Every session starts cold
The decision you settled with Claude on Monday is invisible to Codex on Tuesday, so you re-explain the same context until you stop bothering.
For Claude Code and Codex
An extendable agentic workspace with a brain
Dryft gives Claude Code and Codex a shared local knowledge workspace. Review what your sessions learn, find answers with references and save useful workflows as reusable skills. Add optional stacks when you need more capabilities.
Works with what you already use
- Claude Code
- Codex
- Linux
- Mac
- Windows
The problem
Your agents are clever and forgetful
As you rely on agents for more of your work, it gets harder to keep track of decisions across sessions and check where remembered facts came from.
Memory nobody checked
An agent can remember a model’s guess as a fact. Without a source reference, you have no way to check where it came from.
Before / after
Give your agents knowledge you have reviewed
Without Dryft
- Every session starts from an empty context window
- Decisions live in transcripts nobody reads twice
- Memory tools keep whatever the model concluded
- An answer arrives with no source you can check
With Dryft
- Every session can search one brain of approved notes
- Facts, decisions and open questions become Markdown files
- Nothing enters the brain until you approve it
- Answers cite the notes and sources they came from
How it works
Work, review, ask
Work as you already do
Your active Claude Code or Codex session summarizes useful work as notes awaiting review. Capture depends on supported hooks and your agent trust settings.
Review what stays
Each summary waits for review. Approve it, edit it or reject it; review history records your decisions.
Ask your brain
Search follows links through approved notes only. Answers cite their evidence and say plainly where it is missing or disagrees.
$ dryft brain candidates
$ dryft brain review 9c41 --decision approve
✓ approved → brain/knowledge/
$ dryft brain search "stack ids" --depth 1
$ _Inside Dryft
Share knowledge and workflows across agents
Claude Code and Codex can share one brain, your workspace rules and reusable skills. Core works before you install a stack.
Knowledge you can review and cite
Keep concise notes in local Markdown. Approve, edit or reject captured knowledge before ordinary search uses it, with source references and conflicting evidence preserved.
Rules for how your agents work
Write a workspace Constitution with your preferences, permission rules and boundaries. Refine it through chat or DUI; it guides your agents but is not a security sandbox.
Workflows you can reuse
Review and save useful workflows as workspace skills for your enabled agents. Dryft checks existing skills and remembers proposals you decline.
Projects linked where they live
Register existing folders and have your agent read selected docs to get oriented. Project source stays unchanged, and proposed knowledge goes through review.
Questions you can answer in a browser
Say “use dui” for grouped questions, forms and custom interactive pages in a local browser. Follow up in the same session, or keep using chat.
Use cases
How Dryft fits your work
You are the whole team, so the context is all in your head
Save decisions as notes that your next session can find, whether you use Claude Code or Codex. You spend less time repeating context.
- brain/knowledge/
- /dryft-search
- decision
1 brain
shared by Claude and Codex
~/Dryft
Stacks
Add the capabilities you need
Stacks add workflows, skills, hooks, knowledge, templates and local capabilities. First-party and community stacks use the same format and installer. Dryft works with zero stacks, and installed stack guides are available offline.
Visit the stacks site- Inspect before installingDryft
Search a configured registry and check contents and prerequisites. Install from a local folder or a listing backed by a GitHub source archive. Dryft records where the content came from.
- Approve executable contentDryft
Inspect hooks and capability assets before trusting the exact content. Dryft checks paths, ownership conflicts and declared digests. Trusted code is not sandboxed by Dryft.
- Keep control of updatesDryft
Review and approve updates, or uninstall a stack. Dryft keeps your notes and user work, refuses conflicting edits and supports recovery after interruptions.
Security & privacy
Your brain is a folder you can read
Knowledge is stored locally, as Markdown in your workspace. Claude or Codex still talks to its own service under your own subscription; Dryft adds no hosted copy of what you keep.
You review captured knowledge before ordinary search uses it. Integrations stay scoped to your workspace, and interrupted captures can be recovered.
Summaries, never transcripts
Capture asks for a concise summary with its sources. Transcripts stay with your agent; the brain holds references, not copies.
Review before search
Approve, edit or reject captured notes before ordinary search uses them. Pending and rejected notes stay out of approved answers.
Resume interrupted work
Dryft detects duplicate captures and keeps pending work recoverable, so you can resume after an interruption without adding the same knowledge twice.
Your files left alone
Integrations stay inside your workspace and preserve unrelated agent settings. Linked projects receive no automatic agent registrations.
FAQ
Common questions
The docs cover setup, commands and troubleshooting. They are also installed in your workspace.
No. Dryft uses your active Claude Code or Codex session to summarize useful work. Your agent still uses its own service and subscription.
No. Your active agent captures useful facts, decisions and open questions as concise notes with source references. Full transcripts are not copied into your brain, and captured notes await your review.
No. The brain is Markdown in a workspace on your machine. Your agent still talks to its own service under your own subscription, as it does today.
Claude Code and Codex. Setup detects them, remembers which are enabled, and dryft agent enable adds the other one later without disturbing unrelated configuration. You can also run the core with no agent at all.
Both notes stay. Approval creates a separate note and preserves the existing one. Answers cite the evidence and explain the conflict.
No. Core includes the knowledge brain, reusable skills, linked projects and DUI. Stacks add workflows, skills, hooks, knowledge, templates and local capabilities when you need them.
Yes. A workspace Constitution tells agents how you want them to work, when to ask permission and what boundaries to respect. Refine and approve it through chat or DUI. These are instructions, not a security sandbox.
It can propose skills from repeated workflows in accessible completed sessions and report gaps in the history your agent provides. Nothing is saved without approval. Discovery from real session history still needs end-to-end testing.
We have manually tested Dryft on Linux with Python 3.12. macOS and Windows support remain unverified. Hooks and history access depend on your agent and its trust settings; some checks in live agent sessions are still incomplete.
Your brain is a folder of Markdown you can keep, move or delete. Stacks uninstall completely through the CLI and keep your notes when they go.
Help improve Dryft
Report a bug, suggest an improvement, or help make the docs clearer. Choose a template on GitHub to get started.
Give feedback on GitHubPrefer email or need to share something privately? Write to support@usedryft.com.