Use an ordinary request
After setup and trust, the connected agent can follow the same file-work procedure while handling your usual requests. Questions and planning alone do not create a workspace.
PREVIEW RELEASE0.4.0rc2
Local workspaces
for AI agents.
Retulo gives AI agents a dedicated local workspace for file-producing work. It keeps source material, output versions, decisions, and checks together. Identify the same task in a new conversation to continue from its recorded state.
0.4.0rc2Preview release
macOSLocal filesystem
3.11+Python required
MITOpen source
01 / THE FOUNDATION
Keep the work together.
A task is more than its latest file. Retulo keeps source material, output history, decisions, and checks organized for your agent.
Continue the same task
A stable task ID and workspace connect the work across conversations. Provide the task, workspace path, or an actual artifact reference when continuing.
Keep sources and earlier versions
Reserve output destinations, register saved files, and retain managed previous outputs. Existing code repositories stay in place.
Know what is done
Keep a concise current guide with decisions, remaining work, checks, and approvals. A saved file, checked content, and user approval remain distinct.
Run the file-work continuity example
.venv/bin/python scripts/demo_continuity.py --output-dir /tmp/retulo-continuity-demoUse a new, empty output directory. Separate CLI processes create a guide, inspect the same task again, and save a second version while preserving the source and first result. This checks task continuity; it is not a test of a new Codex conversation.
02 / HOW IT WORKS
Your agent runs the tools.
Retulo keeps the task in order.
Describe the work
The current agent interprets your request and makes the goal and completion criteria explicit.
Connect the task
The agent starts a workspace or identifies the existing one before producing files.
Save and check
Retulo records versions and file integrity. The agent or user evaluates content and records the result.
Continue with context
Read the short current guide, then retrieve detailed records when the next step needs them.
The core runs locally. It does not launch models or agents, execute recipe commands, watch folders, or publish results.
03 / QUICK START
Start from the source checkout.
Unpack the source archive and open a terminal in its project directory. This preview uses source installation.
Install the CLI and MCP core
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.lock
.venv/bin/python -m pip install --no-deps -e .
.venv/bin/retulo --helpThe final command should display Retulo’s available commands.
Preview the Codex integration
.venv/bin/python scripts/install_codex.py
.venv/bin/python scripts/install_codex.py --applyThe first command previews changes. Apply them when ready, then review and trust the exact hooks in Codex. Keep this source checkout in place. Existing MCP connections are not restarted automatically.
Then ask for ordinary work
Create a visitor guide as a Markdown file. Preserve the source material.
In another conversation, include the existing workspace or file reference
Continue this guide in the same workspace. Add a contact section and keep the previous version.
04 / UPDATE AND REPEAT
Change the input.
Keep the useful work.
For declared workflows, Retulo connects inputs, production steps, and exact outputs. The agent can see which steps need work after an input changes.
Rebuild affected steps
The host agent reruns the necessary tools while earlier final files remain preserved. Undeclared dependencies are not automatically discovered.
Reuse a workflow recipe
Use versioned inputs and parameters to create a new task. Its checks start fresh; previous approvals and verdicts are not copied.
Run the complete local example
.venv/bin/python scripts/demo_incremental.py --output-dir /tmp/retulo-report-demoChoose a new, empty output directory. The example creates an HTML report from the checkout’s inventory, updates affected steps, preserves the earlier report, saves a recipe, and creates a new report from another inventory. It makes no model calls and changes no global installation settings.
05 / DOCUMENTATION
Go from the first task
to the full contract.
- READMEOverview, source installation, and everyday use.
- Workflows and recipesDeclared steps, exact inputs and outputs, and new-task reuse.
- IntegrationThe host agent, context hooks, and task coordination.
- QualificationTested environments, actual methods, and current limits.
- RestoreInstallation snapshots and guarded restoration.
- Frequently asked questionsCommon setup, continuity, and validation questions.
A few useful distinctions.
Does Retulo remember every conversation?
Retulo records managed tasks, files, and progress. To continue across conversations, identify the same task or workspace. The core does not reconstruct conversations it has not received.
Does Retulo write or assess the content?
The current agent interprets the request and runs its tools. Retulo manages the workspace and records file integrity, checks, and approvals. Recorded checks do not guarantee content quality.
What is supported today?
The qualified target is macOS with a local filesystem, using Python 3.11 or later. Other operating systems, network storage, multiple hosts, and other clients need separate validation.
What do I need to install and connect?
Install Python 3.11+ and the package dependencies, then configure local workspace and source roots. Use the CLI or connect the MCP server to your existing agent. Codex guidance and hooks have a separate installer; Retulo adds no AI runtime or background service.
06 / HELP AND CONTRIBUTIONS
Make the next task easier.
Start with the guides, share a minimal reproduction for a bug, or describe the work a proposed feature should help you complete.
Preview release. Install from the GitHub source checkout, and read the qualification record for tested environments and current limits.