VT-AIAgent

System-automation agent for your machine, exposed as an MCP server for opencode / Claude Code / pi — and drivable directly via the vtai CLI.

Tools

Safety model

• Dry-run by default (guardrails.dry_run_default) — previews never touch disk.
• Path containment: writes only inside paths.writable_roots, resolved paths defeat ../symlink escapes; guardrails.deny_globs protect secrets.
• Fail-closed shell sandbox: the binary must be a bare allowlisted name with an explicit read/write policy, resolve inside trusted_bin_dirs, carry no exec-enabling flags, and every path argument passes the same checks.
• Kill switch: drop a KILL file in data_dir to abort executing tools.
• Echo-loop breaker: N identical denials in a window escalate the refusal into a CIRCUIT BREAKER instruction the looping model can't miss.
• Budget cap: budget_usd_per_run aborts an orchestrated run when accrued token cost exceeds it; spend is recorded per run.
• Audit: every allow and deny is logged to SQLite in data_dir, each invocation (orchestrated, MCP, or CLI) attributed to a run row.

Quick start

Prerequisites

1. Install

git clone https://github.com/velsrocky/vtai-agent.git
cd vtai-agent
uv sync --dev

uv sync needs uv >= 0.13.1 (the lockfile is revision = 5; the floor is declared as [tool.uv] required-version, so an older uv says so plainly instead of failing on a lockfile parse).

2. Configure

The agent reads one TOML file. Which one is resolved in this order:

1. $VT_CONFIG — explicit override
2. config/settings.toml next to the package (i.e. the clone you just made)
3. ~/.vtaiagent/settings.toml — the per-user file

A fresh clone contains the maintainer's config (their provider, their writable roots), so point the agent at your own file instead of editing the tracked one:

export VT_CONFIG="$HOME/.vtaiagent/settings.toml"
uv run vtai init          # writes a starter config there + creates the dirs

Put that export in your shell profile so every later command (and the MCP server) picks it up.

vtai init creates ~/.vtaiagent (audit DB + data dir), ~/Downloads, and ~/Documents/vt-data, and writes a starter config with safe defaults: dry_run_default = true, deny_globs for ~/.ssh, ~/.gnupg, ~/.aws and **/.env, and writable roots limited to ~/Downloads, ~/Documents/vt-data and ~/.vtaiagent/tmp. It is safe to re-run.

Check which config you ended up with — vtai info names the file and its origin:

uv run vtai info
#   config file     : /Users/you/.vtaiagent/settings.toml (user-level)

VT_LOCAL__BASE_URL=http://localhost:13305/api/v1
VT_LOCAL__MODEL=<a pulled chat model>
VT_ACTIVE_PROVIDER=local

Choosing a model

The planner/judge model comes from active_provider:

• local — any OpenAI-compatible endpoint: Ollama (http://localhost:11434/v1), llama.cpp, vLLM, LM Studio, Lemonade. No API key.
• anthropic / openai — cloud models; put ANTHROPIC_API_KEY or OPENAI_API_KEY in .env.

Set cost_per_1k_input / cost_per_1k_output in the provider block to make budget_usd_per_run meaningful; without them the agent falls back to the model's native price table.

3. Verify

Start with the preflight — it checks uv, your config, the writable roots, what is protected, the optional tools, the model endpoint, the MCP registration and any delegate waiting for review:

uv run vtai doctor

VT-AIAgent doctor — v0.6.2
  ✓  agent version      0.6.2
  ✓  uv                 0.13.1
  ✓  config file        /Users/you/.vtaiagent/settings.toml (user-level)
  ✓  config valid       no issues
  ✓  writable roots     3 present
  ✓  protected paths    4 deny globs
  ✓  shell allowlist    22 binaries
  ✓  model endpoint     http://localhost:11434/v1
  !  optional tools     not installed: ffmpeg (for transcode_media, voice helpers)
  ✓  mcp registration   opencode (global)
  ✓  pending approvals  none

It exits non-zero if something needs fixing. Then, to try the pieces:

uv run vtai info                         # config, provider, guardrails
uv run vtai tools                        # the registered tools
uv run vtai run system_info              # a read-only call, no LLM involved
uv run vtai goal --dry-run "list my Downloads"   # full loop with the model

4. Use it from your agent CLI

scripts/register-mcp.sh                  # opencode (global), claude (user), pi (CLI)

opencode's registration lands in ~/.config/opencode/opencode.jsonc and merges with your existing providers. The script registers an absolute uv path on purpose: a bare uv fails with NotFound: ChildProcess.spawn (uv ...) when the client spawns the server from a daemon with a minimal PATH.

Use

uv sync                                  # install
uv run vtai init                         # create/validate config, dirs, writable roots
uv run vtai info                         # show config
uv run vtai tools                        # list tools
uv run vtai run system_info              # invoke a tool
uv run vtai run run_shell --params '{"command":"ls ~/Downloads","dry_run":true}'
scripts/register-mcp.sh                  # register with opencode / claude

To get the vtai CLI outside the checkout, install it as an editable tool so it tracks your working copy:

uv tool install --editable /path/to/vtai-agent   # -> ~/.local/bin/vtai

Install into another project as a dependency with uv add git+https://github.com/velsrocky/vtai-agent@v0.6.7, or pull the image with docker build -t vtai-agent ..

Registry & distribution

PyPI: uv tool install vtai-agent (or pipx install vtai-agent) — publishes on v* tags via Trusted Publishing with build attestations (see .github/workflows/publish.yml). The <!-- mcp-name: ... --> comment above is the PyPI-package proof the official MCP registry expects; server.json carries the registry manifest (io.github.velsrocky/vtai-agent, stdio) alongside a metadata-consistency gate in the publish workflow. Only install from these sources; a similar repository name does not establish provenance.

Register with the official MCP registry after a PyPI release:

mcp-publisher auth --oauth2-github
mcp-publisher publish server.json

Certification (vtai-cert)

Safety claims are probes, not prose: vtai certify runs the executable guardrail-compliance suite (deny-glob containment, dotdot/symlink escapes, fail-closed shell, injected-argv resistance, dry-run-first, kill switch, audit exactness, echo-loop breaker, and the delegate's verify-then-trust loop with scripted-fake rollback + approval-gate probes) and emits a digest-stamped certificate; exit 0 gates the release in CI. See docs/certification.md.

HTTP transport

vtai serve --transport http --port 8765 --token $VT_HTTP_TOKEN exposes the same guarded tools over MCP StreamableHTTP at /mcp. The bearer token is required when binding anything other than localhost; the server refuses to bind externally without one.

Development

uv sync --dev
uv run pytest tests/ -v
uv run ruff check src/ tests/
uv run python scripts/generate_tool_catalog.py --check   # tool-catalog doc drift

The config file is resolved as described in Configure: $VT_CONFIG → the checkout's config/settings.toml → ~/.vtaiagent/settings.toml. Env overrides use the VT_ prefix with __ nesting (e.g. VT_GUARDRAILS__DRY_RUN_DEFAULT=false). Config is validated at startup (validate_settings) — a bad deny_globs pattern or a missing writable root fails fast, not mid-run.

Windows: the shell sandbox needs the allowlisted binaries to resolve, and a fresh Windows runner does not put Git for Windows' coreutils on PATH. Install Git for Windows and prepend its usr\bin (this is what CI does):

Add-Content $env:GITHUB_PATH "C:\Program Files\Git\usr\bin"   # CI
# locally: add "C:\Program Files\Git\usr\bin" to your PATH

Everything else (install, config, MCP registration) is the same on all three platforms.

Orchestrator

vtai goal runs a bounded ReAct-style agent loop. Each step the model either emits one fenced-JSON tool call (parsed, guarded, executed, result fed back) or a final natural-language answer. Bounds enforced every step: max_steps, kill switch, and budget_usd_per_run. A lonely dry_run: true text blob that fails to parse or misses its tool key is retried with a corrective note instead of ending the run as "success"; the run mode is authoritative — the model cannot override it — and every tool call is persisted as a Step row and audit entry. If a paid provider's spend can't be determined (no cost_per_1k_* config and no native cost), the run fails closed rather than reporting free. The system prompt describes the machine's actual OS.

We use a text protocol rather than native function-calling on purpose: the local Ollama models on this box emit tool calls as raw JSON text even when native tool schemas are provided (verified 2026-10-03 with qwen2.5-coder:7b), so the text protocol behaves identically across local and cloud providers.

Auto mode

auto_mode runs the loop you describe: it dispatches the goal to a coding CLI (opencode by default), reviews the reply against the goal with the configured model, and if the reply is insufficient, sends a follow-up prompt quoting the critique — up to max_iterations. The judge only accepts a bare OK as satisfied (hedged replies fail). Spend is metered against guardrails.budget_usd_per_run and attributed in the run's spend detail: the judge's own LLM calls always; the delegate CLI when it can report it — claude -p --output-format json exposes total_cost_usd, and an unreadable report from a cost-capable CLI fails closed. opencode/pi report no headless spend (documented vendor behavior, not guess-parsed event streams), so their runs carry "cli_usd": "unmeasured" and stay bounded only by max_iterations/timeout. From inside opencode or Claude Code, start the prompt with @vt-ai_automode_on: the MCP server's instructions route that to the tool.

uv run vtai run auto_mode --params '{"goal":"add a unit test for parser.py","working_dir":".","max_iterations":3}'

Delegate coding

delegate_coding snapshots the target repo, runs the chosen CLI headlessly, captures the diff, and on verify_command failure (or timeout / non-zero exit) rolls back. Three guards worth knowing:

• Human approval gate: require_approval defaults to true. Interactively it shows git diff --stat and asks before keeping changes. Non-interactive runs (MCP, scripted, orchestrated) keep the changes in the worktree as a pending approval — the run reports failure so automated loops stop — and vtai approve <id> (keep) or vtai approve <id> --rollback resolves it. While a pending approval exists for a repo, delegate_coding refuses new work there, even with require_approval: false. Pass require_approval: false only for tasks a human has already reviewed.
• pi needs credentials: with no authenticated provider, pi -p exits 0 having emitted a tool-call JSON block and applied nothing. The tool now fails closed with a clear message. Run pi auth check — or prefer opencode / claude — before delegating to pi.

Claude delegations run with --output-format json, so the reply text comes from result and the invocation's total_cost_usd is captured into the tool detail as cli_cost_usd (and audited). Delegations to opencode/pi report "cli_cost_usd": "unmeasured".

uv run vtai approve          # list pending delegate approvals
uv run vtai approve 7        # keep the pending diff
uv run vtai approve 7 --rollback

vtai info shows how many approvals are waiting, so a delegate that was kept for review does not get lost.

Inspect applications (REA capability)

inspect_application integrates REA so the agent can statically explain how a local app works without source code. Because REA's inspection is powerful, it runs entirely inside vtai-agent's guards and is off by default:

[capabilities.rea]
enabled        = true             # enable only after install vetting
binary         = "rea"
command_prefix = ["analyze-javascript-application"]
timeout        = 1800.0
keep_copy      = false

Safety contract, in order: the target must pass deny_globs; it is then copied into a quarantine dir under data_dir (denied entries and out-of-policy symlinks are skipped on the way in, so REA never reads a path you couldn't); REA runs on the copy with a minimal child environment (no inherited API keys) and is killed after timeout; output JSON is parsed and a sha256 evidence digest plus REA's evidence id are written to the audit trail; the optional judge pass (same model and fail-closed policy as auto_mode) verifies the evidence answers the goal before the result reports success. The original target is never touched, and the quarantine copy is deleted unless keep_copy is true. Live-validated 2026-10-21 against the real rea 6.7.0 CLI on a real JS app — spawn, --json envelope, and evidence digits all confirmed; see docs/showcases/rea-live-validation.md.

uv run vtai run inspect_application \
  --params '{"goal":"explain how Notes builds its CSV export header",
             "target":"~/quarantine/Notes-app","dry_run":true}'

Install REA only from official sources (github.com/morluto/rea or the rea-agents package on npm — verify provenance), not mirror sites.

Audit log

Every allow/deny is stored in ~/.vtaiagent/vtai.db. Query it without a SQL client:

uv run vtai audit --limit 20
uv run vtai audit --denied --grep ssh
uv run vtai audit --action delegate --run-id 193

To see one run end to end — the goal, every tool call, what was allowed or refused, and the spend:

uv run vtai runs            # list runs
uv run vtai runs --summary  # aggregate: spend percentiles, tool counts, denial codes (JSON with --json)
uv run vtai show 193        # one run in full

Voice control (agent-listen / agent-speak)

Both helpers talk to a local Lemonade Server (TTS kokoro-v1, STT Whisper-Tiny on Linux/macOS — Windows defaults to Moonshine-Tiny-Streaming because the whisper.cpp recipe misbehaves on some AMD GPUs) at http://localhost:13305. LEMONADE_STT_MODEL overrides the STT model on any platform. Use --seconds N to change the record window, --vad for silence-based early stop, LEMONADE_BASE_URL to point elsewhere.

The scripts are versioned in scripts/; install them with:

uv run vtai voice-sync

voice-sync copies both helpers to ~/.local/bin (override with --dest), marks them executable, and reports installed / updated / unchanged per file. vtai voice-sync --check only reports drift (exit 1 when out of sync) — re-run without --check after pulling script updates. Set VTAI_SCRIPTS_DIR to sync from a different checkout.

Cross-platform notes (verified on macOS 27/ARM, 2026-10-12):

Supported: Linux + macOS + Windows — all three run in CI (Linux and macOS push-blocked; Windows reinstated in 0.6.0 after the suite became portable, with Git Bash coreutils prepended to PATH so the allowlisted sandbox resolves real binaries). rsync-dependent tests skip on runners without rsync and are complemented by a fail-closed counter-test.

• arecord is used for recording on Linux; elsewhere the helpers fall back to ffmpeg (-f pulse / avfoundation), so macOS and Windows work too.
• --vad currently requires Linux arecord; on other platforms use --seconds N.
• macOS mic selection: avfoundation device indexes vary per machine — list them with ffmpeg -f avfoundation -list_devices true -i "" and override via AGENT_LISTEN_FF_INPUT (default audio-only -f avfoundation -i ":0"). Machines with no built-in microphone (many mini PCs) need an external/USB device attached, otherwise recording yields silence (exit 6, no speech).
• Playback tries ffplay → afplay (macOS) → mpg123 → aplay in order.
• The ALSA mic-level fix and vtai-audio-fix systemd unit below are Linux-specific; Windows and macOS handle this in System Settings.
• guardrails.trusted_bin_dirs defaults per OS (C:\Windows\System32, /usr/bin, /bin, /usr/local/bin, /opt/homebrew/bin, and the Linux set); entries missing on the current machine are ignored (validation only fails when none exist), so one config can list both Linux and macOS paths.
• system_info is portable: memory via sysctl/vm_stat and disks via /, $HOME, /Volumes/* on macOS (Metal GPU note; no rocm-smi there).
• transcode_media defaults to libx264 off Linux (VAAPI is Linux-only); an explicit "codec" still overrides.
• disk_audit also measures ~/Library/Caches, ~/Library/Logs, /Library/Caches, and ~/.Trash on macOS; fresh mac installs get vm_stat (instead of Linux-only free/nproc) in the starter allowlist.

Troubleshooting

Setup and MCP

• opencode mcp list shows ✗ vtai-agent failed: NotFound: ChildProcess.spawn (uv ...): the client spawned the server with a minimal PATH (daemons give PATH=/usr/bin:/bin:...), so a bare uv was unresolvable. Re-run scripts/register-mcp.sh — it registers the absolute uv path — or edit the mcp.servers.vtai-agent.command array in your opencode config to the full path of uv (command -v uv).
• opencode mcp list says No MCP servers configured: the registration is per-project. Run it from the checkout, or use the global registration scripts/register-mcp.sh writes to ~/.config/opencode/opencode.jsonc.
• "config exists: …/config/settings.toml" and your settings are ignored: the checkout's config wins over ~/.vtaiagent/settings.toml. Set VT_CONFIG="$HOME/.vtaiagent/settings.toml" and re-run vtai init; vtai info prints the file in use and where it came from.
• uv sync fails with Required uv version ... does not match: install uv >= 0.13.1 (the lockfile is revision 5).

Voice (Linux / PipeWire / ALSA)

• Mic produces silence or clipped garbage transcripts: the internal mic boost is pinned too high and clips the input. Fix (also applied at every login via ~/.config/systemd/user/vtai-audio-fix.service):

  amixer -c 2 sset 'Internal Mic Boost' 0
  amixer -c 2 sset 'Capture' 75%
  

• Headphone jack not detected: the kernel reports no jack on the combo port, so the headphone amp stays off. Check pw-cli enum-params 53 EnumRoute — the analog-output-headphones route shows available: no until a headset is firmly plugged in. Route manually with wpctl set-route 38 1 if detection is flaky.

• Default source is the wrong mic: list routes with pw-cli enum-params 53 Route and select with wpctl set-route 47 <index> (0 = internal mic, 1 = headset mic).

• STT returns [BLANK_AUDIO] or noise transcripts: verify levels with arecord -d 3 -f cd /tmp/t.wav and inspect the peak; near 0 means the selected route carries no signal, full-scale means clipping (lower the boost).

Delegate coding

• Files changed via delegate_coding are still sitting in the repo: that is a pending approval — the changes are kept for your review until you run vtai approve <id> to keep them or --rollback to discard them. Pass require_approval: false only for tasks a human has already reviewed.

Contributing

See CONTRIBUTING.md — setup, the guardrail rules any change has to respect, and the test/lint commands.

Security

Guardrails, what counts as a vulnerability, and how to report one privately: SECURITY.md.

License

MIT — see LICENSE. Use it, fork it, run it on your own machine.