Website | Documentation | Changelog
Sediment is an open-source, self-hosted evidence store for coding agents.
You can't tell which agent output developers kept, which commits it reached, or whether it passed CI. Sediment records model calls, code changes, developer decisions, and check results on your infrastructure, and links them.
- Evaluate agent work: compare models by decisions, retained code, and CI.
- Reuse context: give agents evidence from a previous Session.
- Build training datasets: export examples for fine-tuning, preference training, and reinforcement learning.
Comparing two models on synthetic data, 30 calls each:
$ sediment report model --org acme --compare claude-sonnet-4-5 gpt-5-codex
...
compare: claude-sonnet-4-5 vs gpt-5-codex
metric prop_a prop_b diff h ci_low ci_high z p_value sig n_a n_b
ci_pass_rate 87.5% 72.2% +15.3% +0.388 -9.3% +39.8% 1.25 0.2121 no 24 18 small n, less reliable
attribution_rate 80.0% 60.0% +20.0% +0.442 -2.6% +42.6% 1.69 0.0910 no 30 30
Install the CLI from PyPI with Python 3.12, then start the local server:
pipx install sediment-cli # or: uv tool install sediment-cli
sediment serversediment server needs host libraries: libpq and openssl@3 with Homebrew,
or git ca-certificates libpq5 libxml2 libzstd1 liblz4-1 zlib1g with apt-get
on Debian and Ubuntu. The installer adds them and the CLI for you:
curl -fsSL https://sediment.so/install.sh | sh.
- Quickstart: connect an agent and verify capture.
- Integrations: Claude Code, Codex, Cursor, pi, and Copilot Chat.
- Run it for a team: EC2, your own host, then enroll your team.
- Using a coding agent? Give it
sediment guide.
Architecture and network boundaries · Captured data and privacy
Contributing · Issues · Security
Copyright (C) 2026 PAULSEN'S LLC. Sediment is licensed under
AGPL-3.0. Harness shims under shims/ use the
MIT license.