CodeGPT

Inside the editor the agent searches the codebase by meaning, reads and edits files, creates and deletes them, views the folder structure, runs terminal commands for builds, tests and installs, searches the web and keeps a to do list.

CodeGPT acts as an MCP host: any local or remote MCP server added to its config file, such as Playwright, a database or an internal API, lists its tools in settings, and each tool can be switched on for the agent. Code Graphs connect GitHub, GitLab including GitLab on premises, Bitbucket and AWS repositories, and Studio agents connect to Slack, Discord and Microsoft Teams. Tools work only with models that support function calling.

Sourcedocs.codegpt.co/docs/tutorial-features/toolsread 2026-10-18

Agent mode explores the project, plans the work and then carries it out: complex tasks are broken into logical steps shown as a live to do list, and the agent edits across multiple files, creates new files, components, tests and documentation, runs commands and iterates until the task is done. Work can be paused, reviewed and resumed partway.

Three profiles set the scope of each conversation, from chat with no tools, to search with read only tools, to agent with every tool, and the profile can change during a conversation. Each account runs one interactive session at a time, and nothing hands work to parallel agents.

Sourcecodegpt.co/developersread 2026-10-18

Codebase indexing builds a semantic index of a chosen folder and gives the agent a semantic search tool, and Code Graphs turn a connected repository into a knowledge graph of nodes and relationships the agent can query, one graph on Free and Professional and more on Teams.

Context can also be added by hand from selected code, project files or a Studio agent trained on API documentation, PDFs and repositories, and Marketplace agents bring documentation for popular libraries, frameworks and languages into a conversation. GitHub sync for agent training reads public repositories up to 100 MB, text files only.

Sourcedocs.codegpt.co/docs/tutorial-basics/configurationread 2026-10-18

The agent reads the codebase first and proposes a plan a person can edit before any code is written, and every step needs approval before anything changes. Each tool can be set to enabled, where the person decides when it runs, or to auto approve, and after 20 automatic approvals of risky operations such as editing or deleting files or running terminal commands, manual approval comes back. Search mode limits a conversation to read only tools, and the terminal waits for a yes before running a command that changes anything.

Sourcedocs.codegpt.co/docs/tutorial-features/toolsread 2026-10-18

CodeGPT reports SOC 2 certification, while its self hosted offer is pitched as SOC 2 ready governance. Organizations have an owner plus admin, moderator, member and billing manager roles, and Teams adds pooled credits, central billing and an admin dashboard with usage stats.

With bring your own keys, prompts and code go straight to the chosen provider under the customer's account, and local models keep everything on the machine. Single sign on and enterprise security come through sales, and audit logs, SCIM or controls over which models and tools a team may use are not offered on self serve plans.

Sourcecodegpt.co/about-usread 2026-10-18

As the agent works, a person can watch it read files, weigh options and propose edits, the to do list shows which step is running, and context use is tracked live with no limits. Teams admins see usage stats for the pooled credit balance in a dashboard, and conversations stay in history for the period the plan allows. There is no exportable log of the agent's tool calls and commands, and no tracing or monitoring hookup for agent runs.

Sourcecodegpt.coread 2026-10-18

Project rules hold the standards the agent follows in every conversation, and conversations are kept for a period that depends on the plan, though chats with Marketplace agents are not saved by default. Studio agents keep the documents and repositories they were trained on until someone removes them. The agent does not write its own notes about past decisions or carry preferences from one conversation to the next beyond the rules, the indexed codebase and the chat history.

Sourcecodegpt.coread 2026-10-18

The extension runs in the developer's own editor, and models can run entirely offline through Ollama, LM Studio and other local runtimes so no code leaves the machine. Bring your own keys routes calls to Amazon Bedrock, Azure OpenAI or Google Vertex AI inside the customer's cloud to meet data residency rules, and for enterprises CodeGPT supports self hosted deployments in the customer's own cloud or on premises, with GitLab on premises connectable for Code Graphs. Studio, plans and credits run as a hosted service.

Sourcecodegpt.co/developersread 2026-10-18

The Studio Marketplace offers ready agents loaded with knowledge of popular libraries, frameworks, programming languages and AI APIs, usable as context in the editor, and new Studio agents start from templates with prewritten instructions. Agents built by users cover Django REST APIs, React with TypeScript, AWS infrastructure, Node APIs, machine learning pipelines and full stack work. These are knowledge assistants rather than packaged agents that take actions, and there are no ready made workflows or rule sets for the coding agent.

Sourcedevelopers.codegpt.co/explore-marketplaceread 2026-10-18

The coding agent starts from a prompt or a slash command in the editor panel of VS Code, Cursor, JetBrains IDEs or Android Studio, on code a person has selected or on the whole indexed project. Studio agents, which answer from their own knowledge, can be connected to Slack (one agent per workspace), Discord, Microsoft Teams through Power Automate or a website, and a Git commit reviewer in Studio summarizes a team's commits on every push, or only on commits tagged /codegpt, and sends the report to email, Discord or Slack. Nothing schedules or wakes the coding agent itself.

Sourcedevelopers.codegpt.co/cicd-git-commit-reviewerread 2026-10-18

Models come from more than 15 providers on the customer's own keys, among them OpenAI, Anthropic, Google AI Studio, Mistral, Cohere, DeepSeek, xAI Grok, Groq, Cerebras, Perplexity, Nvidia, Fireworks, Hugging Face, OpenRouter, Amazon Bedrock and Azure, with GitHub Copilot, Claude Code and Gemini CLI as providers and local models through Ollama, LM Studio or Docker. CodeGPT credits reach premium Claude, GPT and Gemini models without keys, Economy models carry the free and unlimited tiers, and the model can be swapped per task while the code stays where it is.

Sourcedocs.codegpt.co/docs/category/-ai-providersread 2026-10-18

The CodeGPT Plus API is a REST interface with Python and JavaScript SDKs for putting CodeGPT Studio agents, with the knowledge and code graphs they were trained on, inside other applications and calling them from code.

It is sold on an API plan at $20 per 1,000 interactions with a $50 monthly minimum plus token use, access is set up by email, and each account has its own API key. The interface reaches the Studio side of the platform.

The coding agent inside VS Code, Cursor or JetBrains has no API of its own, so a CI job or an outside tool cannot hand it a task, and it extends outward only as an MCP host.

Sourcedocs.codegpt.co/docs/codegpt_plus_apiread 2026-10-18

Slash commands generate unit tests for selected code and fix, debug, refactor, explain or document it, and a find problems command looks for issues in the code at hand. In agent mode the agent runs tests and builds in the terminal, reads the output and iterates until the task is complete. There is no suite for checking the quality of an agent or a rule set, and no comparison of results across models apart from switching the model and trying again.

Sourcedocs.codegpt.co/docs/tutorial-features/unit_testingread 2026-10-18

The agent works through files, the terminal and web search, none of which operate a browser or a desktop. A browser can be reached only by adding an outside MCP server such as Playwright to the configuration, which CodeGPT hosts but neither ships nor governs. Vision models can read screenshots a person attaches, but no agent opens a browser, clicks through a page or controls other software on its own.

Sourcedocs.codegpt.co/docs/tutorial-features/mcpread 2026-10-18