Skip to main content
Kheish runs AI agents as a daemon rather than embedding them inside a single process. It is designed for persistent sessions, detached runs, approval-gated actions, tool execution, and multi-agent workflows over a stable HTTP/SSE control plane. The daemon is the center of the ecosystem:
  • the daemon owns the runtime and durable state
  • the daemon includes built-in connector support for HTTP, Slack, and Telegram ingress
  • the Python SDK talks to the daemon from a separate package
  • the separate kheish-connectors repo adds external sidecar connectors around the same daemon
  • capture pushes observations into daemon-owned sources from a separate host-local repo

​
Start in this order

  1. Read Start Here to understand what runs where and which path to take first.
  2. Run Quickstart for one working daemon, one session, and one completed run.
  3. Then choose what you need next:

​
What Kheish gives you

  • Persistent sessions with journaled state, checkpoints, and restart recovery
  • Detached runs that can pause on approvals or structured user questions and resume later
  • Daemon-owned assets and multimodal inputs for text, documents, images, and retained audio
  • Daemon-owned observation sources and materialized capture flows for screenshots, webcam frames, microphone segments, and other external producers
  • Daemon-managed personas that bind immutable persona snapshots into sessions
  • Immutable Playbooks and Flow projections for evidence-backed procedures over normal runs
  • Durable public channels with shared message logs, reactions, and autonomous thread turns over member sessions
  • Daemon-owned projects with durable membership, linked channels, and project-task execution over assigned sessions
  • Multi-agent execution with sidechains, tasks, schedules, mailboxes, and background shell tasks
  • Extensible runtime primitives for providers, tools, hooks, skills, MCP, connectors, and output delivery
  • Operator controls for runtime reconfiguration, debug capture, and control-plane authentication

​
Learn the model

After the first successful run, continue with:
  1. Overview for the product model.
  2. Sessions and runs for the execution backbone.
  3. Playbooks and flows before adding evidence-backed procedures over runs.
  4. Approvals and questions for blocked and resumable work.
  5. Providers and routing for route_id, auth_ref, selector grammar, and route pinning.
  6. Assets and multimodal input before integrating documents or images.
  7. Capture and observations before integrating host-local capture or durable external media ingestion.
  8. Channels and public conversations before modeling shared discussion among humans and agent-backed sessions.
  9. Projects and project tasks before modeling durable coordination across multiple sessions and channels.
  10. Connectors and reply targets before wiring Telegram, Slack, HTTP, or durable session reply defaults.
  11. Personas before binding long-lived identity instructions into sessions.
  12. Security and auth before exposing the control plane beyond localhost.
  13. Running the daemon before operating a real instance.

​
Minimal first-run CLI surface

For the first daemon success, the smallest useful CLI set is:
  • serve
  • runtime get
  • sessions create
  • sessions input --wait
If a run pauses instead of completing, then inspect:
  • approvals list
  • tasks list
  • runs get

​
Audience

This documentation is written for:
  • platform and infrastructure teams operating Kheish as a daemonized service
  • application teams integrating sessions, runs, approvals, or connectors into other systems
  • contributors working on the runtime, daemon, or surrounding integrations