Build an AI agent whose entire conversation history is stored as inspectable Git commits.
Rewind an agent session to an earlier checkpoint when something goes wrong.
Fork a session into a new branch to explore an alternate line of conversation.
Manage subagents whose work is tracked and linked back to a parent session via Git.
| yashneil75/gitlord | agricidaniel/gogh | betta-tech/harness-sdd | |
|---|---|---|---|
| Stars | 46 | 46 | 46 |
| Language | Python | Python | Python |
| Setup difficulty | easy | moderate | easy |
| Complexity | 3/5 | 3/5 | 2/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
GitLord is a Python framework for building and running AI agents that stores every part of a conversation using Git itself as the storage system. Instead of keeping a session's history in a database, each turn of a conversation, whether from a user, the agent, or a subagent, is written as its own Git commit. Because of that, the entire history of an agent session is inspectable with normal Git tools, can be rewound to any earlier point, and can be forked into a new branch of possibilities the same way a codebase can. Under the hood, each session lives on its own Git branch under a dedicated refs namespace, and any subagents spawned during that session get their own nested branches. The actual content of each turn is stored as a JSON file inside the commit's tree, while metadata like the turn number, its type, which agent produced it, and token counts are attached using Git commit trailers. When a subagent finishes its work, its result gets linked back to the parent session through a trailer as well. The project is organized into separate modules that each handle one concern: low level Git operations for building commits and trees, session lifecycle management for creating and resuming conversations, subagent management for spawning and collecting results, context assembly for deduplicating and summarizing history within a token budget, an MCP server integration layer for tool discovery and calling, a model router for translating tool schemas and handling retries across different LLM providers, and an optional vector index built on ChromaDB for retrieval. A command line tool ships alongside the library, letting a user run a session, view its turn history, inspect the branch structure, show the JSON content of a specific turn by its commit hash, rewind a session back to an earlier checkpoint, or diff two turns against each other. Installation is done through pip, with optional extras for MCP server support, LLM routing through litellm, and the ChromaDB vector index. The project is released under the MIT license.
A Python framework for AI agents that stores every conversation turn as a Git commit, making sessions inspectable, rewindable, and forkable.
Mainly Python. The stack also includes Python, Git, ChromaDB.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.