Evercore: Orchestration Engine for AI Agents and Complex Workflows

Python License

📢 [Basic Local Deployment] [External Postgres Deployment] [Production VPS Deployment]

Evercore is an orchestration engine for managing long-running LLM agents.

It makes it super easy to define an agent as a combination of prompt, memory and tools and it manages ticket lifecycles, task dependencies, and distributed worker loops for you.

By decoupling state management and declarative YAML workflows from your core application logic, Evercore provides a robust, pluggable foundation that simplifies the development and scaling of asynchronous, multi-agent systems.

How does it compare to useWorkflow, Temporal, Prefect

Area	evercore	Temporal	UseWorkflow.dev	Prefect
Main value	Embed a workflow engine directly in your app with minimal infrastructure	Maximum durability and correctness for long-running workflows	Durable TypeScript workflows with agent-friendly DX	Fast path to production orchestration for Python/data teams
Workflow functionality	Tickets/tasks, stage transitions, approvals, pause/resume, retries, event inbox	Full workflow model with signals, queries, updates, child workflows	Durable function-style workflows with resume/event patterns	Flows/tasks, retries, state tracking, deployments, automations
Scheduling	Built-in interval/one-shot schedule runner	First-class schedules with advanced policies	Code-level timing/sleep primitives	Strong deployment scheduling and automation features
Operational complexity	Lowest: lightweight and app-embedded	Highest: most powerful, more platform overhead	Medium: modern and fast-moving ecosystem	Medium: managed experience with solid operational tooling

Evercore is: embeddable, lightweight, fully python and comes with batteries included for AI agent flows.

Quick start

1. Configure environment:

cp .env.example .env 2>/dev/null || true
export EVERCORE_DATABASE_URL="sqlite:///./evercore.db"
export EVERCORE_WORKFLOW_DIR="./workflows"
export EVERCORE_DEFAULT_WORKFLOW_KEY="default_ticket"

1. Configure lemlem models:

export LEMLEM_MODELS_CONFIG_PATH="/abs/path/to/models_config.yaml"

Or configure a database-backed model config source for dynamic model/preset updates.

1. Install and run API:

uv sync --project evercore
uv run --project evercore evercore-api

1. Run worker in a second terminal:

uv run --project evercore evercore-worker

EVERCORE_WORKER_ID is optional. If unset, Evercore now generates a process-unique default (evercore-worker-<hostname>-<pid>). For multi-container deployments, explicitly setting a stable unique worker id per container is still recommended.

1. Create a ticket:

curl -s -X POST http://localhost:8010/tickets \
  -H 'content-type: application/json' \
  -d '{"title":"Hello","workflow_key":"default_ticket"}'

1. Enqueue a lemlem task:

curl -s -X POST http://localhost:8010/tickets/<ticket_id>/tasks \
  -H 'content-type: application/json' \
  -d '{
    "task_key":"lemlem_prompt",
    "payload":{
      "model":"openrouter:gemini-2.5-flash",
      "prompt":"Write a short test summary for this ticket"
    }
  }'

Tests (evercore library only)

Run the full standalone Evercore library suite:

uv run --project libs/evercore evercore-test

Run a subset by pattern:

uv run --project libs/evercore evercore-test --pattern "test_worker*.py"

If you run Evercore outside this monorepo, ensure the lemlem dependency is resolvable (for example from your package index, or by adding a local/path source in your own pyproject.toml).

Extending for a new project

1. Add a workflow YAML in workflows/
2. Define your own task keys in your app layer
3. Register custom executors in evercore/executors/registry.py
4. Create tickets/tasks through the API
5. For long-running custom executors, add execute_with_control(ticket, task, control) and check control.should_stop() periodically for cooperative pause/cancel support

LLM Layer (lemlem)

Evercore uses lemlem for model routing and LLM calls.

lemlem supports two model/preset configuration sources:

• YAML/JSON file (for example via LEMLEM_MODELS_CONFIG_PATH)
• database-backed model config service (dynamic runtime loading)