Build intelligent agents.
Run them reliably.
Reliable and Extensible Lifecycle Framework for Intelligent Agents.
Compose models, tools and memory. Keep approvals and receipts. Resume after a restart.
ticket: "INC-001"
run_01J_RALFPause for approval.
Resume from recorded evidence.
Try an illustrative ticket change: review and approve the plan, restart the host, then continue the original task with its recorded approval. Use the buttons below to follow the path.
- 01Review the operationReviewInspect the tool, arguments and digest
- 02Record approvalApproveApproval belongs to the original action
- 03Restart the hostRestartA new process reads the saved records
- 04Continue the runResumeAdopt the result and finish the run
waiting_approvalrun_01J_RALFAction IDaction_change_001Digest55b34230…2ee944ac shortenedToolapply_changeArguments{ "ticket": "INC-001",
"plan": "restart test cache" }ApprovalpendingDemo hostprocess_01Illustrative IDs and digest: approval will bind to this operation.
- Ticket read. The change is awaiting approval.
Review the tool, arguments and digest before approving.
Fixed browser-only data: no Python, model calls or business writes. Real recovery needs original data, matching configuration and business evidence; unknown writes still require reconciliation.
From one business tool.
To a complete agent application.
Start with one concrete task. Add conversations, knowledge and collaboration as needed, with explicit configuration, guides and recovery boundaries.
Models and Python tools
Connect an explicit model protocol and register business functions as tools. Define arguments, output constraints and approval rules.
For: Ticket lookups, order assistants, business actions
Workflows and teams
Express dependencies as static graphs. Add bounded parallelism, plans or teams. Parent and child tasks retain their own results and identities.
For: Stepwise diagnosis, routing, research and review
Durable conversations and memory
Link turns through sessions and retain confirmed information as versioned facts. Select recall and context policies explicitly.
For: Follow-up questions, preferences, multi-turn tasks
Documents and retrieval
Ingest documents with versions and sources. Choose lexical, vector or hybrid retrieval, trace the evidence and evaluate actual results.
For: Runbooks, product questions, knowledge assistants
External tools and references
Opt into MCP, HTTP, fixed CLIs, workspaces and skills. The host selects tools and scope; approval and receipt protocols still apply.
For: Existing services, internal scripts, files and instructions
Configuration, input and recovery
Pin components and execution limits. Handle approvals, user input and uncertain effects through their own controls, then observe and resume original tasks.
For: Version management, diagnosis, controlled execution
Start locally. Compose as needed.
Local + SQLite is the low-dependency starting point. External tools, remote stores and Temporal need explicit configuration within their experimental Alpha scope.
No model key.
Your first local run.
Install from source and generate an offline ticket project. A fixed provider proposes an operation for approval, so you can learn the execution protocol before connecting a real model.
- 01Install the SDK and create a project
Python 3.11+ · local SQLite
- 02Inspect and approve the operation
inspect → approve
- 03Resume the run and inspect its result
resume → succeeded
The first run should pause at waiting_approval. These commands install and start the project; the full tutorial covers installation, approval, recovery and result inspection on one page.
1. Install from source
git clone https://github.com/RalfNick/relf-agent.git
cd relf-agent
git checkout 0b9b6c9682c40285d91994cd831aa93d7e814bf1
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .2. Start the offline project
ralf --format text init ../my-agent
cd ../my-agent
python -m agent_app --ticket INC-001Next: connect a real model and Python tools
Find the complete Python example in the models and tools guide. A real model requires credentials and model selection, and incurs provider usage.
Read the models & tools guideYour next step starts here.
Capabilities overview
Choose capabilities by task, with clear defaults and optional components.
Quickstart
Complete an offline task from installation through approval and recovery.
Models & tools
Choose a model protocol and register your Python tools.
Sessions, memory & RAG
Organize conversations, lasting facts and traceable knowledge retrieval.
Workflows & teams
Compose static graphs, bounded parallelism, plans and teams.
Runtime & deployment
Choose a backend. Manage versions, limits and resource ownership.
Define what happens next.
Start your execution path with a local agent.
ALPHA · APIS & DEPLOYMENT CONTINUE TO EVOLVE