AI Chat

Supported providers

Every conversation remembers which provider created it and automatically restores the model when reopened. Individual messages also store their provider, so you can mix Claude, OpenAI and Gemini answers in the same thread.

Reasoning & effort

Every provider exposes its own extended-thinking scale. Pocket Code surfaces them through a single model so you pick how hard the model thinks per conversation — and per agent in auto-iteration.

• •Tiers are discovered dynamically by probing each official CLI at startup — not hardcoded — so new levels appear automatically as the CLIs update.
• •A single value object carries both formats: the provider's native tier name and the equivalent integer thinking budget — each client uses whichever it needs.
• •Wired the right way per path: thinking block (Claude REST), --effort flag (Claude CLI), reasoning_effort field (OpenAI), -c model_reasoning_effort (Codex), LSP params (Copilot).
• •A pink Reasoning chip in the chat toolbar lists the active model's tiers with descriptions; it hides itself when the model exposes no reasoning.
• •In AI Office auto-iteration each agent role can be assigned its own reasoning tier, falling back to the session default.

Conversation modes

Three levels selectable from the chat top bar:

37 tools across 11 categories

The agent can read, write, search, run commands, control Git and connect to external services. Each category has granular individual permissions per provider and model.

Granular permissions

From Settings → AI → Tool Permissions you decide exactly what each model can do. Granularity per provider, per model or per category, with individual switches per tool. Quick Allow/Deny all by section.

• •4 provider chips (OpenAI, Claude, Gemini, Copilot) with optional scope to a specific model
• •Filtering applied at every layer: tool builders, MCP HTTP server and Copilot LSP
• •Binary toggle to disable all Copilot Language Server tools at once
• •Interactive confirmation on destructive operations (Code and Plan modes)

Attachments & mentions

Multi-file auto-iteration

Describe a complete project ("a notes app with authentication and sync") and the orchestrator generates a plan, presents it for review, and executes the tasks in order with resumable checkpoints.

• •Structured plan: architecture, tech stack, modules, task list with dependencies
• •Task states (Backlog → Todo → In progress → Review → Done) with real-time visualization
• •Specialized workers per task type (documents, mapping, refactor, optimization, audit)
• •Persistent checkpoints: if the app closes, the session resumes where it left off
• •Automatic lint + syntax validation of each file before continuing
• •~40 settings: separate providers for planning/coding/review, parallelism, tests, docs, Git auto-commit

Automated workflows

Define automations triggered by IDE events: after saving a file, on commit, when a test fails, on cron, etc. Each workflow can chain tool invocations or LLM calls.

Skills & custom agents

Quick access to pre-packaged prompts from the / button on the input. Ten predefined skills cover the most common cases:

Create your own agents in .pocketcode/agents/*.md with frontmatter (name, description, model, tools, systemPrompt). Compatible with the Claude Code convention and MCP servers.

Project memory

Persistent information the model reads every turn from .pocketcode/memory/. Four types:

INSTRUCTIONS.md, AGENTS.md, CLAUDE.md or .cursorrules files are also supported — they are automatically injected into the system prompt.

Conversation management

• •Unified sheet: local conversations alongside Claude CLI, Codex CLI and Gemini CLI sessions
• •Pin, rename, deep search (also searches inside message bodies)
• •Auto-compression of context when reaching 85% of the model window
• •Export to Markdown or JSON preserving code blocks and tool calls
• •Stop button during streaming: cancels response and leaves it truncated in history
• •Visible quota: rate-limit windows with countdown (Claude 5h/7d, Copilot monthly, etc.)

Free / Premium comparison