agentxchain

CLI for governed multi-agent software delivery.

The canonical mode is governed delivery: orchestrator-owned state, structured turn results, phase gates, mandatory challenge, and explicit human approvals where required.

Legacy IDE-window coordination is still shipped as a compatibility mode for teams that want lock-based handoff in Cursor, VS Code, or Claude Code.

Docs

• Quickstart
• CLI reference
• Export schema reference
• Adapter reference
• Protocol spec (v6)
• Protocol reference
• Why governed multi-agent delivery matters

Install

npm install -g agentxchain

Or run without installing:

npx agentxchain init --governed --dir my-agentxchain-project -y

Testing

The CLI currently uses two runners on purpose: a 36-file Vitest coexistence slice for fast feedback and node --test for the full suite.

npm run test:vitest
npm run test:node
npm test

• npm run test:vitest: the current 36-file Vitest slice
• npm run test:node: full integration, subprocess, and E2E suite
• npm test: both runners in sequence; this is the CI requirement today

Duplicate execution remains intentional for the current 36-file slice until a later slice explicitly changes the redundancy model. For watch mode, run npx vitest.

Quick Start

Governed workflow

npx agentxchain init --governed --dir my-agentxchain-project -y
cd my-agentxchain-project
git init
git add -A
git commit -m "initial governed scaffold"
agentxchain status
agentxchain step --role pm

The default governed dev runtime is claude --print with stdin prompt delivery. If your local coding agent uses a different launch contract, set it during scaffold creation:

npx agentxchain init --governed --dir my-agentxchain-project --dev-command ./scripts/dev-agent.sh --dev-prompt-transport dispatch_bundle_only -y

If you want template-specific planning artifacts from day one:

npx agentxchain init --governed --template api-service --dir my-agentxchain-project -y

Built-in governed templates:

• generic: baseline governed scaffold
• api-service: API contract, operational readiness, error budget
• cli-tool: command surface, platform support, distribution checklist
• library: public API, compatibility policy, release and adoption checklist
• web-app: user flows, UI acceptance, browser support

step writes a turn-scoped bundle under .agentxchain/dispatch/turns/<turn_id>/ and expects a staged result at .agentxchain/staging/<turn_id>/turn-result.json. Typical continuation:

agentxchain validate --mode turn
agentxchain accept-turn
agentxchain approve-transition
agentxchain step --role dev
agentxchain step --role qa
agentxchain approve-completion

Default governed scaffolding configures QA as api_proxy with ANTHROPIC_API_KEY. For a provider-free walkthrough, switch the QA runtime to manual before the QA step. If you override the dev runtime, either include {prompt} for argv delivery or set --dev-prompt-transport explicitly.

Multi-repo coordination

For initiatives spanning multiple governed repos, use the coordinator to add cross-repo sequencing and shared gates:

npx agentxchain init --governed --template api-service --dir repos/backend -y
npx agentxchain init --governed --template web-app --dir repos/frontend -y
agentxchain multi init
agentxchain multi step --repo backend --role pm

See the multi-repo quickstart for the full cold-start walkthrough.

Migrate a legacy project

agentxchain migrate
agentxchain status
agentxchain step

Command Sets

Governed

Shared utilities

Legacy v3 compatibility

Governed Flow

1. agentxchain step initializes or resumes the run if needed.
2. It assigns the next role for the current phase.
3. It writes .agentxchain/dispatch/turns/<turn_id>/.
4. The assigned role writes .agentxchain/staging/<turn_id>/turn-result.json.
5. The orchestrator validates and either accepts, rejects, advances phase, pauses for approval, or completes the run.

Protocol Conformance

AgentXchain ships a conformance kit under .agentxchain-conformance/. Use it to prove a runner or fork still implements the governed workflow contract:

agentxchain verify protocol --tier 3 --target .

Useful flags:

• --tier 1|2|3: maximum conformance tier to verify
• --surface <name>: isolate one surface such as state_machine, dispatch_manifest, or coordinator
• --format json: emit a machine-readable report for CI
• --target <path>: point at the root containing .agentxchain-conformance/capabilities.json

Important governed files:

agentxchain.json
.agentxchain/state.json
.agentxchain/history.jsonl
.agentxchain/decision-ledger.jsonl
.agentxchain/dispatch/turns/<turn_id>/
.agentxchain/staging/<turn_id>/turn-result.json
TALK.md
.planning/

Runtime support today

• manual: implemented
• local_cli: implemented
• api_proxy: implemented for synchronous review-only turns and stages a provider-backed result during step

Legacy IDE Mode

Legacy mode is still useful if you specifically want one IDE session per agent and lock-file coordination.

agentxchain start                   # Cursor (default)
agentxchain start --ide vscode
agentxchain start --ide claude-code
agentxchain kickoff
agentxchain supervise --autonudge

In this mode, agents hand off through lock.json, state.json, triggers, and TALK.md. For new projects, prefer governed mode unless IDE-window choreography is the goal.

macOS Auto-Nudge

supervise --autonudge is legacy-only and macOS-only.

agentxchain supervise --autonudge
agentxchain supervise --autonudge --send

Requires:

• osascript
• jq
• Accessibility permissions for Terminal and Cursor

Links

• agentxchain.dev
• Quickstart
• CLI reference
• Adapter reference
• Protocol spec (v6)
• GitHub
• Legacy Protocol v3 spec

License

MIT