Everything your team's AI Desk does.
From the first prompt to the promoted branch - how TaskHound turns a pile of solo agent setups into one structure your team controls: your team, its secure Containers, and the projects inside them, with every agent working in the open.
Run any agent. Steer it mid-flight.
Start a session with Claude Code, Codex, Gemini, or Grok in any project. Watch the streaming transcript - assistant text, tool calls, diffs - and type into the run to redirect it, exactly like pairing. Permission requests (run this command? edit this file?) route to you on every surface.
- Two modes per session: TaskHound's structured chat/tool-card UI, or the vendor's real terminal UI streamed to your browser.
- Fork any session - live or historical. The fork starts on a branched worktree and reads the parent's transcript, so handoffs actually hand things off.
- Every session is history: transcripts are searchable and browsable per project, forever.
▸ tool: read knowledge/auth-flow.md
▸ tool: edit auth/session.go (+41 −12)
● permission: run go test ./auth/... ? ApproveDeny
One board. Every agent. Live.
The board is the team's heartbeat: a real-time grid of every running session - who, which agent, which project and branch, elapsed time, last action, pending-permission indicator. Filter by person, project, agent, or status. Jump into any transcript mid-stream. Admins can pause or stop anything.
The board shows agent work, the same way CI shows builds. Teammates watching a session appear as presence avatars - review becomes ambient.
When all concurrent agent slots are busy, new sessions queue on the board in the open - never a silent failure, never a mystery.
Running Pipeline deploys sit across the top of the board with their *.th.works URLs - what's shipped and clickable, right now.
auth-flow.md
pinned · always loaded into new sessions
query-cache-learnings.md
auto-distilled · from session #241 (Priya, Gemini)
deploy-runbook.md
edited by Mara · 2 days ago
The tenth session starts smarter than the first.
Every project carries a persistent knowledge base - docs, notes, and learnings auto-distilled from finished sessions. Agents load the relevant knowledge automatically at session start, whichever vendor they come from. What one person's agent figures out today, everyone's agent knows tomorrow.
- Auto-distillation: session ends → decisions, gotchas, and conventions get summarized into an editable learning.
- Pin the essentials so they're always in context; everything else is retrieved by relevance.
- It's plain markdown in your Container - edit it in the app, from a terminal, or with an agent.
Secure Containers, with a big undo button.
Your team gets secure TaskHound Containers - dedicated, fully isolated environments you control. Projects live inside a Container; agents can install tools, run tests, and build freely in there, not on anyone's workstation. And because every Container snapshots itself, nothing an agent does is permanent until you say so.
Every agent session starts with an automatic snapshot of the project, plus scheduled snapshots on top. Retention scales with your Container tier.
Roll a single project back on a timeline - or restore the whole Container for catastrophic cases. A safety snapshot is taken before every rollback.
Terminal and SSH access whenever you want it. Repos are plain git checkouts, knowledge is plain markdown, and your data exports anytime.
CI/CD for AI agents, with a human gate.
Every project gets two remotes: your real origin (GitHub/GitLab - optional) and hound - the TH Git inside your Container. Agents can only push to hound; they never hold origin credentials. From there, the Pipeline takes over.
-
1 · Push
Agent pushes fix/cart-totals to the hound remote.
-
2 · Build
Pipeline detects the Dockerfile or compose stack and builds automatically.
-
3 · Deploy
Deploys to the sandbox/dev environment inside your TaskHound Container.
-
4 · Review
Live at fix-cart-totals--acme.th.works with diff view and streaming logs.
-
5 · Promote
A developer promotes the reviewed branch to origin. Teams without an origin use TH Git as canonical.
▸ Agent work never reaches your origin without a human decision - that's the whole design.
Every branch gets a URL your client can click.
Pipeline builds - and any manual deploy - go live at https://<branch-or-app>--<team>.th.works in seconds, TLS included. Demos show on the board with status, logs, and resource usage; anyone can open the URL, admins can stop it.
- Access controls per deployment: public, basic-auth, or team-only.
- Agents use it end-to-end: build, deploy, read the logs, fix, redeploy - and post the URL in the transcript.
- Idle demos auto-sleep; caps and per-deployment limits keep your Container healthy.
▸ push received: hound/feature/checkout-v2 ▸ Dockerfile detected - building image… ✓ build complete (42s) ▸ deploying to sandbox… ✓ live: https://checkout-v2--acme.th.works ▸ awaiting review · promote when ready
The Desk follows you: web, phone, and native desktop.
Status is a core feature, not an afterthought. The same board, sessions, and deploys are live on the web dashboard, iOS and Android apps, and native macOS, Windows, and Linux apps - with push notifications when an agent finishes, fails, or needs input.
Check the pack at lunch. Read transcripts, quick-reply to steer your own sessions, approve a pending permission, open a fresh demo URL to show a client.
Everything the web has, plus tray live-status with running-session count, native notifications, one-click terminal to your Container, and the bundled hound CLI.
A session looks the same on every surface - avatar, agent badge, project, status pulse. Deep links resolve on whatever device you're holding.
Built so agents can't hurt what matters.
Your Containers run on fully isolated, dedicated infrastructure - never shared with another team, never multi-tenant. The control plane never executes your code.
Agents hold credentials for the hound remote only - never your origin. Each seat's agent logins live in their own isolated space and are never shared between users.
Team and project variables are encrypted at rest (AES-256-GCM, per-team envelope keys), masked in the UI, and scrubbed from transcripts before storage and indexing.
Role-based access (Owner, Admin, Member), TOTP two-factor, an audit log for privileged actions - and every agent action is on the record in its transcript.
Questions our answers didn't cover? hello@taskhound.com - a human reads it.