Causari — Trace intent. Debug causality.
Skip to content
Star
v0.1.0 — initial public release
Trace intent.<br>Debug causality.
The first content-addressable ledger for AI agents. Every prompt,<br>model, read, write and reasoning becomes<br>an immutable causal event — captured at the wire and on disk,<br>without the agent's cooperation , queryable like git,<br>but for the intent behind every line of code.
Quick start
View on GitHub
BSL 1.1 → Apache 2.0<br>Linux · macOS · Windows<br>~3 MB binary<br>No telemetry
re — causari
causari report
0%
AI Waste
0%<br>survival
$0<br>wasted spend
claude-3.5
27%
gpt-4o
56%
cursor
13%
A real re churn session — how much of your AI spend survived.<br>R to replay
The problem
Your agents are writing rewriting your codebase. Blind.
01
Black box edits
Cursor touched 30 files in 4 minutes. Which prompt wrote that suspicious regex on line 84? You don't know. Nobody does.
02
Silent regressions
You revert one of yesterday's agent changes. Three other agent decisions silently lose the context they were built on. CI is green. Production isn't.
03
Causal amnesia
A test breaks 200 prompts later. Bisect through chat logs? You read for an hour, give up, and rewrite the feature from scratch.
The thesis
Git tracks bytes.<br>LangSmith tracks conversations.<br>IDE checkpoints track snapshots.
None of them connect a line of code to the intent that produced it.
Causari does. We call it intent-addressable code.
The capture engine
Provenance without cooperation.
Every provenance tool before Causari only worked if the agent volunteered its own history. Agents don't. So Causari observes two independent streams — and joins them by content .
re proxy
The wire
A local, OpenAI- and Anthropic-compatible LLM proxy. Point your agent's BASE_URL at it and every prompt, completion, token and dollar flows through Causari on its way to the provider. Streaming passes through live.
re watch
The disk
A passive filesystem recorder. Every change becomes a snapshot. Then the causal join: the lines inserted in your files are searched inside the completions captured moments before. A match is a causal fingerprint — with a confidence score.
re hook
The runtime
Where the agent exposes lifecycle hooks (Claude Code today), capture is native and exact: re hook claude-code wires UserPromptSubmit and PostToolUse so every prompt and every edit is recorded at the source. No inference needed.
A real session — nothing self-reported
$ re proxy
causari: LLM capture proxy listening on http://127.0.0.1:4242
• gpt-4o 42→18 tok $0.0003 "Add JWT refresh logic that rotates every 24h"
$ re watch # another terminal
• 0d47599550 auth.py
↳ intent: "Add JWT refresh logic that rotates every 24h" gpt-4o (confidence 100%, 3/3 lines)
$ re why auth.py:2
introduced by 0d47599550
model: gpt-4o
prompt: "Add JWT refresh logic that rotates every 24h"
→ provenance is now a fact, not a self-report
The experience layer
The same mistake is never paid twice.
Recording the past is half the job. Causari distills every completed task into a signed skill — proven experience your agents recall before they act. Trust is earned, never claimed.
re skill distill
Distill
One command walks the ledger and compresses each task — the prompt that triggered it, the steps taken, the files changed — into a portable skill file. Idempotent, local, instant. Signed with the repo's Ed25519 key at the moment of creation.
● → ◆ → ★
Earn trust
● recorded — distilled, no success signal yet. ◆ verified — evidence attached: exit code 0, or the work survived at the tip of the timeline. ★ proven — verified and recalled 3+ times by agents doing new work. Edit one byte and re skill verify exposes it.
causari_recall
Recall
Through MCP, agents query their own past before acting: signed skills come first, ranked by trust (proven ×4, verified ×2). Every recall bumps the use counter — the loop that turns a verified skill into a proven one. Agents get measurably cheaper over time.
Events become experience — signed, verifiable, reusable
$ re skill distill
distill: 128 event(s) scanned, 7 new skill(s), 12 already distilled
◆ verified 2ce0c7bbda add retry with exponential backoff
$ re skill verify
ok 2ce0c7bbda add retry with exponential backoff
verify: 7 skill(s), every signature valid (Ed25519)
$ re skill show 2ce0
trust: ◆ verified signature: valid
trigger: "add retry with exponential backoff"
evidence: exit_zero=true survived=true
→ next time any agent faces this task, it already knows the answer
$ re skill pull ~/team-skills/
pull: 7 imported, 2 already present, 0 rejected
→ one engineer's verified fix becomes every agent's instinct. No server. Ed25519 only.
Causari Proof
Provenance anyone can verify. Trusting no one.
Mint a signed certificate of your repo's AI provenance — how many agent actions, which agents and models, how much verified experience — bound to the exact ledger by a content digest. Anyone can verify it offline : no server,...