# GLMChat V2.2 User Guide

Updated: 22 July 2026

## 1. First use

1. Open GLMChat on the HTTPS site.
2. Create and confirm a six-digit PIN.
3. Enter the Together API key and use the validation action.
4. Review the privacy and offline explanation.
5. Install the PWA when the browser offers installation.

The API key is encrypted at rest and decrypted only when required. Protected destructive actions may request the PIN again.

## 2. Home tiles

- **Build an Agent:** create a specialist with instructions, memory, tools, approval policy and limits.
- **Run Automation:** create a typed multi-stage workflow.
- **Create Agent Team:** combine specialists using manager-led, sequential, parallel or consensus coordination.
- **Tool Actions:** run approved project and verification actions.
- **Projects:** organise files, instructions, chats, automations and activity.
- **Templates & Prompts:** save and reuse prompts and typed automation definitions.
- **Offline & Sync:** inspect drafts, queued actions, conflicts and retries.
- **Usage & Costs:** review estimates, limits and execution telemetry.

Status badges show running work, offline queues, conflicts or cost warnings without opening each section.

## 3. Guided wizards

Every automation wizard uses the same pattern:

1. Choose the goal or task type.
2. Select the project and visible context.
3. Configure roles, stages or tools.
4. Configure approvals, memory and limits.
5. Review the exact typed configuration.
6. Save as a draft or template, or launch.

Drafts are stored in IndexedDB and can be resumed. Unsupported values prevent launch rather than being silently ignored.

## 4. Chat

- Select a project and only the files that should be included.
- Attach supported images and files visibly before submission.
- Select reasoning and structured-output options when needed.
- Review the estimated maximum cost before sending.
- Use Stop during streaming, then retry or regenerate where appropriate.
- Copy, share or export a conversation as plain text.
- Save a useful response as a prompt or convert it into a preseeded agent or workflow.
- Use Speak to generate TTS playback when the provider is available.

When offline, chat text remains a local draft. GLMChat does not silently submit it after reconnection.

## 5. Projects

Each project provides:

- **Overview:** name, instructions, permissions, budget and retention.
- **Chat:** project-aware conversation with visible selected context.
- **Files:** upload, inspect, select, edit or remove supported files.
- **Automations:** project agents, teams, workflows and run history.
- **Activity:** project changes, outputs and audit information.

ZIP imports are validated for traversal, symlinks, size and decompression limits.

## 6. Agents and teams

An agent definition controls its role, instructions, model behaviour, memory, allowed tools, approvals and execution limits.

A team contains separate workers. Coordination options determine real execution behaviour:

- **Manager-led:** a manager delegates work to specialists.
- **Sequential:** each worker hands its output to the next.
- **Parallel:** specialists work concurrently and results are combined.
- **Consensus:** proposals and reviews are compared before a decision.

The live run screen shows actual workers, assignments, outputs, costs and approval states.

## 7. Workflows and run controls

Workflow stages can define dependencies, assigned agents, tools, expected outputs, retries, cost/token limits, timeouts and checkpoints.

Available controls depend on the current state:

- **Pause:** requests a checkpoint-safe pause.
- **Resume:** resumes from the persisted checkpoint without duplicating completed actions.
- **Cancel:** stops the run.
- **Retry:** retries a permitted failed stage within its limit.
- **Skip:** skips an optional stage.
- **Intervene:** provide guidance, revoke a tool, reassign a worker, approve a budget change or branch from a checkpoint where permitted.

## 8. Tool approvals

Tools are classified as read-only, mutating or destructive. Before approval, GLMChat shows:

- the exact tool;
- validated arguments;
- affected project/files;
- whether the operation is reversible;
- the required approval level.

Destructive operations require recent-PIN reauthentication. Approved arguments cannot be altered afterwards.

## 9. Offline and synchronisation

The Offline & Sync screen shows:

- local drafts;
- pending mutations;
- authentication-paused operations;
- exhausted retries;
- conflicts;
- recently synchronised actions.

Conflicts can retain the local version, retain the server version, or use a supported manual merge. Queued operations preserve identifiers, request hashes and ordering.

## 10. Privacy and recovery

Privacy controls allow supported export, selective deletion, full reset and offline-access revocation. The error-recovery surface can retry, install a waiting update or clear only stale application-shell caches. It does not clear IndexedDB, projects or queued mutations unless a separate destructive action is explicitly confirmed.