Ever wished your agent conversations were as trackable as your code? Gitlord makes it real.
What it does:
- Every agent turn becomes a git commit
- Branch out subagents without breaking your main flow
- Rewind to any point in your conversation history
- Connect tools via MCP (filesystem, search, browser, etc.)
- One interface, any AI provider (OpenAI, Anthropic, local models)
- Full CLI for managing sessions and branches
Why it matters:
- Your agent history is navigable, forkable, and diffable, just like code
- Context management handles token budgets automatically
- Spawn child agents on isolated branches with their own history
- No lock-in: all components are modular and swappable
- Built-in integrations:
- MCP tools: Connect any MCP server (git, filesystem, browser, search). Tools flow to subagents automatically
- RAG: Vector search across your full agent history. ChromaDB-backed semantic queries built in
- Provider abstraction: Switch between any of the 170 providers and nearly 3,000 models, or local models with one line. Mix providers per agent
Build an agent in 4 lines:
from gitlord import Session, SessionConfig
config = SessionConfig(model="claude-opus")
session = Session.create("my-agent", config)
session.add("user", "Refactor our OAuth to the new framework")
Done. Gitlord handles the rest.
Performance improvements (v0.1.0):
- Structured trailers eliminate JSON walks: metadata parsing is now O(1)
- Auto-index updates on every turn, cached at .gitlord/index.json
- New in-memory query layer for fast turn filtering and aggregation:
- Snapshot compression for long-running sessions: compress old turns into JSON, rebase from checkpoint
Repo: https://github.com/yashneil75/gitlord
Landing page: https://yashneil75.github.io/gitlord/
Docs: https://github.com/yashneil75/gitlord/tree/master/docs
I'm curious: what's one pain point with frameworks like LangChain and CrewAI that bugs you in the middle of the night?
MIT licensed. Built for agents that ship.
Starring helps more than you think!
Top comments (0)