Agent Commit
Version control for AI reasoning.
Every LLM turn — messages, tool calls, results, reasoning — gets a content-addressed, queryable commit ID. Like git, but for what your agent is thinking.
pip install agent-commit
The problem
Agents re-derive the same context over and over. Agent A figures out the project structure, then Agent B starts from scratch and does it again. With 5 agents sharing a task, that's 5x the reasoning cost for the same work. Even a single long-running session pays this tax — context window fills up, summarization kicks in, history gets lost.
Reasoning is ephemeral. Once a conversation ends, the chain of thinking is gone. You can't ask "what was Agent A's reasoning on step 3?" — it's buried in a flattened transcript with no structure.
No branching means no experimentation. Want to try a different approach? Restart the whole session or manually piece together context — no way to fork, compare, or merge reasoning paths.
How it works
Agent Commit records every agent turn as an immutable commit with a SHA-256[:16] ID. Commits form a parent chain — giving you a browsable history, instant equivalence checks, and branching/merging for free.
from agent_commit import agent_commit_tool
# Record a turn
result = agent_commit_tool(
action="commit",
messages=[{"role": "user", "content": "Build a REST API"}],
tool_calls=[{"id": "1", "name": "terminal", "arguments": {"command": "pip install fastapi"}}],
tool_results=[{"id": "1", "result": "Installed fastapi-0.115.0"}],
reasoning="Start with the scaffold, then add auth",
metadata={"model": "claude-sonnet-4", "turn": 1},
agent_id="agent-1",
task_id="build-rest-api",
)
commit_id = json.loads(result)["id"] # e.g. "a1b2c3d4e5f6g7h8"
# New agent asks: "what was the last reasoning state?"
ctx = agent_commit_tool(action="reference", commit_id=commit_id)
# Compare two paths
diff = agent_commit_tool(action="diff", commit_a_id="a1b2c3d4", commit_b_id="b2c3d4e5")
# Fork a new approach from an existing commit
branched = agent_commit_tool(
action="branch",
commit_id=commit_id,
new_task_id="build-graphql-api",
)
What you get
Reference — Inject the full reasoning state from any prior commit into a new session. No re-derivation cost.
Diff — See exactly what changed between two commits: tools added/removed, message count delta, reasoning drift.
Branch — Fork from any commit to try an alternative approach. Parent chain stays intact; both paths are explorable.
Audit log — Full timestamp, agent ID, and tool/message state for every turn. Replay or hand off to a human.
MCP server
Works with any MCP-compatible agent — Claude Code, Codex, OpenCode, Hermes, and more — via the built-in stdio adapter:
{
"mcpServers": {
"agent-commit": {
"command": "python",
"args": ["-m", "agent_commit.mcp_server"]
}
}
}
python -m agent_commit.mcp_server
No accounts, no external services. SQLite on disk by default.
Project structure
src/agent_commit/
├── core.py # CommitStore, tool handler, schema
├── mcp_server.py # MCP stdio adapter
└── __init__.py # Public exports
tests/
└── test_agent_commit.py # 24 passing tests
pyproject.toml # zero external dependencies
- Storage: SQLite at
~/.agent_commit/agent_commit.db(or$AGENT_COMMIT_HOME) - IDs: SHA-256 of canonical turn serialization, truncated to 16 chars
- Thread-safe: RLock per store, WAL mode
- Dependencies: Python standard library only
Metadata
Release files for agent-commit 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agent_commit-0.1.0.tar.gz | 12.2 kB | Without provenance | Details for agent_commit-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agent_commit-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for agent_commit-0.1.0-py3-none-any.whl |
Total release size: 21.9 kB
Release files / agent_commit-0.1.0.tar.gz
| Download URL | agent_commit-0.1.0.tar.gz |
|---|---|
| Size | 12.2 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e89a0e0d492942dab61a2d515018c3124d825b3094ba193a893cc48960ddca6d
|
|
BLAKE2b-256 checksum How to use checksums |
943e7673902cffe4b5a150a53128560d31705ae604accd190d5bdebfc9ae4dea
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.7
|
Release files / agent_commit-0.1.0-py3-none-any.whl
| Download URL | agent_commit-0.1.0-py3-none-any.whl |
|---|---|
| Size | 9.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
46e8f778844915fa6d49ff725b8ae0d38015428547e1bf277a52d29123ceaa09
|
|
BLAKE2b-256 checksum How to use checksums |
b30891482f1ddbd8dc7678c4b18c444b93102f3615841eea52a2c3811bc79e5c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.7
|
Security for agent-commit 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.