AI Agents

Run agents across projects. Review each change.

Give an API fix to Claude Code and frontend tests to Codex while you keep coding. Each independent task can have its own Git worktree. Follow approvals, inspect the changes, and decide how each branch lands.

Three full coding agents + AI Chat

Claude Code, Codex, Gemini, and GitHub Copilot.

Run the first three as full coding agents, or use GitHub Copilot in AI Chat. All four use Flame's project context and conversation controls.

Full coding agent

Claude Code

by Anthropic

Run Claude as a complete coding agent with Flame's project context, tools, permissions, and review flow.

Sign in
Claude login, API key, Bedrock, or Vertex
Setup
Installed on demand by Flame
Full coding agent

Codex

by OpenAI

Use OpenAI's coding agent for full repository work while Flame keeps the task isolated and observable.

Sign in
Your existing Codex / ChatGPT login
Setup
Installed on demand by Flame
Full coding agent

Gemini via Antigravity

by Google

Run Gemini models as a full coding agent through Google's Antigravity CLI, with multi-turn context and tool activity in Flame.

Sign in
Your Google sign-in through Antigravity
Setup
Explicit install from Google's official CLI
AI Chat provider

GitHub Copilot

by GitHub

Use GitHub Copilot in Flame AI Chat with the files, diffs, terminal output, tests, and browser context you choose.

Sign in
Your GitHub account with a Copilot plan
Setup
Connect it in AI settings
Agents Manager DashboardLive status, tokens, and tool calls for every running, scheduled, and merging task, with a concurrency limit you set.
The Agents Manager

Manage a task from assignment to review

Choose the project and runtime, monitor progress, handle decisions, and inspect the resulting branch.

Create

Start from a prompt, then choose the project, model, branch, isolation, and landing policy.

Monitor

Follow status, duration, token use, tool calls, and terminal output while each agent works.

Control

Stop, duplicate, restart, retry, or requeue a task without losing sight of its history.

Schedule

Run now, at a specific time, or after another job finishes. Chain dependent tasks into a pipeline.

Isolate

Give a task its own git worktree and branch so parallel agents stay away from your checkout and one another.

Review & land

Inspect the diff, verify the change, then merge locally, push the branch, or push and prepare a PR/MR form for you to submit. Automatic landing is a separate opt-in policy.

How a task flows

From a scoped prompt to a verified change

A worktree separates files. Set separate ports and test databases when tasks need independent runtime state.

Compose

Write a prompt, pick the project, model, branch, isolation, and how the work should land.

Isolate

Each task can run in its own git worktree and branch, so parallel agents do not need to share your checkout.

Work

The agent edits files, runs commands, and tests, with every action gated by your permissions.

Review

Inspect the diff and merge preview, ask questions, and decide how the work should land.

Land

Merge locally, push the branch, or push and prepare a PR/MR creation form for you to submit. Automatic landing requires a separate opt-in policy.

AI chat, in context

Give the agent the evidence behind the task

Use GitHub Copilot alongside your other configured models without giving up Flame's project context, permissions, or conversation controls.

  • Bring in the context that matters: files of any type, selections, diffs, terminal output, scripts, test results, screenshots, and live browser tabs.
  • Choose where the chat works: the current branch, a new branch, any other branch, or its own Git worktree. Move a running chat when plans change.
  • When the agent needs a decision, it asks: you answer with a tap, and it resumes.
  • Tool cards show every edit, command, and test as it happens.
  • Multi-project chat can grant scoped read or write access across several repositories or a whole project group.
  • Switch model and reasoning effort per conversation, mid-thread, and schedule a message for later.
Two levels of coordination

Independent jobs and delegated workers serve different purposes.

Separate implementation jobs

Use Agents Manager for independent changes across projects or branches. For example, one agent fixes API pagination while another adds UI tests. Give each its own worktree, follow its approvals, and review each diff separately.

Workers inside a Claude conversation

On a supported native Claude runtime, turn on Multi-agent mode beside Plan Mode. The lead can delegate read-only research and review while retaining edits, commands, and the final answer. Use Single mode when you want one agent on the task.

Inspect who did what

Open Agents in chat or expand its team summary in the manager. Inspect recorded assignments, activity, reports, and pending requests. Claude and Copilot native worker observation depends on the configured runtime; direct child stop, retry, and follow-up controls are not available.

Match the provider to the workflow

Claude Auto uses read-only workers. Codex and Antigravity retain ordinary agent sessions without Flame-managed native teams. Local and compatible models supply inference, with tool behavior determined by the model and endpoint. Available controls reflect runtime capabilities.

Interactive Plan Mode

Plan first. Review in code. Execute with consent.

Toggle Plan Mode in AI Chat to investigate with read-only tools and get a plan instead of a diff. Settle the decisions, review the architecture diagram, request changes, then approve before execution tools are enabled.

01. Planning

Read-only exploration

The agent investigates with read-only tools. Planning gates deny file writes and shell execution until you approve execution.

02. Review

Decisions first, with a diagram

The Plan tab opens with your decisions and approvals, each option's consequence, the agent's recommendation, an At a glance summary, and an interactive architecture diagram of the change.

03. Refine

Iterative feedback

Comment on a section, ask AI about it, or edit the Markdown. The agent revises the plan in another read-only turn, and Versions shows what changed.

04. Execution

Authoritative execution

Approve & Implement sends your plan and your decisions as the instruction, unlocks write tools, and runs in the same session context.

Hard read-only boundary:Planning and review gates deny write tools and command execution. Approval uses your edited plan. This tool policy is separate from operating-system sandboxing and from the file separation a Git worktree provides.
Bring your own model

GitHub Copilot, cloud APIs, custom gateways, or fully local

Beyond the three full agent runtimes, use GitHub Copilot in AI Chat, connect Anthropic, OpenAI, Gemini, or OpenRouter APIs, route Claude through Bedrock or Vertex, or run local models through Ollama, LM Studio, and compatible endpoints.

Claude
Available Claude models · direct API · Bedrock · Vertex
GitHub Copilot
Use your GitHub Copilot plan and available models in AI Chat
OpenAI API
Your OpenAI models
Gemini API
Google Gemini models
OpenRouter
Models exposed by OpenRouter
OllamaLocal
Models served by your configured local endpoint
LM StudioLocal
Models served by your configured local endpoint
Compatible endpoints
vLLM · LiteLLM · custom gateways
Keep useful conversation history

Bring past chats into the project they belong to.

In Settings → AI → Import and export chats, preview supported Claude Code, Codex, Copilot, Claude, ChatGPT, or Flame histories. Filter and select conversations, map them to projects or global chats, and export selected histories. Imports preserve recorded conversations, not live worker execution.

Beyond chat

AI woven through the whole editor

Use AI where you have evidence: a failed test, terminal error, source diff, browser bug, profile, or asset. Voice input can prepare an editable prompt before you send it.

Automate

Agent hooks

Run your own shell commands when a prompt is submitted, a tool runs, an agent waits for you, or validation fails. Configure them globally, per group, or per project.

Learn more about agent hooks →
Verify

Work validation

Run lint, tests, typecheck, or an AI review after every agent turn, and send failures back with Fix with AI before you commit.

Learn more about work validation →
Schedule

Send later

Queue or steer while the agent works, schedule a message for a date and time, or send it after every chat in the project is idle.

Models

Models per assistant

Choose separate models for commit messages, summaries, Improve writing, Is this safe?, and Help me decide, apart from chat.

Explore

Conversation forks

Explore an alternative from a specific message, open it in another chat tab, and navigate back to the original. Choose a worktree separately when the files need isolation.

Continue

Usage-limit continuation

Schedule a usage-limited turn for a reset or chosen time, with its prompt and context retained. Execution needs the backend running; uncertain sends pause for review.

Assets

Image and SVG edits

Use Ask AI in an asset tab for a crop, resize, color adjustment, or SVG change. Inspect the editor result, undo or save it, and export a separate copy when needed.

Voice

Local voice input

Use on-device Whisper transcription to dictate prompts and trigger supported command intents.

Autocomplete

Inline completion

Fill-in-the-middle suggestions grounded in your project's symbols and nearby code.

Predict

Next-edit prediction

Cursor-style predictions of your most likely next change: jump and accept.

Actions

Focused AI actions

Fix a failed command, review a diff, repair failing tests, resolve conflicts, or improve a prompt from the surface where the issue appears.

Recap

Work summary

A standup-ready recap of commits, branches, project activity, time, CI state, and agent work.

Permissions & safety

Choose which actions an agent can take

Every agent action is governed. You decide what's allowed, where, and whether to remember the choice.

Scopes
  • Global
  • Group
  • Project
  • This chat
Categories
  • File read / write
  • Terminal & git
  • Browser
  • IDE control
Decisions
  • Allow once
  • Deny
  • Save a rule
  • Ask every time
Guardrails
  • IDE kill-switch
  • Protected paths
  • Command safety check
  • Audit log (user vs agent)
Secret & Token GuardrailFlame warns immediately if sensitive tokens or credentials are detected in pending changes.

Put an agent on it.

Download Flame to delegate real work while keeping permissions, diffs, and landing decisions in view.

Free to use on macOS 13 (Ventura) or later, Windows, and Linux.