ctxpipe

Context management for software factories. Open source, MIT licensed.

Your agents can reach everything, but they cannot tell what is true. ctxpipe turns your code, tickets, docs, and incidents into one organized, current git repository of Markdown that you can inspect. Background jobs keep it current, and every coding agent reads it over MCP. Self-host the open-source core, or use the hosted version (early access).

The problem

Every repository, ticket, doc, incident, and Slack thread is one MCP call away. Thus agents read too much, trust the wrong parts, and you cannot see why. More context has not made the fleet better. It has made the fleet noisier.

• Too much. The context window fills with reading, not work.
• Too noisy. Stale runbooks, duplicate docs, and decisions that contradict each other all look equally true to an agent.
• Unorganized. Knowledge is spread across tools and many repositories. Nothing links a decision to the code that it governs.
• Opaque. When an agent makes a mistake, you cannot see what it read or why it believed it.

How it works

Write it down once. Keep it current. Serve it to every agent.

1. Connect: agents read your decisions, not only your code. Link your code repositories, and select what to pull from Linear, Notion, Confluence, Slack, and PagerDuty. ctxpipe writes it to one git repository as Markdown, with links back to the code. Each team or product gets its own Workspace.
2. Keep current: agents stop acting on stale docs. ctxpipe reads a linked repository again on every push. Each fact links back to its source. A fact expires when the code no longer backs it. Each job makes one commit that you can revert.
3. Serve: every agent starts from the same facts. Claude Code, Cursor, Codex, OpenCode, and VS Code search the same Workspace over MCP: knowledge, code, and graph. ctxpipe returns sources, not answers. Your agent reasons with its own model and cites the sources.

For the full flow, see Getting started.

Where ctxpipe fits

Same models as everyone else. Your factory is the difference.

Teams that scale with AI build a harness around the models: agents, review bots, and internal tools. Under all of it is what your organization knows about how you build. Most teams make that layer by hand, and it goes stale when its maintainer moves on. ctxpipe is that layer: open source, and kept in your own repository.

Designed for transparency

Nothing that your agents read is hidden from you.

• It is open source. The core is MIT licensed, including every job that writes to your Workspace. Read it, run it, or fork it.
• Context is files in a repository. Every fact that your agents read is a Markdown file in a git repository that you own. Open it in any editor, or point any tool at it.
• History is git. Every change is a commit that you can read, blame, or revert. Jobs commit small and often. Workspace chat edits arrive as pull requests.
• The knowledge stays with you. If you stop using ctxpipe, the Workspace repository stays where it is, with its full history.

Features

Feature	What it does
Git-backed Workspaces	One Markdown repository per team or product, in your git.
Six connectors	GitHub, Linear, Notion, Confluence, PagerDuty, and Slack.
Background upkeep	Reads code on every push, and expires facts that lose their source.
Knowledge graph	Typed links between facts, code, and their sources.
Inspectable views	Files, Diff, Graph, and the history of every commit.
Workspace chat	Answers with sources, and proposes changes as pull requests.
One MCP endpoint	Seven read-only tools that return sources, not answers.
Scoped access	Read-only, and limited to the organizations that you approve.
Isolated sandboxes	Each agent holds only the repositories and keys that it needs.
Self-hosting	Docker Compose, Terraform, or AWS CDK.

Get started

Hosted (early access)

We operate ctxpipe for you. We help you set up your first Workspace, connect GitHub and your tools, and connect your agents to it. Model calls go only to zero-data-retention endpoints, and ctxpipe does not train on your code.

Request early access, or book a demo.

Self-host

The MIT core runs on Docker Compose for evaluation, and on Terraform or AWS CDK in production. It works with any OpenAI-compatible model provider, including Azure and Bedrock. When you self-host, the servers of ctxpipe are not in the data path. Use the self-hosting docs as the deployment runbook, not this README:

• Self-hosting overview
• Quickstart
• Docker Compose
• Terraform
• AWS (CDK)
• Production readiness
• Upgrades

For the AWS CDK construct, see packages/aws-cdk/README.md.

Connect your coding agent

Run the ctxpipe CLI in your repository:

npx ctxpipe init

The wizard writes the ctxpipe MCP config for Claude Code, Cursor, Codex, OpenCode, or VS Code. Any other MCP client can use the same endpoint. The wizard can also add local memory: Markdown lessons in .ai/memory that your agents keep with the code. For local memory only, run npx ctxpipe memory init. See packages/cli/README.md.

Documentation

• Product docs
• Getting started
• Workspaces
• Connections
• MCP setup
• Local memory
• Self-hosting
• CLI

Support and contributing

• Product or early access questions: ctxpipe.ai, or book a demo
• Docs and setup questions: docs.ctxpipe.ai
• Bugs and feature requests: GitHub issues
• Community chat: Discord

To contribute, see CONTRIBUTING.md. It covers local development, the repository layout, and the common commands.

License

ctxpipe uses the MIT License. See LICENSE for the binding terms. For more about the open-source core, see the open-source guide.