Tip
Start with the docs site: curd.aerospeedsta.dev
CURD is a code-intelligence control plane for humans and agents.
It helps you:
- understand a codebase structurally, not just textually
- mutate code safely inside a shadow workspace
- inspect graph impact before committing changes
- author repeatable workflows in
.curdsource scripts - compile those workflows into governed plan artifacts
You can use CURD as:
- a serious solo coding tool
- a supervised agent backend
- an internal execution substrate for larger agentic systems
CURD is the backend and control plane.
The broader product direction around it is:
CURDthe code-intelligence, session, plan, and execution substrateCURD-CODEX(future)an agent-facing coding surface with CURD as the backendWikiCURD(future)a human-first GPUI management surface intended to bring developers and PMs closer to the real execution state of the codebase
That split is intentional:
- CURD owns correctness, routing, sessions, and code intelligence
- agent surfaces use CURD instead of bypassing it
- management surfaces stay human-first instead of turning into another opaque planning tool
Most coding workflows still make you choose between:
- fast but sloppy local automation
- heavy but awkward governed automation
CURD tries to keep both layers:
- the source layer stays readable
- the execution layer stays explicit
That is why CURD has:
search,graph,read, andedit- workspace shadow sessions
- profiles and runtime ceilings
.curdsource scripts- compiled plan artifacts
0.7.1 is the release where CURD starts to feel like a real platform instead of a loose bag of tools.
Highlights:
- BM25/FTS-backed ranked search with safer indexing behavior
- stronger separation between
curd-coreandcurd - shared validation across CLI, REPL, and MCP
- profile-gated runtime behavior on top of
lite/fullceilings - real
.curdauthoring flow:run checkrun compileplan editrun
- plugin hardening for
.curdland.curdt - broader package-manager and launcher support
- much clearer local and public docs
See CHANGELOG.md for the full release detail.
CURD has four important layers:
-
code intelligence
searchgraphread
-
safe mutation
edit- workspace sessions
- shadow diffs
-
authoring
.curdscripts
-
governed execution
- compiled plan artifacts
- profiles
- policy
If you only want local coding help, you can stay mostly in layers 1 and 2.
If you want agentic workflows, layers 3 and 4 become important.
curd init
curd doctor .
curd search alpha
curd read src/lib.rs::alpha
curd graph src/lib.rs::alpha --direction both --depth 2curd workspace begin
curd edit src/lib.rs::alpha --code "pub fn alpha() {}"
curd workspace diff
curd workspace rollbackcurd run check fix_alpha.curd
curd run compile fix_alpha.curd
curd plan edit <plan-id>
curd workspace begin
curd run fix_alpha.curd
curd workspace commit.curd is the human-facing workflow language.
It supports:
use profile ...use session ...arglet- multiline strings
- tool-call statements
sequenceatomicabort
It also supports structured explainability comments:
# explain:# why:# risk:# review:# tag:
Compiled plan artifacts preserve that metadata for review and execution context.
CURD has two runtime ceilings:
fulllite
And profile-based capability gating in settings.toml.
That means:
- the runtime ceiling defines the outer boundary
- the active profile defines what the actor can do
- policy can still deny a request
This is what makes CURD usable for both:
- indie/local workflows
- enterprise agent harnesses
Search is not an afterthought in CURD.
0.7.1 adds BM25/FTS-backed ranked retrieval to improve top-down entry into a graph-shaped codebase. The graph gives structural context; ranked search gets you to the right place faster.
Use:
searchto enter the codebasegraphto understand neighborhood and impactreadto materialize the exact code you care about
CURD supports two signed plugin package formats:
.curdlfor language ecosystem packages.curdtfor native tool packages
This release also wires .curdl build_system metadata into real build detection and adds clearer CLI plugin management flows.
CURD is distributed via native OS packages, language-bound wrappers, and container images.
| OS | Method | Command |
|---|---|---|
| macOS / Linux | Homebrew | brew tap aerospeedsta/curd && brew install curd |
| Windows | Winget | winget install curd |
| Windows | Scoop | scoop bucket add curd https://github.com/aerospeedsta/curd-scoop.git && scoop install curd |
| Arch Linux | AUR | paru -S curd-bin |
| Debian / Ubuntu | APT | sudo dpkg -i curd_amd64.deb (Download from GitHub Releases) |
| Fedora / RHEL | DNF | sudo rpm -i curd_x86_64.rpm (Download from GitHub Releases) |
Run CURD instantly without system-level installation using our language wrappers:
# Python (via uvx)
uvx --from curd-python curd --version
# Node.js (via bunx)
bunx --bun curd-node --versionServe the CURD MCP server via Docker:
docker run -it --rm -v $(pwd):/workspace aerospeedsta/curd:latest- Python:
pip install curd-python - Node.js:
npm install curd-node
For manual pre-compiled binaries and advanced setup, visit curd.aerospeedsta.dev/setup.
- docs/README.md
- docs/GETTING_STARTED.md
- docs/SCRIPTS_AND_PLANS.md
- docs/TOOLS_AND_PROFILES.md
- docs/CONTROL_PLANE_INTERFACES.md
- docs/PLUGIN_EXTENSIONS.md
- docs/PACKAGE_MANAGERS.md
- examples/README.md
curd-coreAuthoritative business logic and engines.curdControl plane, CLI, REPL, routing, and MCP surface.curd-pythonThin Python binding.curd-nodeThin Node.js binding.
CURD is now a serious backend for supervised and structured agentic coding workflows.
It is still in beta, but 0.7.1 is the point where the architecture is much more honest:
- one control plane
- safer sessions
- shared validation
- better docs
- real script-to-plan flow
It also now makes more sense as the center of a larger product family:
- CURD as the substrate
- CURD-CODEX as the agent execution surface
- WikiCURD as the GPUI management layer
That is a stronger position than treating each of those as separate disconnected tools.
Copyright (C) 2026 Aerospeedsta.
The current prototype of CURD is released under the GNU General Public License v3 (GPLv3).
This repository and its contents are strictly prohibited from being used for training, fine-tuning, or augmenting large language models or other AI systems without explicit written permission from the author.