flightdeck-connect 0.1.1

The MCP rail for NorthGate AI FlightDeck — provision governed MCP tools into Claude Code, Cursor, and Codex; run governed agent lanes; verify signed evidence packs offline.

flightdeck-connect

The MCP rail for NorthGate AI FlightDeck — the governed multi-agent engineering console. Provision governed MCP tools into the coding tools you already use (Claude Code, Cursor, Codex), run governed agent lanes, and verify signed evidence packs offline.

(Import package and console-script names retain the legacy conductor naming — FlightDeck was formerly Conductor, and existing integrations are never broken by a rename.)

After pip install flightdeck-connect you get:

• flightdeck-connect — detect the AI coding tools on this machine and idempotently provision the FlightDeck MCP server into each (Claude Code project .mcp.json, Cursor mcp.json, codex mcp add), profiled by task, chain-recording the provisioning when a board is given.

• conductor — a stdlib-only HTTP client for the /v1/conductor/* API.

• conductor-mcp — a stdio JSON-RPC Model Context Protocol server that exposes 34 tools (21 conductor_* + 13 flightdeck_*, scoped at runtime via MCP_TOOL_SET=flightdeck|conductor|all) to MCP-aware agent runtimes (Claude Code, Cursor, Codex, VS Code agents). The flightdeck_* family implements the FlightDeck Agent Contract v1 loop — get_handoff → read_exam → prepare_workspace → submit_return.

• flightdeck-verify — standalone offline evidence-pack verifier (stdlib-only; exit 0/1/2; --json).

• flightdeck-gate — pre-merge git gate for verify-gated merges (chain walk + accept-gate conjuncts + reference-transaction guard).

Install

pip install flightdeck-connect

Zero runtime dependencies (pure Python stdlib). Python 3.8+.

Auth

The CLI authenticates one of two ways:

Header	Env var	Source
X-Agent-Token	CONDUCTOR_AGENT_TOKEN	Service token issued by /v1/conductor/projects/{id}/agent-tokens. Preferred for agents.
X-JWT-Token	CONDUCTOR_USER_JWT	A standard TrustSHIELD access token. Fallback for humans.

export CONDUCTOR_AGENT_TOKEN=cdtr_...
conductor whoami

CLI quickstart

# Probe auth + caller mode
conductor whoami

# Project + task ops
conductor project list
conductor project create --name "TrustSHIELD Conductor delivery"
conductor task list --project <project-id>

# Workflow transition (Policy Mesh gated)
conductor transition TASK-0040 \
    --action submit_review --to-state review_ready \
    --actor-member-id <member-id> \
    --reason "tests pass, ruff clean" \
    --evidence gitlab_mr=https://gitlab.com/.../merge_requests/126

# Release pack — download + offline verify
conductor export --project <project-id> --out release.tar.gz
conductor verify release.tar.gz
# -> Exit 0 + "OK: pack integrity verified." == nothing tampered.

conductor --help lists every subcommand. --json on any command gives machine-readable output for scripting; the default is friendly text.

Local board mode

Local board mode is the zero-cloud bridge for IDEs. It operates directly on a filesystem board folder:

<Project Conductor folder>/queue/TASK-*.md

No CONDUCTOR_AGENT_TOKEN or CONDUCTOR_USER_JWT is required.

BOARD="C:/Users/ericg/OneDrive - NorthGate Strategic LLC/Project Conductor"

conductor --json local list --board "$BOARD"
conductor --json local list --board "$BOARD" --state READY
conductor --json local read TASK-0001 --board "$BOARD"
conductor local handoff TASK-0001 --board "$BOARD"

Create and update packets:

conductor --json local create \
  --board "$BOARD" \
  --title "Audit the installer handoff" \
  --objective "Verify artifacts and append evidence." \
  --owner Codex \
  --reviewer "Claude Code"

conductor --json local signal TASK-0001 \
  --board "$BOARD" \
  --state REVIEW_READY \
  --action submit_review \
  --verdict "Audit completed; report attached." \
  --next "Reviewer checks the evidence and accepts or requests changes."

MCP server (Claude Code / Cursor / Codex)

The conductor-mcp script implements the Model Context Protocol (2024-11-05) over stdio. Sample config snippets:

Claude Code (~/.config/claude/claude_desktop_config.json or project- level .mcp.json):

{
  "mcpServers": {
    "conductor": {
      "command": "conductor-mcp",
      "env": {
        "CONDUCTOR_AGENT_TOKEN": "cdtr_..."
      }
    }
  }
}

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "conductor": {
      "command": "conductor-mcp",
      "env": {
        "CONDUCTOR_AGENT_TOKEN": "cdtr_..."
      }
    }
  }
}

After restart, the agent's tool list includes 21 conductor_* tools — conductor_project_list, conductor_transition, conductor_export, conductor_verify, plus local-board tools such as conductor_local_list and conductor_local_signal. Tool-execution failures surface as isError: true in the MCP response (per spec); protocol/transport failures use the JSON-RPC error channel.

Custom API endpoint

If you're targeting an environment other than the default (https://airec-api-dxol6hwotq-uk.a.run.app):

export CONDUCTOR_API_BASE=https://your-conductor-api.example.com
# OR pass per-invocation:
conductor --api-base https://your-conductor-api.example.com whoami

Offline pack verification

conductor verify <pack.tar.gz> works without API access. It extracts the archive in a temp dir, walks SHA256SUMS to verify every file's hash, then walks audit-chain/chain.jsonl to confirm per-tenant chain integrity (monotonic chain_position + prev_event_hash threading). Either kind of tampering surfaces a non-zero exit and a specific error.

The check is identical to the embedded audit-chain/verify.py script that ships inside every pack — conductor verify is the more convenient hand-typed form.

Source

Mirror of packages/api/src/conductor/{cli,mcp}.py from the upstream TrustSHIELD repo. See Conductor-Build-Plan-2026-05-23.md sections 4.2 + 5 for the full design.