RTMX

The open source client for RTMX Sync — what comes after Git for humans and agents.

Git versioned the file. Sync is the room: a live matrix of intent where humans and agents claim work, converge, and ship. This repo is the Apache 2.0 CLI and MCP surface — how you enter that room. Local CSV-in-git still works offline; Sync is how teams stay live.

Explore Sync → · Plans · Read the backstory

Install

brew install rtmx-ai/tap/rtmx        # macOS / Linux (CLI + broker binary)
scoop install rtmx                    # Windows
go install github.com/rtmx-ai/rtmx/cmd/rtmx@latest

Or download a binary from releases. Release archives include both rtmx and the verify broker daemon.

The packaged verification broker enables automatically on formula/package install and upgrade. If it is still missing, the first rtmx verify elevates once. Steady-state rtmx verify does not prompt:

rtmx broker status
# manual / fleet: rtmx broker install --prefix="$(brew --prefix rtmx)"

rtmx login
rtmx sync

Self-hosted: rtmx sync --sync-url sync.example.com/sync/acme/main (the client adds wss://).

What It Does

Command	What it does
rtmx sync	Join a Sync room (hosted or self-managed)
rtmx login / checkout	Managed auth and Team Checkout
rtmx status	Completion dashboard across all requirements and phases
rtmx next --one	Pick the highest-priority unblocked requirement
rtmx verify	Run tests and cross-reference against requirements
rtmx health	Lint your RTM: orphaned tests, circular deps, stale refs
rtmx backlog	Prioritized work items with critical path analysis
rtmx mcp-server	Scientific workflow MCP tools (loop_tick, delivery_check, trades, …)

33 commands total. Run rtmx --help for the full list.

The AI Workflow

flowchart TD
    A["Requirements\n(CSV in git)"] --> B["rtmx next\nPick unblocked requirement"]
    B --> C["Agent writes\ncode + tests"]
    C --> D{"rtmx verify\nTests pass?"}
    D -- yes --> E["Status updates\nautomatically"]
    D -- no --> C
    E --> F["rtmx status\nTeam sees progress"]
    F -.-> B

    style A fill:#d1fae5,stroke:#059669,color:#065f46
    style B fill:#d1fae5,stroke:#059669,color:#065f46
    style C fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    style D fill:#fef3c7,stroke:#d97706,color:#92400e
    style E fill:#d1fae5,stroke:#059669,color:#065f46
    style F fill:#d1fae5,stroke:#059669,color:#065f46

An agent runs rtmx next --one, gets a specific requirement, writes code and tests against the spec, and rtmx verify confirms the work is done. No human triages, assigns, or updates a ticket.

Why CSV in Git

block-beta
    columns 1
    block:header["database.csv -- Pull Request #42"]
        columns 1
    end
    block:removed["- REQ-AUTH-003, auth, mfa, TOTP-based MFA, ..., missing"]
        columns 1
    end
    block:added["+ REQ-AUTH-003, auth, mfa, TOTP-based MFA, ..., complete, test_totp_flow"]
        columns 1
    end
    block:context["  REQ-AUTH-004, auth, session, Session timeout, ..., partial"]
        columns 1
    end

    style header fill:#6b7280,stroke:#4b5563,color:#ffffff
    style removed fill:#dc2626,stroke:#b91c1c,color:#ffffff
    style added fill:#16a34a,stroke:#15803d,color:#ffffff
    style context fill:#9ca3af,stroke:#6b7280,color:#ffffff

• Human-readable diffs in PRs -- one row changed, one requirement done
• Works everywhere -- offline, air-gapped, no database, no API
• AI agents parse it without an SDK -- it's a file
• git blame tells you when and why every requirement changed
• No vendor lock-in -- it's your data in your repo

MCP Integration

flowchart LR
    subgraph agents["AI Agents"]
        direction TB
        A1["Claude Code"]
        A2["Cursor"]
        A3["Custom Agent"]
    end

    subgraph mcp["rtmx mcp-server"]
        direction TB
        T1["status"]
        T2["backlog"]
        T3["next"]
        T4["verify"]
        T5["health"]
        T6["markers"]
        T7["deps"]
        T1 ~~~ T2 ~~~ T3 ~~~ T4 ~~~ T5 ~~~ T6 ~~~ T7
    end

    subgraph repo["Git Repository"]
        direction TB
        DB[".rtmx/database.csv"]
        Tests["Test files"]
    end

    A1 -- "JSON-RPC" --> mcp
    A2 -- "JSON-RPC" --> mcp
    A3 -- "JSON-RPC" --> mcp
    mcp --> DB
    mcp --> Tests

    style agents fill:#f0fdf4,stroke:#059669,color:#065f46
    style mcp fill:#d1fae5,stroke:#10b981,color:#065f46
    style repo fill:#f3f4f6,stroke:#6b7280,color:#111827
    style A1 fill:#f0fdf4,stroke:#059669,color:#065f46
    style A2 fill:#f0fdf4,stroke:#059669,color:#065f46
    style A3 fill:#f0fdf4,stroke:#059669,color:#065f46
    style DB fill:#d1fae5,stroke:#10b981,color:#065f46
    style Tests fill:#d1fae5,stroke:#10b981,color:#065f46

Spine tools (status/backlog/next/claim/…) plus a scientific cycle surface (loop_tick, decompose, hygiene, cycles, webs, context, delivery_check, trade_*). Prefer these over shelling arbitrary CLI. No generic run_cli tool. Mutations require agent_id.

Git-native requirements RTM for agents (next, claim, verify) — not another Jira/Linear MCP. Prefer marketplace / one-click install — the GIF is the agent calling tools, not a human pasting mcp.json. (Positioning)

Claude Code (first-party marketplace):

claude plugin marketplace add rtmx-ai/rtmx
claude plugin install rtmx@rtmx

Cursor (one-click MCP deeplink): Add RTMX in Cursor · cursor.directory/plugins/rtmx

Codex:

codex mcp add rtmx -- rtmx mcp-server --stdio

One-shot (marketplace text + offline emit): rtmx bootstrap agents — see docs/BOOTSTRAP_AGENTS.md.

Emit (Claude/Cursor/Codex + long-tail gemini/windsurf/continue/gastown): rtmx plugin emit --list · rtmx plugin emit --all — see docs/LONG_TAIL_EMIT.md

Fallback paste (.cursor/mcp.json / generic MCP clients):

{"mcpServers":{"rtmx":{"command":"rtmx","args":["mcp-server","--stdio"]}}}

Details: docs/CLAUDE_MARKETPLACE.md, docs/CURSOR_MARKETPLACE.md, docs/CODEX_MCP.md.

Test Verification (2.0.0)

rtmx verify auto-detects your test framework. No configuration needed.

Go, Python/pytest, Rust/Cargo, Node.js/npm, Java/Gradle, Java/Maven, Elixir/Mix, Swift, Dart, Ruby -- 10+ frameworks supported.

Breaking in 2.0.0: COMPLETE is earned only with a verifying attestation signature. A results file alone (rtmx verify --results) cannot set COMPLETE, and rtmx status counts COMPLETE only when the signature verifies. That is the verification moat — agents cannot impersonate the verifier. The broker daemon ships in the same package and enables via install hooks or a one-time elevate on first rtmx verify when missing.

Release notes: v2.0.0 · CHANGELOG.md

Dogfooding

RTMX manages its own requirements. 219 requirements across 24 phases, auto-verified in CI on every push. Run rtmx status in this repo to see it.

Read the backstory -- how and why this tool was built.

Benchmark Results

Independent testing by intent-bench measures whether providing structured requirements to coding agents improves implementation effectiveness. Results from N=5 runs per condition using Claude Sonnet 4:

Experiment	Complexity	Control	Treatment (RTMX)	Fisher p
url-shortener	Baseline (10 reqs)	100% completion	80% completion	1.000
task-manager	Standard (13 reqs, depth 5)	20% completion	80% completion	0.206

On the baseline task, the intent layer adds overhead without improving completion. On the standard task with complex dependency chains, completion increases from 20% to 80% and token variance drops (CV 0.85 to 0.46).

Full data, methodology, and reproduction instructions: intent-bench.github.io/intent-bench

Technical Details

• Single static binary -- Go, CGO_ENABLED=0, zero runtime dependencies
• Linux, macOS, Windows -- amd64 and arm64
• 2 external dependencies (Cobra + YAML parser)
• GPG-signed releases with SBOM
• Apache 2.0

Contributing

See CONTRIBUTING.md for development setup, building, and testing instructions.

Migrating from Python CLI

Deprecation Notice: The Python rtmx CLI is deprecated and reached end-of-life on 2026-09-25. Run pip uninstall rtmx and install the Go binary above.

See docs/MIGRATION.md for full migration steps.

Support

• Discussions: github.com/rtmx-ai/rtmx/discussions
• Documentation: rtmx.ai
• Issues: github.com/rtmx-ai/rtmx/issues
• Email: [email protected]