Agents

An AI staff that escalates, not improvises.

Agents in GPTree hold roles, run routines, use your connected tools, and stop for your approval the moment a decision needs a human. Autonomy is off until you turn it on, and it is double-gated when you do.

Start with a Pro trialSee plans

Agents are a Pro and Team feature

The company

Structure first. Autonomy second.

Your agents form a small company: an org chart with roles, crews for repeatable jobs, and meeting rooms that are real conversation trees you can open, read, branch, and join at any time.

Roles and the org chart

Each agent has a role, a scope, and a place in the chart. Delegation flows down; anything unusual escalates up, ending at you.

Crews

Groups of agents assembled for a job, from templates or from scratch, with versioning and publishing when a crew is worth keeping.

Meetings you can read

Agents meet in rooms that are ordinary conversation trees. Open one mid-meeting, read the turns, branch a tangent, or close it down.

Routines

Set the cadence. Review the output.

Routines run on a schedule, fire from a webhook, or wait for a manual click. The classic: a briefing agent that reads your connected inbox and calendar at 7:00 and posts the day's brief before you sit down.

  • Cron-style schedules, webhook triggers, manual runs
  • Run history with every output kept on the tree
  • Notifications when a run needs your attention
Pipelines

Deterministic steps for repeatable work.

Pipelines are saved multi-step procedures your agents can run: research this, draft that, compare the drafts, produce a final. They are versioned, so a pipeline that works keeps working, and you can roll back the one you broke.

  • Multi-step, deterministic, and reusable across agents
  • Versioned with one-click rollback
  • Built on the same Branch, Compare, Merge loop you use
Decisions

A ledger, not a vibe.

Every meaningful call an agent makes lands in the Decisions ledger. When two decisions contradict each other, GPTree flags the pair for you to reconcile. Every run writes a hash-chained audit timeline, so what happened is never a matter of memory.

  • First-class decision records with provenance
  • Contradiction detection with a reconcile flow
  • Hash-chained, tamper-evident audit timeline
The safety model

Autonomy you grant, not autonomy you hope for.

Off by default, double-gated on

Agents start with no autonomy. Turning it on takes an explicit grant, and high-risk actions still stop for approval every time, with an approve-with-edits option so you can fix a draft instead of rejecting it.

Write actions always ask

Reading your tools is one thing. Sending, posting, and editing are another: every write action through a connector waits for your explicit approval before it runs.

Credentials encrypted

Connector tokens, webhook secrets, and API keys are encrypted at rest. Connections are per-user, and disconnecting revokes access immediately.

Everything on the record

Runs, tool calls, and escalations write to a hash-chained audit timeline. If a question ever starts with "what did the agent do", the answer is a lookup, not an investigation.

Connectors

Your tools, on the tree.

Connect the tools your work already lives in. Agents read from them during routines, and chat can act on them inline, always with approval on writes.

GmailSlackGitHubNotionGoogle CalendarLinearAsanaJiraDayzie+ 1,000 more in the gallery

Connections are per-user and revocable in one click. Write actions always ask first.

For developers

GPTree Graft: your local machine, on a branch.

A single Python client that runs code, reads files, and drives local agents on your machine, while every tool call lives on a branch you can compare, replay, or merge. Outbound connection only; no ports exposed.

  • pip-installed bridge, connected over an outbound WebSocket
  • Code and terminal runs recorded on the tree
  • Pair it with the Chrome extension to branch from ChatGPT
Setup guides in Learn

OpenClaw bridge

Connect local OpenClaw agents and manage their reasoning from the tree: branch-scoped context and visual orchestration for agents you already run.

Pro

Cursor dispatch

Send background coding tasks to Cursor agents with your own key, and keep the results on branches you can compare before merging.

Pro

PentestGPT runtime

For security research teams: drive authorized testing workflows through the same branch, approve, and audit loop as everything else.

Pro
Get started

Never re-explain yourself.

Start free in under a minute. No card required. Use built-in credits or bring your own keys.

Start freeSee pricing