taucode — the coding agent that remembers

Memory

Notes that get more useful every session.

Most coding agents forget you when the terminal closes. tau writes down what it learns, per project, and brings back only what the current turn needs.

Signup uses a token bucket, 5 req/min per IPfact

Migrations are plain SQL in db/migrationsfact

Tests run with bun testfact

gate·retrieve2 of 14 facts

Facts

Semantic memory stored per project. Before each turn a Retrieval Gate decides whether facts are needed at all, and what to search for.

today · 14:02Added rate limiting to /login, reusing the /signup middleware.

yesterday · 18:40Agreed never to edit an applied migration.

sep 26 · 10:15Set up the eval gate for CI.

Episodes

After every turn tau writes a one-sentence summary and consolidates them over time. That's how it remembers what you did last week.

name: add-endpoint
triggers: route, endpoint, api
steps:
  1. read an existing route
  2. copy its middleware stack
  3. add a test next to it

create_skill

Skills

Procedural memory: named instructions with trigger keywords, injected when your prompt matches. tau can write new ones itself.

How it works

Ask. Recall. Approve.

Type a request in your terminal. tau pulls in the memory that matters, plans with your model, and stops for a yes before it touches anything.

Any provider. Log in to OpenAI with your ChatGPT account, or paste an Anthropic API key. Swap models mid-session with /model.
A permission gate. Every write, edit and shell command asks first. Reads run freely.
A live dashboard. Reasoning, memory, tool calls, traces and cost at localhost:7777.

tau — ~/projects/my-app

taucode · dashboard → http://localhost:7777
  /model   pick a model
  /auth    switch provider

[Memory] 3 facts · 12 episodes · 2 skills

›

Design

Fully local. Less on purpose.

Everything tau does happens on your machine where you can see it. Each of these choices has an ADR in docs/adr/.

No cloud sandbox

Tools run directly on your machine, which is why every write, edit and shell command asks you first.

No hosted service

There's no account to create and no server of ours between you and the model. You bring your own provider.

No spoofed logins

Claude uses a real API key. Logging in with a Pro/Max subscription would mean impersonating another product, and that gets accounts banned.

Install

Zero to a working agent in four steps.

taucode runs on Bun. You need bun, git, and an OpenAI account or Anthropic API key.

1Install Bun

Skip this if bun --version already works.

macOS / Linux

curl -fsSL https://bun.sh/install | bash

Windows

powershell -c "irm bun.sh/install.ps1 | iex"

2Clone and install

Pull the repo and install its dependencies.

git clone https://github.com/Saikat-D023/tau-code.git

cd tau-code && bun install

3Connect a provider

OpenAI uses a browser login against your ChatGPT/Codex account. Claude takes an API key. Credentials are saved to ~/.taucode/auth.json, outside the repo.

bun run auth

4Start tau

Opens the chat loop in your terminal and the live dashboard at localhost:7777.

bun start

Reference

Commands

Everything you can type, run or configure.

`/model`	Opens an interactive picker for models on your current provider.
`/model <name>`	Sets the model directly on your current provider.
`/auth`	Logs in to or switches provider (OpenAI OAuth or Claude API key) without leaving the session.
`exit` · `quit`	Ends the session. Ctrl+D also works.
anything else	Sent to the agent as a turn. Tool calls print inline as `↳ tool {args}`.

Start a session that remembers.

Clone the repo, connect a provider, and run bun start.

Get started Read the source

`bun start`	Starts the CLI and the dashboard on :7777.
`bun run auth`	Runs the provider login flow before you start.
`bun run eval`	Runs the deterministic eval suite (bun test assertions).
`bun run eval:judge`	LLM-as-judge scoring of agent responses.
`bun run eval:gate`	Fails when eval scores drop below `TAU_EVAL_THRESHOLD`. Meant for CI.
`bun run trace`	Points you at the JSONL traces in `.tau/traces/`.

`read_file`no prompt	Reads files in your project.
`write_file`asks first	Creates or overwrites a file.
`edit_file`asks first	Makes targeted in-place edits.
`bash`asks first	Runs a shell command directly on your machine.
`search_web`	Looks things up on the web.
`manage_memory`asks first	Adds, updates or removes project Facts.
`create_skill`asks first	Saves a reusable procedure with trigger keywords.
`update_soul`asks first	Edits `.tau/SOUL.md`, the project's standing persona and preferences.

`OPENAI_API_KEY`	Optional OpenAI key, used instead of the OAuth login.
`ANTHROPIC_API_KEY`	Your Anthropic key for the Claude provider.
`TAU_GRAPH_WORKFLOWS=1`	Turns on the Triage Graph: greetings get a quick reply, and everything else runs the full loop. If the graph errors, tau falls back to the plain loop.
`TAU_DIR`	Overrides where project state (`.tau/`) lives.
`TAU_GATE_MODEL`	Model used by the retrieval gate.
`TAU_JUDGE_MODEL` · `TAU_EVAL_THRESHOLD`	Judge model and pass threshold for evals.