The interactive terminal
Opens in the current directory. Approvals appear in the transcript where the call is; many sessions, one screen.
$ cd my-project && qqExplore the TUIv0.1.6Open source. Written in Rust.AI coding agents
AI coding agents
in one binary.
A terminal UI, a headless runner, and a local server, sharing one runtime. Every action passes a policy you can read, and every event is kept, so sessions survive restarts and many agents can run at once.
qqmy-project › add dry-runanthropic/claude-sonnet-512% ctx$0.04
▌YOU
▌Add a --dry-run flag to the CLI and a test for it.
●Readsrc/cli.rs212 lines0.0s
●Search"fn parse_args"3 hits0.0s
●Runcargo test -p cliexit 04.1s
QQ
I’ll add the flag next to
--verbose and gate the write behind it.◇Editsrc/cli.rsawaiting approval
◇ approval needed
src/cli.rs
@@ -18,3 +18,4 @@
18 #[arg(long)]
19 verbose: bool,
20+ #[arg(long)]
21+ dry_run: bool,
y oncea sessionw workspacen deny
preparing tool call? help^K commands^O detail
✎Answer the approval above
Your agent works. You stay in control.
FROM YOUR TERMINAL TO YOUR PIPELINE
Three surfaces, one runtime.
Work interactively. Run unattended. Share a local server.
qq run
The headless runner
The same agent in a script or CI job. Read-only unless you say otherwise; JSONL out, meaningful exit codes.
$ qq run --approval auto --format jsonlExit codes: 0 1 2 3 4 5 130
"Add a --dry-run flag and a test"
qq serve
The shared local server
Several clients, one runtime, over HTTP and SSE. Random loopback port by default.
$ qq serve --bind 127.0.0.1:4711Run a server
Same tools. Same policy. Same SQLite store. Everywhere.
A POLICY, NOT A LEAP OF FAITH
Approval you can read.
A real bash parser grades every shell command: allow, prompt, or forbidden. A held call shows you what will run, where, and why it asked; answer once, or grant it for the session or the workspace.
Forbidden means forbidden. No mode, grant, or delegate overrides it.
Understand the policy| verdict | example | auto (TUI default) |
|---|---|---|
| ● allow | cargo test | runs |
| ◇ prompt | npm install x | asks |
| ✕ forbidden | sudo … · curl … | sh | refused |
KEEP THE WORK IN VIEW
Many agents, one screen.
The sidebar groups sessions by what you should do — NEEDS YOU, WORKING, IDLE, DONE — with unread counts, not by when you opened them. Ctrl+G jumps to the next one that needs you.
Alt+A / Alt+D approve or deny another session's oldest held call without leaving yours. Sub-agents appear under their parent, bounded by the roster and depth you configure.
Work with sessionsqqmy-project › plan the releaseanthropic/claude-sonnet-5$0.11
▌YOU
▌Plan the 0.2 release and split the work.
QQ
Checklist drafted: twelve items, three need code.
Starting two helpers: one writes the regression tests, one surveys the callers of the changed API.
generating? help^K commands^O detail
›Ask QQ...
YOUR MODEL. YOUR CREDENTIALS.
Bring your own model.
Or any OpenAI-, Anthropic-, or Gemini-compatible endpoint: a gateway, LiteLLM, or a local model server.
Credentials live in your OS keyring, stored once with qq auth login. Config files hold names, never secrets; a repository’s config cannot reach a credential you did not register.
Layer your configuration: packs → global → trusted project → QQ_CONFIG and the environment.