v3.5.0Node 20+ · MIT · runs locally

The coding agent
that shows its work.

Sentinel is a local coding agent for the terminal. It streams from the LLM you pick, edits files through sandboxed tools, and prints what every turn cost.

Get startedRead the docs
default cost, Groq free tier
$0
sandboxed local tools
19
providers, one client
12
servers, 0 telemetry
0

What it is

Does one thing well.

Most terminal coding agents either do everything or are a hosted service. Sentinel is a fast, cheap, transparent coding chat: the whole brain is one loop in src/agent/loop.js.

  1. 01

    input

    your turn

  2. 02

    stream

    any provider

  3. 03

    tool call

    sandboxed, in-process

  4. 04

    result

    fed back

↺ repeat until done or MAX_ITERATIONS · yieldstexttool_calltool_resultfinisherror

One streaming client covers

  • OpenAI
  • Anthropic
  • Gemini
  • Groq
  • Mistral
  • DeepSeek
  • xAI
  • Together
  • Fireworks
  • OpenRouter
  • Ollama
  • LM Studio

  • No servers, no telemetry

    The agent loop runs in-process. Nothing leaves your machine except LLM API calls.

  • Cost-aware by default

    Every turn prints tokens and USD. Free models default to $0. Context auto-compacts at 40k.

  • Sandboxed tools

    Reads, edits, grep, bash and more, scoped to the project root, path traversal rejected, every write checkpointed.

  • Sessions as files

    Chats persist as plain JSON. List, switch, delete, export and compact whenever you want.

  • MCP stdio server

    Expose health, ask and review-diff to Claude Desktop, Cursor or any MCP client.

Permissions

Six modes. Fixed allowlists.

Each mode maps to a tool allowlist the model cannot talk its way out of. Switch with Ctrl+M or /mode.

Mode guide

/mode build

Actually making changes

Every tool is available. Each write still creates a checkpoint, so /undo and /redo work across turns.

File edits
allowed
Shell
allowed

Beyond chat

Built to be left running.

An engineer on an engagement states the goal, respects the budget, and leaves a runbook behind. These commands do the same, from files already on disk.

  • sentinel outcome "the sync is flaky"

    Vague ask in, judgeable contract out

    Turns a request into current state, one measurable target, the exact verification command, blast radius, rollback and unknowns, then works to it.

    CURRENT STATE  retries hit 3 before the lock clears
TARGET         0 failures across 50 sync runs
VERIFICATION   npm test -- sync · exit 0
ROLLBACK       revert checkpoint #14
UNKNOWNS       is the lock TTL configurable?

  • sentinel watch "keep the sync green"

    A presence, not a cron job

    Wakes on a failing command, a moved file or a new HEAD. Steer it from another terminal; the instruction lands on the next tick. Budget-checked before every wakeup.

    $ sentinel watch "keep the sync green" \
    -t "command:npm test" -t git
▸ tick 7 · trigger command:npm test (exit 1)
$ sentinel steer "also check the retry path"

  • blast-radius gate

    Asks once where being wrong is expensive

    Migrations, workflows, lockfiles, auth, billing, infra. The first write is refused until the agent cites the file:line that justifies it and names the rollback.

    ◆ blocked  db/migrate/0042_add_index.sql
  a migration is rarely undone by reverting it
  required: justifying file:line · exact rollback
✓ opened   for the rest of this turn

  • sentinel budget --usd 25 --deadline 2h

    Spend that outlives the process

    A ceiling persisted per project. ask, goal and outcome all honour it, and the loop stops hard at the limit rather than letting one more call land.

    active  ████████  $12.40 of $25.00 (50%) · 1h 59m left

  • sentinel handoff <runId>

    The runbook, not the diff

    Writes HANDOFF.md from the recorded trajectory: what changed, what was verified, what was tried and rejected, claims to distrust, what is still fragile. No model, no API key.

    # .sentinel/HANDOFF.md
## What was verified
## Tried and rejected
   incremental parser patch, broke on nested arrays

  • sentinel risk "npm publish"

    Permission by novelty

    Command shapes are graded per repo. git commit and git push never share a shape, and --force is never collapsed into a placeholder.

    green   git commit -m <msg>   approved shape
yellow  npm run bench          new, not destructive
red     git push --force       always asked

Watch

It shows its work.

The agent refuses a risky write, then answers with the line it read and the command that undoes it. A real session, shot as it ran.

Sixty-nine seconds, no narration over the footage. The terminal output is Sentinel's own, at its real speed.

sentinel launch film · 1920×1080 · 69s

Documentation

Read it in an afternoon.

Every page is short and task-shaped.

Install

Four commands.
Then just type.

Export a Groq key for the free default, or point it at Ollama and skip the key entirely.

Full installation guide
terminal
git clone https://github.com/KunjShah95/SENTINEL-CLI.git
cd SENTINEL-CLI
npm install
npm link

export GROQ_API_KEY=gsk_…
sentinel