cloud storage fees. Pay only for context bandwidth delivered.

Cloud Vector DB

Storage costs, compute costs, egress fees add up.

Built for

Engineers who think in long sessions

CodeFathom is built for developers whose work requires sustained context across hours, not minutes.

CLI agent power users

Developers who use Claude Code, Codex CLI, or similar tools for serious coding sessions.

"My agent forgets what we decided two hours ago."

Ideation-heavy engineers

Developers who use agents for architecture exploration, refactoring planning, and design discussions.

"By the time I'm ready to implement, the context is gone."

Security-conscious teams

Teams that can't or won't push proprietary code to third-party cloud services.

"We need code search but can't use cloud vector DBs."

How it works

From repo to retrieval in minutes

Four steps to give your agent persistent codebase memory.

Install CodeFathom

Download and run the local service. No cloud account required for indexing.

Index your repo

Point at your codebase. CodeFathom builds embeddings and a symbol graph locally.

Connect your agent

Add CodeFathom's MCP server to your agent's config. One line in your settings.

Query and build

Your agent now retrieves code context via MCP tools. Stay in flow without pasting files.

Ready to try

Keep the plan alive.

Start your 14-day trial. Full access to all MCP tools. No code leaves your machine during indexing.

Start free Talk to sales