Review with context
Know which lines an agent wrote — and with which model — before you approve the pull request.
Local-first AI code provenance
Git shows what changed. Tellur shows how AI participated.
Line-level attribution for AI-assisted code: which agent, which model, which session — sealed in a tamper-evident log that never leaves your machine unless you say so.
One command, one machine-wide setup. Free and open source — Apache-2.0 core.
Captures from Claude CodeCodexCursorVS CodeJetBrainsGemini CLIWindsurfAntigravityAiderCopilotDevin
Why teams record provenance
Three jobs, one quiet recorder.
See the whole team
AI share, sessions, and review gaps across every repository, without asking anyone to change how they work.
Prove it later
Tamper-evident evidence with SLSA and SPDX exports, ready for the audit you'll get next year.
01 · Install
Running before your coffee cools.
One command installs a checksum-verified CLI, deploys the editor packages it detects, and walks through the setup wizard. There is no per-repository step — repositories activate on first agent use.
zsh — ~/work
macOS / Linux
curl --proto '=https' --tlsv1.2 -fsSL https://github.com/sydneyvb-nl/tellur/releases/latest/download/install.sh | bash Windows
irm https://github.com/sydneyvb-nl/tellur/releases/latest/download/install.ps1 | iex 02 · Capture
From keystroke to evidence.
Agents report through their own lifecycle hooks; editors capture saves. Tellur attributes the resulting lines — and hovering any line answers “who wrote this, and with what?”
src/auth.rs — tellur-demo
Explorer
src/lib.rs
src/auth.rs
src/limiter.rs
Cargo.toml
Agent session — Claude Code
You
Add rate limiting to the login endpoint.
Claude Code · claude-sonnet-5
I'll check the limiter before the user lookup and return an error
when attempts are exceeded.
Claude Code
Done — 4 lines added to
src/auth.rs.
Tellur: capturing saved · event sealed #1285 Rust · UTF-8
Step 1You ask, the agent starts a session.
Step 2The agent writes; hooks report every event.
Step 3On save, lines get a source. Hover to inspect.
Step 4Your edits stay yours — attribution follows.
03 · Provenance
Evidence nobody can quietly edit.
Every event is appended to a JSONL log and sealed into a SHA-256 hash
chain. Try it yourself: tamper with one event and watch
tellur verify catch it.
session.start
agent=claude-code model=claude-sonnet-5
prev genesis hash a3f19c
prompt.submit
prompt=sha256:9be1f0… (content not stored)
prev a3f19c hash 5d20e7
file.write
src/auth.rs +4 lines
prev 5d20e7 hash c471b2
policy.check
sensitive-paths: pass · review: required
prev c471b2 hash 08fe93
session.end
events=14 duration=6m12s
prev 08fe93 hash e6a4d1
04 · Team Hub
One control plane for every repo.
The optional self-hosted hub aggregates events and attribution during
git push. Developers keep working locally; leads and
auditors get the organization-wide picture.
hub.acme.dev/app/orgs/platform — Tellur Team Hub
acacme platform
Overview Repositories Sessions Policy Audit AI-assisted lines 0%
Sessions this week 0
Review coverage 0%
AI share by repository
Recent activity
claude-codemira pushed 3 sessions · payments-api
codexjon pushed 1 session · web-frontend
cursorana pushed 2 sessions · infra
review gapPR #412 — 61 AI lines, no human review yet
exportSLSA provenance bundle ready · job #42
Step 1Attribution rolls up from every repo.
Step 2Sessions stream in on git push.
Step 3Review gaps surface before they ship.
Step 4Exports and audit history stay on tap.
SSO, SCIM & RBAC
OIDC login, SCIM-provisioned users and groups, org and per-repo roles.
Central policy
Distribute sensitive-path and review rules; developers pull them automatically.
Tamper-evident audit
The hub keeps its own hash-chained audit history of every change.
Audit-ready exports
SLSA v1.0 and SPDX 2.3 exports run as durable jobs, per repo or per org.
05 · In practice
What teams use it for.
PR risk reports
tellur pr-report combines changed-line attribution, model evidence, policy findings, and review coverage.
AI policy audits
Sensitive-path rules, review requirements, and central policy distribution — checked where the code is written.
Incident forensics
Who wrote this line, in which session, from which prompt? Answer in seconds, months later.
Adoption insight
Which agents and models your team actually uses, and how much of the diff they produce.
Supply-chain evidence
Export SLSA v1.0 provenance, SPDX 2.3 AI annotations, and Git-notes authorship your tooling can consume.
Designed not to get in the way
Trust is the product.
Your code stays put
Capture and inspection work entirely offline. No account, no telemetry, no cloud dependency.
Privacy by default
Prompt content is hashed, not stored. Secret-looking values are redacted before anything is persisted.
Unknown stays unknown
Missing evidence is never converted into “human-written”. Confidence and source are always recorded.
Opt out per repo
One file stops capture, note publication, and hub sync for a repository. No questions asked.
Get started
Ship AI-assisted code with receipts.
Install once, keep working the way you already do. Provenance accrues in the background — until the day you need it.