AgentHydra

One tab for every local AI session

AgentHydra is a local dashboard for AI coding agents that brings Claude Code, Codex, and OpenCode session history into one browser tab. The README explains the gap it fills: when you run AI tools in more than one place, nothing shows all of that local history at once, so you alt tab to remember which account is which, whether a session is still going, and what you asked it to do. AgentHydra reads the session data those tools already write and displays the conversations together, newest first, filterable by provider or recency, with Claude sessions also filterable by project or desktop instance. Clicking one lets you read the conversation and live tail it while it runs. Codex is read from its rollout JSONL files, while OpenCode CLI and Desktop share one local SQLite session store. For Claude, a message box at the bottom lets you type straight back into a session without finding its terminal, and you can send the same message to several Claude sessions at once. Codex and OpenCode support is read only. An optional ChatGPT handoff turns a task and repository into a bounded Markdown context file and opens ChatGPT, with common secret files omitted, though you still review the attachment. Raw Claude or Codex files can be opened in your editor or downloaded under their real title.

Queue and analytics

Beyond viewing history, AgentHydra lets you queue Claude runs, each with its own prompt, working directory, model, effort, permission mode, and account. You can run one on demand or let a scheduler drain the queue, one at a time or a few at once, with spacing so you are not hammering anything. Any queued run can be given a start time, so a delayed task is a checkbox rather than a cron job. The README says two things make this survive real use: runs reattach after a restart, so quitting the app or an auto update does not kill in flight work, and an optional Orchestrator watches every live session each minute and publishes what needs attention, such as a finished chat waiting, a context that outgrew its thread, a dirty repo, an account near its weekly wall, or a thread killed by a restart, with a reviewer chat checking each action. A rate limit is not a dead end either; stopped sessions can resume after the window resets, or move to another account with headroom. The analytics view breaks cost down by day, model, project, and account, and shows when in the week you work and which tools get used. Costs use published list prices and the CLI can print the same numbers as JSON. The whole store stays small because no message text is kept.

Privacy and install

AgentHydra is local first with no cloud service and no account signup. It reads the local stores your tools already write and talks to localhost, and optional external handoffs open the provider and leave the final upload to you. Install on Windows is a single line that places the app under the local app data folder without administrator rights and verifies the download against the release's published SHA-256 before unpacking, with no way to skip that check. Other platforms download a build from the releases page, and a source checkout uses Bun with a build and start command. The UI runs at localhost on port 7787. A fake mode sets an environment variable so dispatch uses a stand in for the claude CLI and touches nothing on your quota or repos, while instance actions still act on real desktop instances. The one thing that leaves the machine on its own is a periodic update check against a LunarWerx relay that mirrors GitHub's feed; it sends the app version, a coarse OS tag, and a random per install id, and never a hostname, username, file path, account, or email, though it derives a coarse location and network ASN. Setting an environment variable opts out and sends the check straight to GitHub's API with no telemetry. The project is compared openly with Claude Squad and Conductor, which launch new worktree sessions, while AgentHydra reads existing history and adds a queue and analytics on top.

Editorial conclusion

The project is written in TypeScript, runs as a local daemon with no account or cloud service, and is released under the MIT license.