tret

An open-source AI harness platform for non-technical knowledge work.

tret is what a coding agent is for engineers, built instead for analysts: a workbench where AI does rigorous, auditable knowledge work — structured verdicts, evidence extraction, deliverable drafting — under rules that make its output trustworthy enough to put in front of a client, a credit officer, or an auditor.

The flagship domain pack is climate risk assessment; the core is domain-agnostic and packs are pluggable.

Why tret is different

Five rules are built into the architecture, not just the prompts:

1. The AI never invents numbers. Values enter a run only by retrieving dataset rows (lookup_dataset) or by invoking vetted deterministic methods (run_method) — pack-authored Python the agent can parameterize but never write, each execution manifest-pinned (params, code hash, output hash). Every cited value is mechanically cross-checked against what was actually retrieved or computed; anything else is a validation error, not a hallucination that slipped through.
2. Outputs are drafts until a named human approves. The approver is stamped from the login session — there is no API field to claim approval.
3. Every run is fully auditable: which model ran, why the router picked it (verbatim reasoning, candidates considered, prompt version, and which objective was in force), the doctrine version (content hash), every tool call, the token breakdown of the prompt itself, cost — and the estimated energy and carbon the run drew.
4. Doctrine-as-context. The reasoning rules live in versioned markdown files the agent must follow and cite by heading.
5. Honest uncertainty. insufficient_data is a respectable verdict; missing data becomes an explicit data request, never a guess. A run that never lands a valid verdict ends completed_without_output — a status that admits there is nothing to read, instead of one that implies there is.

Quickstart

One command on macOS or Linux — it checks for Docker, clones tret into ~/kith-tret, builds and starts it, and opens your browser:

curl -fsSL https://tret.kithailab.com/install.sh | bash

Re-run it later to update in place (it won't touch a checkout you've edited). That URL serves install.sh from this repo verbatim, so https://raw.githubusercontent.com/kith-ai-lab/tret/main/install.sh is the same script — read it before piping it to bash if you'd rather not do that blind.

Verify before running. Checking the download against tret.kithailab.com itself proves nothing — that's the same host that could serve you a bad file. install.sh.sha256 is committed in this repo instead, so fetch it from GitHub's raw content host and verify against that:

curl -fsSLO https://tret.kithailab.com/install.sh
curl -fsSLO https://raw.githubusercontent.com/kith-ai-lab/tret/main/install.sh.sha256
shasum -a 256 -c install.sh.sha256   # must print: install.sh: OK
bash install.sh

That raw.githubusercontent.com fetch only works once this repo is public — today it's private, so the URL 404s. Until then, the clone path just below is the way to actually inspect the script before running it; install.sh.sha256 still verifies once the repo goes public.

Or clone the repo and run the same script from your own checkout — git's own integrity checking stands in for the checksum here:

git clone https://github.com/kith-ai-lab/tret.git && cd tret
less install.sh          # read it
./install.sh             # then run it

Not a terminal person? Double-click start-tret.command (macOS) or start-tret.bat (Windows) and tret sets itself up — Docker check, first-run config, browser open. docs/easy-start.md is the plain-language walkthrough, stop-tret the off switch. Or skip installing anything and deploy to Render with one click.

Already have a checkout, or want to drive it yourself:

cp .env.example .env       # defaults are fine — provider keys are added in-app
docker compose up --build

On first login tret asks for a provider key (one OpenRouter key alone works; Anthropic and Moonshot too) and stores it encrypted — no config file editing. Prefer no cloud at all? Set TRET_LOCAL_BASE_URL to your own inference server and tret runs with no cloud provider — from the compose stack that is http://host.docker.internal:11434/v1 for a host Ollama, not localhost.

Open http://localhost:5180, log in ([email protected] / tret-admin by default — change in .env), and you're in a seeded demo:

• a Chat front door — just ask ("Is the vendor flood score for Alder Point still trustworthy?") and the assistant recognizes the request, triggers the right specialist harness task via its run_harness_task tool, and reports the draft verdict back — with the delegated run fully audited and its finding waiting in the approval queue. It can also fan several independent specialist tasks out at once (delegate_parallel) and brief short-lived, read-only subagents for a quick lookup (spawn_subagent). Every delegated run, however it was started, is a real Run with its own audit trail — none of this is a hidden side channel — and the turn's own cost cap covers the whole tree, not just the top-level run.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗