All pages
Docs · AI features
trmnly-code CLI
Availability: macOS. Ships inside the app bundle; the Linux/Windows equivalent is the portable core, not this CLI.
trmnly-code is TRMNLY's coding agent as a headless CLI — the same
engine, tool surface, approval rules, and audit log as the in-app
Agent, invoked from any shell. With no flags it runs fully
local against Ollama (qwen3-coder:30b), so it works offline and
costs nothing.
Install
Settings → AI Agent → trmnly-code CLI → Install symlinks the CLI
into /usr/local/bin (or ~/bin if that's not writable) — no sudo,
ever. From a shell, trmnly-code --install / --uninstall do the same.
Modes
trmnly-code "fix the failing test in auth_test.go" # one-shot
trmnly-code # interactive REPL
trmnly-code review 128 # review a PR
trmnly-code batch prompts.txt # one turn per prompt
Useful flags:
| Flag | Does |
|---|---|
--model / --provider / --reasoning-effort |
Pick the model. A vendor/model slug implies OpenRouter |
--continue, --resume <id> |
Pick a previous session back up |
--prompt-file <path> |
Read the prompt from a file |
--auto-approve <patterns> |
Pre-approve matching tool calls (e.g. write_file:tests/**) |
--tools-only |
No prose — run tools, print results |
--scope <dir>, --workspace <name> |
Confine the run to a subtree or a project worktree (the flag keeps its old name) |
--branch-guard |
Refuse write-shaped tools while HEAD is on main / master / prod / release/* |
--no-telemetry |
Disable usage recording for this session |
--mcp |
Print MCP server status and exit |
The REPL
Inside the interactive loop, slash commands cover the common moves:
/model (/m) and /provider to switch models, /api-key to store a
provider key, /new for a fresh session, /diff and /review for the
working tree, /allow to manage auto-approvals, /stats and /cost
for usage, /mcp for server status, /clear, /help, /exit. The
line editor has history search and tab completion.
Per-repo config
Commit a .trmnly/code.toml to pin defaults for everyone who runs the
CLI in that repo:
model = "claude-sonnet-4-6"
provider = "anthropic"
reasoning_effort = "medium"
allowed_tools = ["read_file", "grep", "bash"]
ignore_globs = ["dist/**", "node_modules/**"]
auto_approve = ["write_file:tests/**"]
scope = "packages/web"
Flags override the file; the file overrides the app's defaults.
Safety
Everything the in-app agent enforces applies here: WRITE tools prompt for approval (in the terminal), the audit log records every call, and the branch guard layers on top of the agent's existing protections. See Agent safety.
When a tool's output carries credentials, they are scrubbed before
anything is printed, and that result gets a ✂ N secrets redacted from
output line — so you know something was taken out, rather than
wondering why a token you expected isn't there. The notice belongs to the
readable transcript, so --quiet and --json leave it out.