LLM CLI Gateway

llm-cli-gateway

"Without consultation, plans are frustrated, but with many counselors they succeed." — Proverbs 15:22 (LSB)

Secure local control plane for AI coding agents.

llm-cli-gateway lets supported MCP clients operate Claude Code, Codex, Gemini/Antigravity, Grok Build, Mistral Vibe, Cognition Devin, Cursor Agent, and configured HTTP API providers through one user-owned gateway while preserving native CLI sessions, local credentials, durable async jobs, validation receipts, and review workflows.

Why developers try it: use the client you are already in to delegate work to local coding agents, scope remote execution to registered workspaces, gate risky actions, survive disconnects, and collect auditable review evidence without turning those agents into a generic chat proxy.

Current signals: CI and security workflows pass on main, OpenSSF Scorecard is published, OpenSSF Best Practices is passing, releases use Sigstore signing, and the package is MIT licensed.

Quick Start

npm install -g llm-cli-gateway

Or use directly with npx from an MCP client:

{
  "mcpServers": {
    "llm-gateway": {
      "command": "npx",
      "args": ["-y", "llm-cli-gateway"]
    }
  }
}

What It Provides Today

llm-cli-gateway is a single-user MCP control plane for operating AI coding agents from supported local or remote clients. It is more than a thin CLI wrapper:

• Runs registered provider CLIs and configured HTTP API providers through consistent sync and async MCP tools.
• Persists long-running jobs, supports restart-safe result collection, deduplication, cancellation, and sync-to-async deferral.
• Tracks sessions, real CLI resume paths, structured response metadata, and cache telemetry.
• Supports cache-aware promptParts, including explicit Claude cache_control when opted in.
• Can run supported provider requests inside gateway-managed git worktrees for isolated multi-agent review and implementation loops when using the local file-backed session manager. PostgreSQL-backed sessions reject this filesystem-local feature before creation. Grok, Devin, and Mistral require an explicit provider-native sessionId for a gateway worktree; fresh, createNewSession, and resumeLatest-only worktree requests fail closed because they cannot durably reselect it. Materialization suppresses repository, system, and global Git hooks, configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands.
• Ships personal-appliance setup surfaces: HTTP transport with bearer-token auth, doctor --json, setup UI artifacts, provider setup snippets, Docker fallback, and checked release bundles.
• Remote web connectors use MCP OAuth discovery and authorization-code setup with static client or shared-secret gates. Client secrets are generated locally, stored only as hashes, and printed only by explicit copy-once commands.
• Provider CLI requests can select registered workspaces by alias via workspace; every HTTP/tunnel request must use a registered alias, session workspace, or [workspaces].default before provider execution. Local unrestricted filesystem access is the stdio transport.

Workflow Assets

The repo ships agent-ready workflow skills under .agents/skills for async orchestration, session continuity, multi-LLM review, implement-review-fix loops, retrospective evidence walks, secure approval-gated dispatch, and Personal Agent Config Kit operations. Nine caller-facing skills are bundled in the published npm package: async-job-orchestration, multi-llm-review, session-workflow, secure-orchestration, implement-review-fix, retrospective-walk, public-demo-session, least-cost-routing, and personal-agent-config-kit. Machine-readable DAG-TOML plans live under docs/plans and setup/install-plan.dag.toml for workflows that need deterministic sequencing and verification gates.

Skill packs can be updated outside the core npm release by placing skill directories in local, operator-controlled paths. The gateway loads bundled skills first, then [skills].paths, then LLM_GATEWAY_SKILLS_PATH, then ~/.llm-cli-gateway/skills when it exists; later roots override earlier skills by name. Each skill is a directory containing SKILL.md. A root may also carry skill-pack.json to pin expected SKILL.md hashes:

[skills]
paths = ["/opt/llm-cli-gateway/skills"]

export LLM_GATEWAY_SKILLS_PATH="/opt/team-skill-pack:/opt/incident-skill-pack"

{
  "name": "team-pack",
  "version": "1.0.0",
  "skills": [
    {
      "name": "incident-retrospective",
      "sha256": "<sha256 of incident-retrospective/SKILL.md>"
    }
  ]
}

The loader is intentionally local-only: it never fetches remote Markdown at startup. To update a pack, install or replace files through your package manager or deployment system, then restart the gateway so the advertised skills://... resources refresh.

The next documentation focus is provider-specific skill and DAG-TOML pairs for each outbound CLI and API-provider family: Claude, Codex, Gemini/Antigravity, Grok, Mistral Vibe, Devin, Cursor Agent, OpenAI-compatible endpoints, Anthropic Messages, and xAI Responses. The implementation plan is tracked in docs/plans/provider-workflow-assets.dag.toml, with each provider asset expected to cover install/login checks or token-env checks, session behavior, approval modes, cache/telemetry surfaces, failure modes, and a smoke-test gate.

Trust & Supply Chain

• CI runs build, lint, format, tests, package checks, and npm audit.
• Security CI runs actionlint, zizmor, shellcheck, typos, osv-scanner, gitleaks, and lychee.
• GitHub release installer artifacts are checksummed and signed with Sigstore keyless signing.
• npm releases use a generated prod-only shrinkwrap and release security audit; GitHub Actions Trusted Publishing exchanges the job's OIDC identity for short-lived npm publish credentials.
• The npm package intentionally ships a generated, prod-only npm-shrinkwrap.json so registry installs resolve the audited release tree. Release gates regenerate it from package-lock.json, compare for parity, and run a registry-fidelity consumer install before publishing.
• Socket behavioural alerts are documented in socket.yml and under "Security Considerations" below. shellAccess and shrinkwrap are reviewed package capabilities/configuration for this CLI appliance, not hidden install behaviour.

Personal MCP Appliance

The personal-appliance contract keeps that surface intentionally narrow: one trusted user runs the gateway on a machine or volume they own, connects one MCP endpoint, and lets supported clients operate local coding agents through workspace-scoped, approval-gated, auditable requests.

The product contract is documented in docs/personal-mcp/PRODUCT_CONTRACT.md. It defines the single-user scope, security posture, target support matrix, and provider-support verification gates. Public setup guides must not claim ChatGPT, Claude web, Claude Desktop, Codex, Gemini CLI, Gemini web, or Grok inbound support until the corresponding provider/client path has been verified.

This project does not provide hosted multi-tenant credential custody. Provider credentials stay on the user's machine or user-owned deployment volume.

For a single developer who works on several workstations and repositories, the Personal Agent Config Kit guide explains the Git-synchronised personal baseline, repository overlays, immutable context stamps, and workstation-safe provider continuity model. The baseline directory is confined to a non-symlinked descendant of the local home directory, and every publish or sync revalidates its configured origin fetch and push URLs against the HTTPS/SSH-only policy. The Kit is intentionally local-caller-only; Kit provider execution and recovery of an unadmitted attempt require healthy durable SQLite or PostgreSQL async-job admission. Kit mode supports Claude, Codex, and Mistral; every other provider fails closed while it is enabled. Kit scope never inherits the gateway process cwd: Claude and Mistral require a registered workspace selection or configured default, while Codex can also use an explicit absolute workingDir. Relative Kit workingDir values are rejected before filesystem or Git inspection. It disables cross-model validation tools while enabled and is not an HTTP/OAuth configuration or remote execution feature.

For a retention-pinned non-Kit Claude MCP request configuration, follow the local-only same-host recovery procedure. It accepts no SQL or arbitrary-path override. A valid recovery invocation emits JSON and exits 0 after success or 2 for a safe refusal; invalid usage or unavailable durable storage exits 1.

Release-readiness history is tracked in docs/personal-mcp/RELEASE_READINESS.md. Dogfooding evidence (which target LLMs guided setup, what unsafe suggestions were captured, and which findings were deferred from the initial personal-appliance rollout) is in docs/personal-mcp/DOGFOODING_RESULTS.md.

Current personal-appliance artifacts include:

• Streamable HTTP startup: LLM_GATEWAY_AUTH_TOKEN=<token> npm run start:http
• Machine-readable diagnostics: npm run doctor
• Go bootstrapper: installer/ with setup, doctor --json, start, stop, status, repair, upgrade, uninstall, print-client-config, and verified bundle download commands.
• Release packaging: on the public mirror, the release workflow builds Linux binaries on GitHub-hosted ubuntu-latest and builds Windows/macOS binaries on their GitHub-hosted platform runners. The private upstream uses its internal self-hosted Linux runner. Each release publishes checksummed platform bundles with the gateway, production dependencies, and a managed Node runtime; see installer/packaging/README.md.
• Docker Compose fallback: docker/personal.compose.yml + docker/Dockerfile.personal for users who already manage containers.
• Local setup UI artifact: setup/ui/index.html
• Provider setup snippets: setup/providers/
• Cross-validation tools: review_changes, validate_with_models, second_opinion, compare_answers, red_team_review, consensus_check, ask_model, synthesize_validation, job_status, job_result, and validation_receipt (plus the validation-receipt://{validationId} resource). review_changes and durable receipts require a SQLite or PostgreSQL validation-run store and are absent in Personal Agent Config Kit mode.

Install / Upgrade / Uninstall (single binary)

Windows PowerShell:

$Version = '<version>'
$Base = "https://github.com/verivus-oss/llm-cli-gateway/releases/download/v$Version"
$InstallDir = Join-Path (Join-Path $env:LOCALAPPDATA 'Programs') 'llm-cli-gateway'
$ExeName = "llm-cli-gateway-$Version-windows-amd64.exe"
$BundleName = "llm-cli-gateway-bundle-$Version-windows-amd64.tar.gz"
$Exe = Join-Path $InstallDir 'llm-cli-gateway.exe'
$Checksums = Join-Path $InstallDir 'SHA256SUMS'
$ChecksumBundle = Join-Path $InstallDir 'SHA256SUMS.sigstore.json'
New-Item -ItemType Directory -Force $InstallDir | Out-Null
Invoke-WebRequest -UseBasicParsing "$Base/$ExeName" -OutFile $Exe
Invoke-WebRequest -UseBasicParsing "$Base/SHA256SUMS" -OutFile $Checksums
Invoke-WebRequest -UseBasicParsing "$Base/SHA256SUMS.sigstore.json" -OutFile $ChecksumBundle
cosign verify-blob $Checksums --bundle $ChecksumBundle --certificate-identity "https://github.com/verivus-oss/llm-cli-gateway/.github/workflows/release-installer.yml@refs/tags/v$Version" --certificate-oidc-issuer "https://token.actions.githubusercontent.com"
if ($LASTEXITCODE -ne 0) { throw "Sigstore verification failed for SHA256SUMS" }
function Get-ReleaseSha256($Name) {
  $line = Select-String -Path $Checksums -Pattern "^[a-fA-F0-9]{64}\s+$([regex]::Escape($Name))$" | Select-Object -First 1
  if (-not $line) { throw "No SHA256SUMS entry found for $Name" }
  return (($line.Line -split "\s+")[0]).ToLowerInvariant()
}
if ((Get-FileHash $Exe -Algorithm SHA256).Hash.ToLowerInvariant() -ne (Get-ReleaseSha256 $ExeName)) { throw "Checksum mismatch for $ExeName" }
$env:RVWR_GATEWAY_BUNDLE_URL = "$Base/$BundleName"
$env:RVWR_GATEWAY_BUNDLE_SHA256 = Get-ReleaseSha256 $BundleName
& $Exe setup
& $Exe stop
& $Exe install-bundle
& $Exe start
& $Exe status
& $Exe doctor

The Windows installer keeps a stable llm-cli-gateway.exe command in %LOCALAPPDATA%\Programs\llm-cli-gateway and adds that directory to the user PATH. Do not script against release-versioned exe names after install.

# After downloading the binary that matches your OS/arch from a release:
cosign verify-blob SHA256SUMS --bundle SHA256SUMS.sigstore.json \
  --certificate-identity "https://github.com/verivus-oss/llm-cli-gateway/.github/workflows/release-installer.yml@refs/tags/v<version>" \
  --certificate-oidc-issuer "https://token.actions.githubusercontent.com"
sha256sum --check SHA256SUMS            # verify before run (or `shasum -a 256 --check` on macOS)
chmod +x llm-cli-gateway-<ver>-<os>-<arch>
./llm-cli-gateway-<ver>-<os>-<arch> setup
./llm-cli-gateway-<ver>-<os>-<arch> install-bundle    # uses the platform bundle URL/SHA256
./llm-cli-gateway-<ver>-<os>-<arch> start
./llm-cli-gateway-<ver>-<os>-<arch> doctor

# Upgrade: replace the binary, set the new bundle env vars, run upgrade.
./llm-cli-gateway-<new>-<os>-<arch> upgrade

# Uninstall: dry-run first, then run with --yes.
./llm-cli-gateway-<ver>-<os>-<arch> uninstall
./llm-cli-gateway-<ver>-<os>-<arch> uninstall --yes

Docker fallback:

LLM_GATEWAY_AUTH_TOKEN=$(openssl rand -hex 32) \
  docker compose -f docker/personal.compose.yml up -d
docker compose -f docker/personal.compose.yml run --rm doctor

Features

Core Capabilities

• Multi-LLM Orchestration: Unified interface for Claude Code, Codex, Gemini, Grok, Mistral (Vibe), Devin, and Cursor Agent CLIs
• Session Management: Track gateway session metadata and provider-specific continuity with persistent storage
• Gateway-owned worktrees: Run supported sync or async provider requests inside a managed git worktree with the local file-backed session manager. Same-session reuse requires same-host durable ownership plus a matching live Git registration and gateway branch; manager-level named path collisions fail closed. PostgreSQL-backed sessions reject worktrees before creation because a different database-connected host cannot safely own filesystem cleanup. Grok, Devin, and Mistral require an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests fail closed. A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks, configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. Session deletion and TTL eviction hide a durably owned worktree session while cleanup runs. If Git removal fails, the file-backed store retains a durable cleanup-pending tombstone, blocks reuse, and retries cleanup when that store is registered on the owning host. The record is finalized only after verified Git removal.
• Token Optimization: opt-in prompt/response compaction via optimizePrompt / optimizeResponse, both defaulting to false. The 44% / 37% figures are v1.0.0 measurements on a specific corpus, not a guarantee: src/optimizer.ts is a deterministic phrase-stripper, and nothing measures a ratio at runtime.
• Correlation ID Tracking: Full request tracing across all LLM interactions
• Cross-Tool Collaboration: LLMs can use each other via MCP (validated through dogfooding)

Observability

• Durable Flight Recorder: Provider requests and responses follow [persistence].backend: SQLite writes to ~/.llm-cli-gateway/logs.db by default, while PostgreSQL writes them to the configured database with the other durable subsystems. A PostgreSQL failure leaves the recorder unavailable or degraded and is reported generically on health surfaces; it never falls back to SQLite. Existing rows are not migrated or dual-read when the backend changes. Records include correlation IDs, duration, and token usage where the provider emits it (today: claude on stream-json/json, codex, and the API providers; grok, gemini, mistral, devin and cursor emit no usage on their CLI wire). Two exclusions are worth knowing: cross-LLM validation seats write no flight-recorder row at all, so llm_request_result cannot read one back by correlation ID, and repository-review seats are additionally excluded by design so review evidence is not retained in a non-expiring table. The retry_count and circuit_breaker_state columns are constants on the async path, which is the production path; retry and circuit breaking apply only to the direct-execute fallback. Read history through llm_request_list and llm_request_result so per-principal ownership checks remain in force. For human browsing of a SQLite deployment: datasette ~/.llm-cli-gateway/logs.db
• Structured Metadata: Tool responses include machine-readable structuredContent (model, cli, correlationId, sessionId, durationMs, token counts)
• Cache observability resources: cache-state://global, cache-state://session/{id}, and cache-state://prefix/{hash} MCP resources return aggregate cache hit/miss/savings — tokens and hashes only, no prompt text. session_get includes a cacheState block when the session has prior requests.
• Provider capability inventory: provider_tool_capabilities and provider-tools://catalog expose the gateway request fields, supported/degraded provider controls, local skill/tool discovery, and safe config-surface hints for Claude Code, Codex CLI, Gemini/Antigravity, Grok CLI/API, Mistral Vibe, Cognition Devin, and Cursor Agent. doctor --json includes a compact provider_capabilities summary for setup assistants.

Cache-aware operation

Every *_request and *_request_async tool except devin_request / devin_request_async and cursor_request / cursor_request_async accepts an optional promptParts field that structures the prompt for better cache hit rates (the Devin and Cursor headless paths take a plain prompt only). The gateway concatenates the parts in canonical order (system → tools → context → task) so that the stable prefix bytes precede the volatile task tail unchanged across calls, letting each provider's automatic prompt-caching land on the same content hash each time.

{
  "promptParts": {
    "system": "You are a helpful code reviewer.",
    "tools": "You have access to Read, Grep, Bash.",
    "context": "<long stable context block — file dumps, etc.>",
    "task": "Review the changes in src/foo.ts for security issues."
  }
}

prompt and promptParts are mutually exclusive — pass exactly one.

Per-CLI capability matrix (prefix discipline is automatic via promptParts for all providers except Devin and Cursor, which have no promptParts surface; explicit levers are provider-specific):

CLI	Prefix discipline	Explicit lever(s)
claude	yes	promptParts.cacheControl + outputFormat: "stream-json" (Anthropic cache_control breakpoints on stable blocks; ttl="1h" forced)
codex	yes	none (OpenAI implicit)
gemini	yes	none (implicit server-side)
grok	yes	compactionMode / compactionDetail (context compaction: `summary
mistral	yes	none (implicit)
devin	no	plain prompt only
cursor	no	plain prompt only

Claude example (explicit cacheControl)

claude_request({
  promptParts: {
    system: "You are a helpful code reviewer.",
    context: "<long stable file dump>",
    task: "Review the diff.",
    cacheControl: { system: true, context: true }, // task is never marked
  },
  outputFormat: "stream-json",
});

Gateway emits the stream-json stdin path with cache_control: {type:"ephemeral", ttl:"1h"} on marked blocks only.

Grok example (compaction)

grok_request({
  promptParts: { system: "...", context: "...", task: "..." },
  compactionMode: "segments",
  compactionDetail: "balanced",
});

Emits --compaction-mode segments --compaction-detail balanced.

See docs/personal-mcp/PROVIDER_CACHE_SURFACES.md for full surfaces, telemetry differences (e.g. Grok -p vs ACP), exact stream-json payload shapes, and cross-LLM review notes.

Opt-in flags (all default off) live under [cache_awareness] in ~/.llm-cli-gateway/config.toml.

Reliability & Performance

• Retry Logic: Exponential backoff with circuit breaker for transient failures
• Atomic File Writes: Process-specific temp files with fsync for data integrity
• Host-protection backpressure: bounded HTTP session lifecycle (max sessions + idle reaper), global and per-provider job-execution limits with a bounded FIFO queue, and a configurable per-job output cap (default 50MB). See Host-protection limits.
• NVM Path Caching: Eliminates I/O overhead on every request
• Long-Running Jobs: Non-time-bound async execution via *_request_async + polling tools

Security & Quality

• Comprehensive Testing: 1,700+ tests covering unit, integration, and regression scenarios with real CLI execution
• Input Validation: Zod schemas prevent injection attacks
• No Secret Leakage: Generic session descriptions only (file permissions 0o600)
• No ReDoS: Bounded regex patterns prevent catastrophic backtracking
• Type Safety: Strict TypeScript with comprehensive error handling
• Supply-chain hardening: a dedicated .github/workflows/security.yml runs actionlint, zizmor, shellcheck, typos, osv-scanner, gitleaks, and lychee on every push and PR (see SECURITY.md for the threat model)

Provider capability surface

Every provider is reachable through the same request, session, job, and validation machinery, but the underlying CLIs differ in what they natively expose. The table records what actually shipped per provider; discover the live surface at runtime with provider_tool_capabilities, list_models, and the provider-acp://<provider> / provider-tools://<provider> resources.

Provider	CLI request tools	Native ACP	Live model discovery	Admin surface
Claude Code (claude)	claude_request / _async	None (CLI-first; no ACP entrypoint at its tracked version)	model aliases, reasoning-effort levels, fallback model	read-only via provider_admin_list / provider_admin_run
OpenAI Codex (codex)	codex_request / _async, codex_fork_session	None (codex-cli advertises mcp-server / app-server transports, not native ACP)	codex debug models	read-only via provider_admin_list / provider_admin_run
Gemini / Antigravity (gemini, agy)	gemini_request / _async	None (agy exposes no ACP entrypoint; legacy Gemini CLI ACP evidence does not transfer)	agy models	read-only via provider_admin_list / provider_admin_run
xAI Grok (grok)	grok_request (sync transport: "acp") / _async	Native via grok agent stdio	grok models + ~/.grok/config.toml	read-only via provider_admin_list / provider_admin_run
Mistral Vibe (mistral)	mistral_request (sync transport: "acp") / _async	Native via vibe-acp	Vibe config plus the VIBE_ACTIVE_MODEL active model and agent profiles	read-only via provider_admin_list / provider_admin_run
Cognition Devin (devin)	devin_request (sync transport: "acp", agentType: summarizer|review) / _async	Native via devin acp	--model / DEVIN_MODEL	read-only via provider_admin_list / provider_admin_run
Cursor Agent (cursor)	cursor_request (sync transport: "acp") / _async	Native via cursor-agent acp (companion-owned)	model aliases	read-only via provider_admin_list / provider_admin_run

• Native ACP is reported honestly. grok, mistral, devin, and cursor expose a native ACP entrypoint, so provider-acp://<provider> carries the negotiated initialize capability set and the derived session-method availability, and the sync *_request accepts transport: "acp" (fails closed unless [acp] and the provider's runtime_enabled gate are set). ACP routing is sync-only: the *_request_async variants always run the CLI transport and do not accept transport: "acp" (nor Devin's agentType); async ACP parity is a later phase. ACP workspace selection is gateway-owned: an explicit ACP workspace must be a registered alias. A fresh remote ACP request uses that alias or [workspaces].default; a remote resume is fixed to its recorded canonical alias and cwd, and a different or unbound workspace is rejected. Local ACP may omit workspace; each unscoped process then gets a fresh private 0o700 neutral directory that is removed after the process exits, never a shared predictable temp path. claude, codex, and gemini have no native ACP entrypoint at their target CLI versions; their provider-acp:// records report native: false with no methods and no adapter-as-native masquerade, and they expose no transport: "acp" selector.
• Managed approval is Claude-only today. approvalStrategy:"mcp_managed" is executable only by the Claude CLI adapter, which launches Claude with a request-scoped generated MCP configuration and --strict-mcp-config. It permits only provisioned, gateway-owned MCP definitions and rejects dynamic npx, ambient-PATH, and Codex-config overrides. Codex, Gemini, Grok, Mistral, Devin, and Cursor reject mcp_managed before launching a provider because their current adapters cannot isolate ambient MCP configuration. For those adapters, use approvalStrategy:"legacy"; approvalPolicy has no effect.
• ACP has its own permission bridge. approvalStrategy:"mcp_managed" and any approvalPolicy are rejected when transport:"acp" is selected. Use the Claude CLI transport for managed approval. ACP host services fail closed, gated by the tool-call category the agent declares: write is denied unless allow_write_host_services is set, execute unless allow_terminal_host_services is set, and any other category is denied outright, including fetch and every unrecognised kind, so a new upstream tool kind cannot be auto-approved by default. Approval is expressed only by selecting an agent-offered single-use allow option; an agent that offers only a persistent allow_always grant is denied, so ACP never maps a raw CLI bypass input to a standing permission grant. Note the boundary is this category gate: it constrains what the agent may ask the gateway to do, not what the agent does in its own process.
• Resources are generated from the provider registry for every CLI provider: models://<provider>, sessions://<provider>, provider-acp://<provider>, provider-tools://<provider>, and provider-subcommands://<provider>.
• Model discovery is live and account-aware: the discovery listed above reaches models://<provider> and list_models, degrading to static registry facts when a live probe is unavailable (a resource read never spawns a CLI).
• Admin surfaces are discovery-driven and output-redacted. provider_admin_list and provider_admin_run are read-only for every provider. State-mutating admin operations are exposed only through provider_admin_mutate, gated behind [admin] allow_mutating_cli_admin_ops, the remote cli:admin scope, an approval gate, and an audit record. Mutating ACP session operations are likewise gated behind [acp] allow_mutating_session_ops.
• Validation commands work across every provider: review_changes, validate_with_models, second_opinion, compare_answers, red_team_review, consensus_check, ask_model, and synthesize_validation, with canonically hashed immutable receipts via validation_receipt and the validation-receipt://{validationId} resource. review_changes captures a complete, hashed Git artifact and starts repository-bound read-only reviewers.

Prerequisites

Node.js >= 24.4.0 is required (engines.node in package.json). The SQLite backend uses Node's built-in node:sqlite module, so there is no native binding to compile and no install scripts run. The 24.4 floor is where allowBareNamedParameters defaults to true, which the SQLite persistence layer relies on.

Running the source-tree release audit also requires Bash and flock from util-linux. This prerequisite was verified against commit 242e7669565ecc7c71183f5f7133791a1384ca7c. Release automation runs the audit on Ubuntu. On macOS, install a compatible flock before running npm run security:audit or the full npm run check gate.

Before using this gateway, you need to install the CLI tools you want to use:

Claude Code CLI

# Installation instructions for Claude Code
# Visit: https://docs.anthropic.com/claude-code
npm install -g @anthropic-ai/claude-code

Codex CLI

npm install -g @openai/codex
codex login

Gemini (Google Antigravity CLI)

The Gemini provider runs through Google Antigravity CLI (agy).

curl -fsSL https://antigravity.google/cli/install.sh | bash
# Docs: https://antigravity.google/docs

Grok Build CLI (xAI)

curl -fsSL https://x.ai/cli/install.sh | bash
grok login   # OAuth flow; for headless auth, set XAI_API_KEY
# Docs: https://docs.x.ai/build/overview

Mistral Vibe CLI

# Pick one — the gateway's cli_upgrade auto-detects which one you used.
curl -LsSf https://mistral.ai/vibe/install.sh | bash
pip install mistral-vibe
uv tool install mistral-vibe
brew install mistral-vibe

vibe --setup
# Complete the API-key setup locally. Do not paste the key into a chat.
# Current Vibe defaults session logging to enabled. If an older config disabled it,
# edit ~/.vibe/config.toml and set:
# [session_logging]
# enabled = true

Vibe-specific notes:

• Model selection is via the VIBE_ACTIVE_MODEL environment variable — Vibe has no --model flag. The gateway discovers ~/.vibe/config.toml / VIBE_MODELS, injects VIBE_ACTIVE_MODEL only when a model is explicitly requested or Vibe config needs recovery, and retries once after a model-not-found failure with refreshed discovery.
• permissionMode is the Vibe --agent name. Builtins are default | plan | accept-edits | auto-approve; Vibe also accepts install-gated builtins (e.g. lean) and custom agents from ~/.vibe/agents. Requests pass the selected name through for Vibe to validate. mcp_managed is not available for Vibe.
• Tool controls use Vibe's native flags. The gateway emits one --enabled-tools <tool> flag per allowedTools entry and one --disabled-tools <tool> flag per disallowedTools entry. Vibe applies disabled tools after enabled-tool filtering.
• Usage telemetry is best-effort. Vibe does not emit token or cost data in programmatic stdout. When the gateway knows Vibe's native session UUID, it reads ~/.vibe/logs/session/session_<...>/meta.json for usage and cost. A missing, malformed, or not-yet-known session log leaves those fields empty.
• No self-update: cli_upgrade --cli mistral detects whether you used pip / uv / brew and dispatches the matching upgrade command. Running vibe update is not a thing.

Installation

As an MCP server (npm)

npm install -g llm-cli-gateway

Or use directly with npx:

{
  "mcpServers": {
    "llm-gateway": {
      "command": "npx",
      "args": ["-y", "llm-cli-gateway"]
    }
  }
}

From source

git clone https://github.com/verivus-oss/llm-cli-gateway.git
cd llm-cli-gateway
npm install
npm run build

Usage

As an MCP Server

For clients that already support local stdio MCP servers, add a configuration like:

{
  "mcpServers": {
    "llm-cli-gateway": {
      "command": "node",
      "args": ["/path/to/llm-cli-gateway/dist/index.js"]
    }
  }
}

Stdio is the recommended path for unrestricted machine-local development access. HTTP MCP, including localhost HTTP and tunneled HTTPS, is treated as remote-capable for provider execution: provider tools must resolve a registered workspace alias, a session workspace, or [workspaces].default before spawning a CLI. Remote clients should pass relative workingDir, addDir, and include-directory values inside the selected workspace, and may resume only gateway-tracked sessions they own. Raw native provider session IDs are local-only. Disabling auth or using a no-auth connector path is not a filesystem bypass.

For a local CLI request with no resolved workingDir, registered workspace, or gateway-managed worktree, the child runs in a fresh private 0o700 temporary directory that is removed after the process exits. It never inherits the gateway repository cwd or its provider-native instruction context. The gateway canonicalizes the temp root and rejects or relocates it when any ancestor contains .git, AGENTS.md, AGENTS.override.md, Agents.md, AGENT.md, CLAUDE.md, Claude.md, CLAUDE.local.md, .claude/CLAUDE.md, .claude/rules/, .cursor/rules/, .cursorrules, GEMINI.md, or .vibe/config.toml, including through a symlinked or custom TMPDIR beneath that context. The list covers entries a provider discovers by walking up from its cwd. User-scope configuration such as ~/.claude/settings.json is deliberately absent: it loads on every invocation regardless of cwd, so relocating the workspace would not isolate it. Provider-native resumeLatest operations that use a cwd-scoped latest-session pointer therefore require an explicit workingDir, workspace, or configured default workspace and fail closed when none is available. Use explicit target selection whenever several repositories are active at once.

CLI request schemas accept prompts up to 100,000 characters, but operating systems also impose byte limits on individual argv elements. Codex new and resume requests stream the exact prompt over stdin. codex_fork_session remains argv-bound and rejects an oversized UTF-8 prompt before spawn as non-retryable input_too_large. Other providers whose current CLI contracts require an argv prompt use the same admission rule. Every other caller-controlled argv value is checked on its final encoded form too, including serialized agent/schema JSON, joined tool lists, instruction overrides, paths, model names, and native session IDs. The final spawn boundary checks every argv element plus the aggregate resolved command line against a conservative platform-specific byte budget and a 2,048-element cap. The aggregate byte budget excludes the environment but reserves headroom for it; on Windows, pre-resolution admission assumes the smaller npm .cmd/.bat wrapper limit until command resolution proves a native executable. Native session and resume flags on non-Kit requests are included before workspace, session, provider-artifact handoff, or durable job side effects. Claude Kit projects its eventual argv before materializing its compiled context artifact or allocating a durable Kit session. An embedded NUL byte in the command or any argv element is rejected before spawn as non-retryable invalid_input. Caller-facing results, long-lived job memory, durable job args, and async flight rows use a fixed invalid-argv marker; the optional duplicate durable payload is suppressed. None retains the rejected vector or Node's value-echoing native error. Native E2BIG, including an environment-driven failure, is normalized without retaining the native spawnargs. The gateway never truncates instructions or other values to make them fit. For stdin-backed requests, a clean provider exit is accepted only after the complete payload write callback succeeds. A closed or still-pending pipe becomes a fixed, non-sensitive incomplete-delivery failure; timeout, cancellation, and provider nonzero exits remain authoritative.

This generic stdio example is not provider-support verification for the Personal MCP Appliance. Client-specific setup guides for ChatGPT, Claude web, Claude Desktop, Codex, Gemini CLI, Gemini web, and Grok remain gated by the provider-support matrix in docs/personal-mcp/PRODUCT_CONTRACT.md.

Available Tools

Cross-LLM Validation Tools

The personal-appliance surface exposes simplified validation tools for non-developer clients. These tools start provider CLI jobs through the durable async job manager and return normalized provider status plus raw job references.

• validate_with_models: ask two or more providers to independently validate a question.
• review_changes: capture one complete Git review artifact, fence repository content as untrusted data, and start read-only independent reviewers. See Repository change review.
• second_opinion: ask one provider to review an answer.
• red_team_review: challenge a plan, answer, or document for risks and failure modes.
• consensus_check: check whether providers agree with a claim.
• ask_model: ask one provider through the simplified surface.
• synthesize_validation: run an explicit judge model after provider results have been collected. General validation requires the caller's question and terminal normalized results. A review_changes run instead reloads its exact owned durable results from validationId; caller-supplied question/results are ignored.
• list_available_models: list the models each provider CLI exposes through the simplified surface.
• job_status and job_result: poll and collect validation job outputs.
• validation_receipt: retrieve the canonically hashed immutable receipt of a terminal cross-LLM validation run by validationId (returns minted | pending | verification_failed | expired_unminted | not_found, own-or-not-found). verification_failed means a stored receipt exists but disagrees with its durable run, which is a defect to investigate; expired_unminted only ever means absence. format: "markdown" renders a human-readable report; includeRawResponses inlines complete provider answer text when the linked job still exposes identity-verified output. Registered only when the attached job store provides the durable validation-run store capability (sqlite and postgres).

The same receipt is also exposed as the validation-receipt://{validationId} MCP resource (same durable gate and own-or-not-found owner scoping).

The validation report preserves per-provider disagreement. Optional judge synthesis is explicit about which provider produced the judge job.

Repository change review

review_changes accepts an absolute local workingDir or a registered workspace, then resolves scope: "auto" | "uncommitted" | "branch" | "commit". It can take an explicit Git base, literal repository-relative paths, stance: "standard" | "adversarial", reviewer models, an optional judgeModel, trustCursorWorkspace, and fail-closed artifact/prompt byte ceilings.

Cursor refuses to review a directory it does not trust, so a cursor seat is granted --trust only when the reviewed directory is a registered [[workspaces.repos]] path whose providers include cursor (a gateway worktree beneath one counts, and the nearest enclosing registration decides). On any other directory the cursor seat is skipped with an actionable reason, and the rest of the roster still runs. trustCursorWorkspace: true accepts the grant for one durable review run. The consent applies to Cursor seats in that run's reviewer roster and to its planned Cursor judge when synthesize_validation launches it later. It is bound to the run owner, repository, and planned judge, so another principal or review run cannot replay it. That is a real decision rather than a formality: a trusted folder is also where cursor loads project rules and AGENTS.md, so the repository under review gains some influence over its own reviewer, which the fenced review prompt otherwise forbids. The artifact keeps committed, staged, unstaged, and regular non-ignored untracked file evidence separate. It forces tracked diffs to remain readable even when in-tree attributes mark them as non-diffable. The review-evidence.v2 artifact exposes committedPatch, stagedPatch, and unstagedPatch independently; each segment carries its sorted path inventory, encoding, exact byte length, SHA-256 identity, and content. This prevents an index change and its worktree-only reversal from canceling out. The artifact is collision-fenced, byte-counted, SHA-256 identified, race-checked, and never truncated. In auto mode, a diverged branch is reviewed from its merge base with working-tree evidence included. Otherwise, a dirty tree selects uncommitted changes, while a clean tree falls back to the last commit (HEAD^..HEAD) without working-tree evidence. Unsafe untracked file types or a repository mutation during capture cause a refusal.

The tool starts asynchronous provider jobs and returns a validationId, exact artifact and prompt identities, file inventory, and one rawJobReference per reviewer. Poll those references with validation job_status and collect them with validation job_result, not the similarly named llm_job_* tools. If a judge was requested, wait for every reviewer to become terminal, then call synthesize_validation with the validationId and the same workingDir or workspace selector. Continue collecting results for progress and human visibility, but do not pass them as review evidence: for a review_changes run, the gateway ignores caller-supplied question and providerResults, reloads the exact owned durable linked terminal jobs, and reconstructs requested but unavailable seats as skipped. General validation synthesis still requires a caller-supplied question and terminal normalized results.

The review surface is registered only with durable SQLite or PostgreSQL job and validation-run storage. Each CLI review job retains the exact fenced prompt in its expiry-bound payload_json; its persisted argv contains only a hash marker. The non-expiring flight recorder does not receive repository-review prompts. Configured HTTP/API reviewer seats require explicit allowApiUpload:true because the complete artifact leaves the local CLI boundary. Remote HTTP/OAuth workspace reviews reject API reviewer uploads even with that flag. Treat the durable job store as sensitive until the configured job retention expires. When judgeModel is an HTTP/API provider, review_changes binds that explicit consent, the judge provider, the resolved repository, and the caller identity to the durable validationId. The later synthesize_validation call must provide that id and the same repository selector. The stored judge, repository, owner, and upload consent are authoritative. The gateway atomically claims the planned judge once, so concurrent or repeated synthesis cannot start a second judge. A follow-up argument cannot grant or override upload consent.

LLM Request Tools

claude_request

Execute a Claude Code request with optional session management.

Parameters:

• prompt (string, optional*): The prompt to send (1-100,000 chars). *Exactly one of prompt or promptParts is required (mutually exclusive)
• model (string, optional): Model name or alias (use list_models for available values; supports latest)
• outputFormat (string, optional): Output format (text|json|stream-json), default: stream-json — the gateway parses NDJSON usage events for token/cost observability; override to text only when you want unparsed stdout
• sessionId (string, optional): Specific session ID to use. Under mcp_managed, native continuation is a high-risk input because it can inherit an unverified provider posture; it requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but does not select a full-permission profile.
• continueSession (boolean, optional): Continue the active session. It has the same managed-approval requirement as sessionId. Because Claude --continue selects by cwd, it requires workingDir or a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default. That workspace may optionally supply a gateway worktree. The request fails closed when no selection supplies a stable cwd.
• createNewSession (boolean, optional): Always create a new session
• forkSession (boolean, optional): Fork the resumed session instead of appending to it. Under mcp_managed, it is a high-risk native-fork input that requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but stays bounded.
• allowedTools (string[], optional): Restrict Claude tools to this allow-list. A non-empty allow-list is a high-risk managed input because it can change the tool posture.
• disallowedTools (string[], optional): Explicitly deny listed Claude tools
• permissionMode (string, optional): Claude permission mode (default|acceptEdits|plan|auto|dontAsk|bypassPermissions); preferred over dangerouslySkipPermissions. bypassPermissions is a direct full-permission request under mcp_managed.
• dangerouslySkipPermissions (boolean, optional): Deprecated, maps to permissionMode: "bypassPermissions"; permissionMode wins when both are set. It is a direct full-permission request under mcp_managed.
• agent (string, optional): Named sub-agent to run as. A non-empty value is a high-risk managed input because it can change tool and permission posture.
• agents (string, optional): Inline agent definitions JSON. A non-empty value is a high-risk managed input for the same reason.
• systemPrompt / appendSystemPrompt (string, optional): Replace or extend the system prompt. A non-empty value is a high-risk managed input.
• systemPromptFile / appendSystemPromptFile (string, optional): Replace or extend the system prompt from a file. A non-empty file path is a high-risk managed input.
• safeMode (boolean, optional): Start Claude with local customizations disabled, including CLAUDE.md, skills, plugins, hooks, MCP, commands, and agents. true is a high-risk managed input.
• bare (boolean, optional): Start Claude in minimal mode, skipping local customization discovery. true is a high-risk managed input.
• debugFile (string, optional): Write Claude debug output to a file. A non-empty path is a high-risk managed input.
• maxBudgetUsd (number, optional): Budget cap in USD for the request
• maxTurns (integer, optional): Agent-loop turn cap
• effort (string, optional): Reasoning effort (low|medium|high|xhigh|max)
• fallbackModel (string, optional): Auto-fallback model when the default is overloaded
• jsonSchema (string, optional): JSON Schema literal constraining structured output
• addDir (string[], optional): Additional workspace directories. A non-empty value is a high-risk managed input.
• noSessionPersistence (boolean, optional): Ephemeral session (not persisted to disk)
• settingSources / settings / tools (optional): Setting sources to load, settings JSON path/literal, built-in tool restriction. Non-empty setting sources, settings, or tool selections are high-risk managed inputs.
• pluginDir / pluginUrl (string[], optional): Load Claude plugins from local directories or URLs. Non-empty values are high-risk managed inputs.
• excludeDynamicSystemPromptSections (boolean, optional): Trim dynamic system prompt sections
• approvalStrategy (string, optional): "legacy" (default) or "mcp_managed". Managed mode uses acceptEdits by default and forces strictMcpConfig:true, so Claude uses only the gateway-generated MCP configuration. A direct full-permission request requires all of an explicit caller request, an approval-manager approval, and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. Other high-risk inputs require the approval and operator setting too, but remain bounded and do not themselves select full permission.
• approvalPolicy (string, optional): "strict", "balanced", or "permissive"
• mcpServers (string[], optional): Names of MCP servers to expose to Claude (default: none). Legacy requests resolve names from the local registry or Codex MCP config; unknown names are reported as unavailable. Under mcp_managed, Claude uses only the generated configuration and only registry entries explicitly provisioned as gateway-owned local commands are eligible. Dynamic npx launchers, ambient-PATH commands, and Codex-config overrides are rejected. Configure and deploy the managed entries in the gateway environment.
• strictMcpConfig (boolean, optional): In legacy mode this defaults to false; set true to require only the generated MCP config and fail if requested servers are unavailable. Under mcp_managed, the gateway forces it to true and a caller-supplied false cannot weaken that boundary.
• optimizePrompt (boolean, optional): Optimize prompt for token efficiency (44% reduction), default: false
• optimizeResponse (boolean, optional): Optimize response for token efficiency (37% reduction), default: false
• correlationId (string, optional): Request trace ID (auto-generated if omitted)
• idleTimeoutMs (integer, optional): Kill a stuck process after output inactivity; 30,000 to 3,600,000 ms. Idle enforcement applies only when outputFormat is stream-json; it is ignored for text/json, which produce no output until the run completes
• worktree (boolean|object, optional): Run inside a gateway-owned git worktree (slice λ). A worktree requires a registered workspace selected explicitly, through caller-owned session metadata, or by the configured default; it never inherits process cwd or combines with workingDir, addDir, or includeDirs. Materialization suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, so filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. Requesting a worktree is a high-risk managed input that requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1, but remains bounded.
• promptParts (object, optional): Cache-aware structured prompt { system?, tools?, context?, task }; mutually exclusive with prompt
• forceRefresh (boolean, optional): Bypass dedup and force a fresh CLI run, default: false

Workspace boundary: stdio callers may use machine-local paths directly. HTTP/tunnel callers must pass workspace or rely on a configured default/session workspace; path fields are then validated relative to that workspace. [workspaces].allow_unregistered_working_dir is an inert legacy key: it is still accepted so old configs keep loading, but nothing reads it at either value, and setting it now logs a warning at startup. It never allowed arbitrary HTTP working directories or additional directories.

Response extras:

• approval: Approval decision record when approvalStrategy="mcp_managed"
• mcpServers: Requested/enabled/missing MCP servers for this call

Example:

{
  "prompt": "Write a Python function to calculate fibonacci numbers",
  "model": "sonnet",
  "continueSession": true,
  "optimizePrompt": true,
  "optimizeResponse": true
}

codex_request

Execute a Codex request with optional session tracking.

Parameters:

---

View the full README on GitHub