FlexiGPT is a local-first BYOK AI workspace for repeatable LLM workflows with model choices, Agents, Workspaces, MCP servers, native and provider-compatible tools, Skills, and private local history across multiple providers.
- Power users, developers, writers, consultants, teams, and interface builders who use LLMs for repeatable work.
- I want to be in the driver's seat when using LLMs.
- I want to use LLMs as tools while staying in control of what I do and how I do it. Most systems seem to be designed to let their tooling and workflows drive the work.
- Keeping reusable LLM setup in local catalogs and source-backed files where supported: models, prompts, Agents, MCP servers, Skills, and Workspace context.
- Assembling every request from that setup, with its exact input inspectable before it is sent. Nothing gets injected behind your back.
- Controlling execution through tool policy, approval, manual execution, automatic execution where allowed, and Run controls.
- Keeping a record of each run - input, output, and what it changed - so it can be reviewed, diffed, and repeated.
- Exposing all of it through one uniform interface, so any application can drive the same setup.
- Building use-case UIs on top of that interface, starting with chat.
- Download the latest release from GitHub Releases.
- macOS:
.pkg - Linux:
.flatpak - Windows:
.exe
- macOS:
- Install the package. Detailed installation steps are in Installation.
- Launch FlexiGPT.
- Get an API key for a hosted provider, or start a local LLM server you control.
- Open Models and add the key alongside the matching provider/model setup.
- For local endpoints that require a non-empty key, add a harmless placeholder key for the local provider.
- In Models, enable a provider/model or use Add Provider to copy the closest provider setup for your endpoint.
- Open Chats.
- Choose a built-in Agent for a reusable starting setup, or choose a model directly.
- Attach files, folders, notes, PDFs, URLs, or code when the model needs source material.
- Optionally select a Workspace for recurring repository or folder context, or an already configured MCP server for connected tools, resources, or prompts.
- Send.
Good first workflows:
- Use a home screen workflow card such as Develop a Feature, Review Code, or Investigate a Bug.
- Attach only the relevant source material.
- For code changes, start with a repo path or changed files and let the Spec Driven Development Agent inspect, scope, implement, and verify the change.
- Review the prefilled opening text and adjust it for your task before sending.
- Export, edit, and import an Agent when you need a durable variation.
FlexiGPT does not bill you directly. Usage costs and limits come from the provider account behind the key you configure.
FlexiGPT does not proxy LLM calls through a FlexiGPT-hosted service. Requests go directly to the provider or endpoint you configure.
- Built-in support for OpenAI, Anthropic, Google Gemini, xAI, Mistral, Hugging Face, OpenRouter, and local/self-hosted runtimes such as LocalAI, LM Studio,
llama.cpp, Ollama, SGLang, and vLLM. - Custom endpoints across OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google GenerateContent-style APIs.
- Curated built-in providers and models so you can start quickly without manually defining endpoints or defaults first.
- Provider and model configuration is managed in Models, where local users can copy/fork a built-in provider, adjust the endpoint, headers, compatibility type, and model defaults, then use the forked provider in Chats.
- API keys are entered in Models and stored securely through the OS keyring, not in plain-text exported settings.
- One interface for chats, tabs, reusable Agents, Workspaces, MCP servers, models, attachments, tools, Skills, search, and exports.
- Build repeatable workflows by combining Agents, Workspaces, model choices, attachments, tools, Skills, and connected services.
- Skills can seed starter drafts, carry instruction-only behavior, or manage reusable workflow state.
- Switch providers or models as you iterate.
- Multi-tab conversations with local history search and resume flows.
- Export the current conversation as JSON.
- Agents can prepare a model choice, instructions, opening text, tools, Skills, and connected services without locking the chat.
- Workspaces provide reusable repository or folder context without reattaching the same project material.
- MCP servers can contribute selected tools, resources, resource templates, prompts, and instructions from configured local or remote services.
- MCP tool results can render policy-controlled MCP Apps inside the chat.
- Tools can be attached per conversation or per message and configured for manual review or auto-execution.
- Native FlexiGPT tools are built-in Go implementations. Provider-compatible tools can also appear for the selected provider. Custom HTTP tool definitions are not supported.
- When an eligible auto-execute tool is called, FlexiGPT can run it and submit the result back to the model.
- Keep tools manual when you want tighter control over execution.
- Markdown rendering with syntax-highlighted code blocks.
- Mermaid diagram rendering with zoom and source or image export workflows.
- KaTeX math rendering.
- Citations, token usage, and per-message request/response details for inspection and debugging.
- Message-level controls for copying, inspection, and follow-up iteration.
- Local conversation storage and full-text search.
- File, folder, image, PDF, and URL attachments.
- Bundled offline docs shipped inside the app.
- Conversations, Agent Plugins, Agent files, Workspace setup, MCP server catalogs, and configuration are stored locally.
- Selected request context is sent to the provider or endpoint you choose when you send.
- Use your own provider accounts. FlexiGPT does not proxy or bill model usage.
- Develop bounded features and enhancements from local repo context with a spec, implementation steps, edits, and focused verification.
- Review code, diffs, and PRs for correctness, security, reliability, maintainability, and test gaps.
- Investigate bugs from logs, stack traces, failing outputs, source files, and config.
- Refactor code, design tests, implement tests, explore codebases, and review architecture with built-in software Agents.
- Use read-only Agents for review/investigation and manually review any built-in tool with write or other external side effects.
- Start from the home screen's Develop a Feature, Review Code, or Investigate a Bug workflow card when available.
- Inspect the loaded Agent and revise its draft, model, context, tools, Skills, and MCP selections before sending.
- Export and re-import an Agent when you need a durable variation.
- Getting Started
- Chat Workspace
- Composer Context
- Agents
- Workspaces
- MCP Servers
- Providers and Models
- Privacy, Data, and Troubleshooting
- Everyday Recipes
- Data storage: local artifact storage backed by
JSONandSQLitefiles in the local filesystem. - Go backend.
- Wails desktop application platform.
- Official Go SDKs by OpenAI, Anthropic, and Google GenAI.
- Vite and React Router v7 frontend in TypeScript.
- DaisyUI with Tailwind CSS for styling.
- Tooling: GolangCI-Lint, Knip, OxLint, Prettier, and GitHub Actions.
Developer setup is documented in devsetup.md.
Copyright (c) 2024 - Present - Pankaj Pipada
All source code in this repository, unless otherwise noted, is licensed under the Mozilla Public License, v. 2.0. See LICENSE for details.