ant — Antaḥkaraṇa governed-AI operator CLI

Author — Deepak Soni. A tiny, dependency-free, cross-platform CLI (Windows / macOS / Linux) that prompts your on-box copilot (the Antaḥkaraṇa model + control ring) from any terminal — Claude-Code style. It investigates a machine, reasons, runs read-only commands immediately, and asks y/N before any state-changing action; catastrophic commands (rm/dd/mkfs/reboot/…) are hard-blocked server-side.

PyPI: https://pypi.org/project/antahkarana-cli/ · command: ant

Install

pip install antahkarana-cli          # or: pipx install antahkarana-cli

First run

ant connect http://YOUR-BOX:8090     # point it at your copilot (saved to ~/.ant/config.json)
ant login dk-xxxxxxxx                 # the copilot requires an API key (dk- prefix); ant refuses without one
ant                                   # interactive session: banner + prompt
ant "what is listening on this server?"   # one-shot
ant status                            # connection + model health

Commands

ant                         interactive REPL
ant "<prompt>"              one-shot
ant connect <url>           save the copilot endpoint
ant login <dk-key>          save an auth token (sent as Bearer)
ant logout                  remove the saved key
ant status                  connection + model health
ant new | ant forget        archive / clear conversation memory
ant audit verify <file>     verify a hash-chained audit bundle LOCALLY (offline governance)
ant version

In the REPL, slash commands: /connect /login /logout /model /status /save [path] /new /sessions /resume <id> /forget /audit verify <file> /version /help /exit.

Governance: verify an audit bundle offline

A governed copilot (antahkarana 4.0 platform) can export a signed, hash-chained audit bundle. ant audit verify <bundle.json> recomputes the whole chain on your machine — no trust in the box that produced it, no dependencies — and if a record was altered it tells you the exact index where the chain breaks (exit code 1, so it drops straight into CI):

ant audit verify audit-bundle.json
#  ✓ audit chain intact — 128 record(s), hashes verify end-to-end
#  ✗ audit chain BROKEN at record index 41   (tamper localized)

Governance: run the non-bypass proof offline

A governed copilot (antahkarana 5.0 self-governing) ships a machine-checked guarantee that no action reaches a sink without a capability token minted by the control ring. ant verify runs that proof on your machine — a zero-dependency, stdlib-only copy of the checker, so you don't trust the box's word for it. It exhaustively checks the non-bypass invariants over the whole governance skeleton; --teeth also fails three deliberately-broken models (a checker that can't fail proves nothing):

ant verify --teeth
#  ✓ INV-1/4/5 hold over all 291 reachable states
#  checker teeth (each broken model MUST fail):
#    INV-1 bypass          INV-1 violated - counterexample in 4 step(s)
#    INV-4 ignore-control  INV-4 violated - counterexample in 8 step(s)
#    INV-5 token-reuse     INV-5 violated - counterexample in 5 step(s)

This is behaviorally identical to the SDK's atk verify and antahkarana.formal.check() — same 291 states, same counterexamples — just with no dependency on the SDK.

What you can ask

• what OS is this and what's listening on its ports? → runs read-only, reports
• create a file /tmp/demo.txt → proposes it, asks [y/N], runs only if you approve
• rm -rf /tmp/x → hard-blocked by the control ring
• assess example.com → passive web-recon → prioritized report
• summarize this file /path/to/notes.md → reads a LOCAL file on your machine
• write the summary to a file ~/Desktop → saves LOCALLY as plain text, prints the path

Features (0.6.0)

• Offline non-bypass proof — ant verify [--teeth] runs the machine-checked non-bypass invariant locally (zero-dependency, stdlib only); exhaustive over 291 reachable states, and --teeth fails three broken models. Behaviorally identical to the SDK's atk verify.
• Offline audit verification — ant audit verify checks a hash-chained audit bundle locally, zero-dependency, and localizes any tamper to the exact record index (matches the box's own verifier).
• Multi-tenant aware status — shows tenant/tier and governed-fleet size when the box reports them.
• API-key auth — refuses with "API key required" unless a dk- key is set.
• Live progress — a "thinking…/running " spinner while the agent works (TTY only).
• Memory across runs — your conversation persists in ~/.ant/session.json, so ant remembers what you did; /new archives and starts fresh, /sessions + /resume revisit past ones, /forget clears.
• Local file read — paths you mention are read on YOUR machine and inlined for the copilot.
• Local file write — when you ask it to save/write, the reply is written locally as clean text (not markdown); give a folder/path or it defaults to ./ant-output-<time>.txt, and the path is printed.

How it works

ant is a thin client over the copilot's /api/chat — the model (e.g. Qwen2.5-14B), the governed run_shell, the web-recon tools, and the control ring all live on the box (or a remote SSH target the copilot is pointed at). The box governs and audits every command; local file read/write happens on your own machine by the CLI.

Requires a running Antaḥkaraṇa copilot (see ../copilot/).