Agentic Ops is a workflow runtime and control plane for operational AI agents: it queues incoming work, launches an isolated Claude Code agent per task with the tools and memory that workflow needs, brokers human approvals/clarifications, and gives you a web UI over all of it. Workflow implementations (prompts, skills, connector instances, integration policy) live in your own workflow repository; this repo is the platform they run on.
- Workflow packages and bundle assembly — a workflow is
agent.yaml+settings.json+.mcp.json+ skills/hooks/agents, merged with shared platform/team assets into a versioned, checksummed bundle. - Claude Code runtime harness — ephemeral per-task containers with
subagents, human-in-the-loop approvals, a built-in clarification
(
AskUserQuestion) flow, and a sandboxed-credential deny-list so secrets never reach the model's shell. - MCP server framework — a public catalog of core (message, memory, platform self-service) and integration (Salesforce, Jira, Splunk, CloudWatch) servers, plus a pattern for adding your own.
- Source connectors and task queue — pluggable ingestion adapters (Pub/Sub, ServiceNow polling, ...) that turn external events into queued tasks, with alert coalescing.
- Two-way messaging and human oversight — agents and operators can exchange threaded messages, start work from configured keywords, request clarifications, and approve or reject sensitive actions. Mattermost and Slack are supported through real-time provider integrations and a pluggable messaging interface.
- Hindsight memory integration — long-term recall and pattern reflection across past incidents, plus per-agent memory volumes backed up to object storage.
- Web control plane — tasks, session replay, schedules, approvals, analytics, and platform catalogs (MCPs, connectors, memory, workflow-repo sync/versioning).
- Docker Compose and Helm deployment packages, with a provider-neutral
object-storage layer (
s3orgcs).
See docs/architecture.md for the full component breakdown and a diagram of the request/task flow (connector or webhook → task queue → session-manager → runtime container → MCP servers/message bus/approvals → completion → memory sync → control-plane UI).
cp examples/workflow-repo/platform-config.example.yaml platform-config.yaml
docker compose -f deploy/docker-compose.yml up --buildThis runs the base stack against examples/workflow-repo as a starter workflow repo. To point at your own workflow repo instead:
HOST_PLATFORM_CONFIG_FILE=/path/to/workflow-repo/platform-config.yaml \
HOST_WORKFLOW_REPO_PATH=/path/to/workflow-repo \
docker compose -f deploy/docker-compose.yml up --buildFor a guided setup that also handles secrets, remote workflow repos, and
Kubernetes targets, use make bootstrap — see
docs/deployment.md.
| Doc | Covers |
|---|---|
| docs/architecture.md | Components and the request/task flow. |
| docs/configuration.md | The three config layers, full platform-config.yaml reference, bootstrap env reference. |
| docs/connectors.md | Connector model, available connectors, and how to add one. |
| docs/mcps.md | Core vs. integration MCP servers and their config. |
| docs/memory.md | Long-term memory (Hindsight), per-agent project memory, hooks, and weekly reflection. |
| docs/deployment.md | make bootstrap, Compose/Helm, bundle sync and versioning. |
| docs/security.md | Permissions, sandbox/credential controls, bundle integrity, approvals, encrypted secrets. |
| docs/workflow-authoring.md | How to build a workflow package (agent.yaml, skills, hooks) and its bundle. |
This repository intentionally ships no company-specific workflows, Salesforce policies, private connectors, customer memory-bank maps, or private deployment overrides — those belong in a workflow repository. See CONTRIBUTING.md.
python -m pytest -q tests/unit # unit suite, no infra required
python -m ruff check . # lint
python -m ruff format --check . # format check
cd control-plane-ui && npm ci && npm run build # control-plane UImake test runs the full unit + service (real Postgres) + runtime (real
Docker) suite; see the Makefile (make help) for every target.
See docs/roadmap/ for planned work and CONTRIBUTING.md for how to contribute.
Apache License 2.0. See LICENSE.