Repository navigation
Hermes Talk 0.22.0: run your Desktop sessions by voice
Run your Desktop sessions by voice: choose a project, start work in a new conversation, check another agent's progress, and send or steer instructions while your original call stays connected.
New: Desktop session and project controls
In an OpenAI Realtime call in Hermes Desktop, try:
- “Show my projects and start a session in this directory.”
- “Check what the other agent is working on.”
- “Tell that session to check the parser next.”
- “Steer the running agent toward the smaller fix.”
- “Check my capabilities, scheduled jobs, and messaging setup.”
Talk uses the host's actual session and project APIs. It verifies the target conversation, keeps operations on the captured connection/profile, and leaves the voice owner unchanged. New conversations inherit the profile's model and reasoning settings and use Standard speed. An opening message makes the new session appear in the sidebar.
Progress checks return saved replies, running/waiting state, task lists, and in-progress text when the host exposes it. Queued instructions are reported as accepted, not completed. If a write's network reply is lost, Talk reports an unconfirmed outcome and tells you to inspect before retrying; replaying the same tool call does not dispatch it again.
Fixes that came out of live testing
- Stop → Connect works in the same Desktop conversation. Stopping releases the call's microphone lease without permanently cancelling its controller.
- Hermes-managed ChatGPT sign-in is recognized. Readiness checks use the host's explicit read-only Codex OAuth resolver, including credential-pool logins.
- Capabilities work on current Hermes. Discovery accepts the renamed public skill-catalog functions instead of falling back to an unavailable HTTP server.
- Typed messages work during unbound realtime calls. They stay on the active data channel, and their item IDs fit the provider's 32-character limit.
- GPT-Live doctor no longer asks for an ElevenLabs key. ElevenLabs validation applies only to cascade mode. Thanks to @Sterling-Automation for #198.
- The Ruff dependency conflict is removed. Plugin extras no longer impose a competing exact Ruff pin on Hermes's dependency workspace.
Also included since 0.21.1
- All 28 documented Grok built-in voices, plus an expanded experimental speech-tag reference. Thanks to @zoidypuh.
- An optional Talk presence/command bus for integrating the call with another UI. Thanks to @gabrieljudah.
- More reliable steering receipts when a provider omits the echoed steer token, and isolation of ambient host registries in tests.
Verified with a real SSH backend
Tested with a Windows Desktop client and Ubuntu 26.04 SSH backend, both running stock Hermes v0.22.5, using OpenAI gpt-realtime-2.1 through ChatGPT/Codex OAuth.
Live canaries verified a new sidebar session, a second session in a selected project directory with an actual terminal working-directory check, a follow-up to another session, and steering that changed a running task's final reply. Saved results were read back after the Desktop backend restarted. The operator also confirmed spoken replies, background completion announcements, and interruption stopping speech.
The release goes through the Linux/Windows Python 3.11–3.13 test matrix, pinned Ruff, CodeQL, the Hermes plugin guard, packaging checks, and independent adversarial review. See PR #200 for validation and scope.
Install or update
Fresh install:
hermes plugins install --ref 7ad833e4fcedb2c9a9947d84d7fd89bdbd15d256 --enable TheSmokeDev/hermes-talkExisting pinned/catalog installation: preserve local plugin edits, then install this release explicitly:
hermes plugins install --force --ref 7ad833e4fcedb2c9a9947d84d7fd89bdbd15d256 --enable TheSmokeDev/hermes-talkFor an unpinned Git installation:
hermes plugins update hermes-talkRestart the gateway/dashboard and Desktop so they load the updated code. In Desktop, enable Hermes Talk's Desktop switch under Capabilities → Plugins, open a conversation, and choose Talk → Connect. For SSH, install the plugin on both the Desktop machine and remote host. Hermes v0.22.5 requires the full commit SHA shown above for --ref; a version tag is not accepted. This pins the published v0.22.0 source and avoids an older community-catalog pin.
Current boundaries
These new controls apply to unbound OpenAI Realtime calls in Desktop and stay within the selected connection/profile. They do not grant arbitrary screen control, external messaging sends, or scheduled-job writes. Hosts without a project catalog report directories observed in saved sessions instead. The common messaging report is configuration status, not a delivery test.
GPT-Live, the canonical recipient picker, and shared voice-history attachment still require their additional host capabilities. Direct gateway-URL deployments and macOS were not qualified end to end in this test run. See the Desktop guide and voice-command card.
Full changelog: v0.21.1…v0.22.0.