Claude Code, Codex, OpenCode, OpenClaw or Hermes?
Run all five and compare them yourself.
Orkas runs all five as local sessions in one desktop app, so the table below is what a host actually observes when it drives them — not a feature list copied from their docs. You pick per task instead of picking once.
What actually differs between the five
Every row below is a contract Orkas has to honour to drive that CLI, so it is measured from the driving side rather than claimed from a docs page. A dash is a limit of that CLI's transport, not a verdict on its model.
| Capability | Claude Code | Codex | OpenCode | OpenClaw | Hermes |
|---|---|---|---|---|---|
| Resume a session | Native | Native | Native | Same session id | Orkas replays the history |
| Durable instructions | Native field, per run | Native field, per session | First user turn, per session | First user turn, per session | First user turn, per run |
| Binds to a project directory | ✓ | ✓ | ✓ | — | — |
| Gets the Orkas tool bridge | ✓ | ✓ | ✓ | — | — |
| Reads and writes Agent memory | ✓ | ✓ | — | — | — |
| Accepts a message mid-run | ✓ | ✓ | — | — | — |
| Permission control from Orkas | Inherit / Ask / Full | Inherit / Ask / Full | Full access only | Inherit only | Inherit / Ask / Full |
Claude Code and Codex are the two that take the full Orkas contract: project binding, tool bridge, Agent memory, mid-run messages and a per-Agent permission mode. OpenCode takes the bridge but runs in full-access mode only. OpenClaw and Hermes are session-only, so Orkas keeps them for self-contained turns.
One conversation, multiple bounded workers
Multi-agent is useful when a task spans research, code, files, analysis, and follow-up. The lead agent keeps the global goal while sub-agents handle narrower work with their own skills.
Lead agent owns the goal
It understands the request, plans the work, recruits sub-agents, and keeps the final result coherent.
- plan
- coordinate
Sub-agents do focused work
Each sub-agent can carry a role, skill set, and context boundary, so complex tasks do not collapse into one overloaded thread.
- roles
- context
Skills make actions real
Agents call skills for file work, code execution, knowledge search, connectors, and other concrete operations.
- skills
- tools
Sub-agents become workflows
Recurring work can be saved as a sub-agent, then summoned again from chat with the same instructions and boundaries.
- reuse
- workflow
For work that needs more than one assistant persona
Research
Split collection, synthesis, and critique across different sub-agents.
Engineering
Let one agent inspect context while another edits, tests, or explains diffs.
Operations
Turn reports, checklists, and follow-ups into repeatable agent routines.
Data work
Separate extraction, cleaning, analysis, charting, and final explanation.
Choosing between the CLI backends
Which should I use: Claude Code, Codex, OpenCode, OpenClaw or Hermes?
Claude Code and Codex are the only two that accept the full host contract — project binding, tool bridge, Agent memory, mid-run messages and a per-Agent permission mode — so they suit long repository work. OpenCode takes the tool bridge but only in full-access mode. OpenClaw and Hermes are session-only, which fits self-contained turns. In Orkas you install several and pick per task.
Can I run Claude Code and Codex at the same time?
Yes. Each external agent you add gets its own local CLI session, so several can run in parallel in one window and you read their output in the same chat. Claude Code and Codex also accept a new message while a run is still going; the other three finish the turn first.
Are sub-agents the same as workflows?
They can be. A sub-agent can carry instructions, skills, and output expectations for a recurring process, so you can summon it again instead of rebuilding the flow.
Is Orkas local-first and open source?
Yes. Orkas is open source and local-first. Agent configs, chats, generated files, and encrypted model keys stay on your machine by default; model calls go directly to your provider.
Can I switch backends part-way through a project?
Yes. Each external agent is a separate local session, so adding a second backend does not disturb the first. What does not carry over is the session itself: Claude Code and Codex resume natively, OpenCode resumes too, OpenClaw reuses a session id, and Hermes has no resume handle at all — for that one the host replays the visible history into a fresh session.
Build your own AI team on the desktop
Open source, local-first, and built around reusable agents and skills.