# slang-workflows [Health: Active] **Category:** πŸ’» Developer Tools **Repository:** https://github.com/shofer-dev/claude-code-slang-orchestrator **GitHub Stars:** 1 **npm Downloads (last month):** 153 **Views:** 0 **Installs:** 0 **Upvotes:** 0 **Directory Page:** https://allmcps.com/mcp/slang-workflows ## Description Provable, declarative multi-agent workflows for Claude Code (typed .slang, non-LLM executor). ## Tools Capabilities this server exposes over MCP: - **list_workflows** β€” Discover `.slang` files (name, title, params, agent count). - **get_slang_grammar** β€” Concise grammar cheatsheet + example, so an LLM can **generate** a workflow to run inline. - **validate_workflow** β€” Parse + static analysis (deadlocks, unknown refs, orphan outputs) without running. Accepts a `name`/`path` **or** inline `source`. - **run_workflow** β€” Run a workflow by `name`/`path` **or** inline `source` (rejects parse/static-analysis errors first). Synchronous by default; `background:true` returns a `workflow_id` immediately to poll live. - **get_workflow_state** β€” Serialized `FlowState` (per-agent status, round, budget) by `workflow_id` β€” live during a `background` run. - **get_topology** β€” Run topology as a Mermaid **flowchart** (status-colored snapshot) by `workflow_id`. - **get_trace** β€” Execution **trace** as a Mermaid **sequenceDiagram** + raw event log (who staked/routed to whom, commits, escalations, terminal) by `workflow_id`. ## Claude Desktop Quick Installation Remote MCP endpoint (confidence: high). Install path detected from listing signals. Add as a URL/SSE server in your client: ```json "mcpServers": { "slang-workflows": { "url": "https://shofer.dev" } } ``` ## Documentation & README ``` β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•—β–ˆβ–ˆβ•— β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•— β–ˆβ–ˆβ–ˆβ•— β–ˆβ–ˆβ•— β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•— β–ˆβ–ˆβ•”β•β•β•β•β•β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ•”β•β•β–ˆβ–ˆβ•—β–ˆβ–ˆβ–ˆβ–ˆβ•— β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•”β•β•β•β•β• β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•—β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•‘β–ˆβ–ˆβ•”β–ˆβ–ˆβ•— β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ–ˆβ•— β•šβ•β•β•β•β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ•”β•β•β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•‘β•šβ–ˆβ–ˆβ•—β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•‘β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•—β–ˆβ–ˆβ•‘ β–ˆβ–ˆβ•‘β–ˆβ–ˆβ•‘ β•šβ–ˆβ–ˆβ–ˆβ–ˆβ•‘β•šβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ•”β• β•šβ•β•β•β•β•β•β•β•šβ•β•β•β•β•β•β•β•šβ•β• β•šβ•β•β•šβ•β• β•šβ•β•β•β• β•šβ•β•β•β•β•β• provable, declarative multi-agent workflows for Claude Code a non-LLM executor runs your typed .slang files ``` # slang-workflows Run **provable**, `.slang`-driven **multi-agent workflows** inside Claude Code. A non-LLM state machine (the *slang executor*) runs inside an MCP server and coordinates agents; each agent is a Claude **Agent SDK** session. You declare the collaboration in a typed `.slang` file and the executor *enforces* it β€” typed output contracts, static analysis, tool-scoping, and provable termination β€” instead of leaving coordination to the model. The top-level Claude Code session only *triggers* and *observes*; it never makes a coordination decision. Claude Code's native **dynamic workflows** also codify orchestration (Claude writes a JS script); slang's difference is that the structure is **enforced and statically analyzable**, and runs render as **Mermaid topology + trace diagrams**. See the [benchmark](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/benchmark/) for the A/B/C head-to-head. > **Origin.** The `.slang` Workflow engine was born in [**Shofer**](https://shofer.dev) β€” the > AI-agent VS Code extension β€” where the deterministic, non-LLM executor was first built. This > plugin brings it to Claude Code. > Design and rationale: [`DESIGN.md`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/DESIGN.md). Language reference: [`slang_specs.md`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/slang_specs.md). Privacy: [`PRIVACY.md`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/PRIVACY.md). ![slang-workflows running implement-feature.slang deterministically β€” the guaranteed review loop, 0 coordination LLM tokens, converged](https://raw.githubusercontent.com/shofer-dev/claude-code-slang-orchestrator/HEAD/media/slang-workflows.gif) ## How it works ```mermaid flowchart TB You(["You"]) CC["Claude Code
top-level session"] subgraph server["slang-workflows Β· MCP server"] direction TB SL["your .slang workflow
agents Β· stakes Β· contracts Β· converge"] EX["slang executor
deterministic Β· non-LLM state machine"] MB[("FlowState + mailbox")] SL -->|"parsed once"| EX EX <--> MB end subgraph Agents["Agent SDK sessions β€” the agents your .slang declares"] direction LR A1["Agent A"] A2["Agent B"] A3["Agent …"] end You <-->|"ask Β· approve escalations"| CC CC -->|"run_workflow"| EX EX -->|"topology + trace diagrams"| CC EX ==>|"dispatch stake β€” one session per agent, resumed"| Agents Agents ==>|"typed result β†’ output-contract check + retry"| EX EX -.->|"escalate @Human"| CC classDef brain fill:#eaeaff,stroke:#5b5bd6,stroke-width:2px; class EX brain; ``` The top-level Claude Code session only **triggers and observes** β€” it never makes a coordination decision. The non-LLM **executor** reads your `.slang` file, dispatches each *stake* to an Agent SDK session (one long-lived session per agent, resumed across rounds), checks every result against its **output contract** (retrying on failure), routes it through the **mailbox**, and repeats each round until the workflow's `converge` condition or round budget is met. Runs render as Mermaid **topology** and **trace** diagrams; `escalate @Human` surfaces back to you in the normal chat. ## Use cases Concrete workflows (each is a `.slang` file you run with `run_workflow`): - **Implement a feature with a human design-approval gate** (`implement-feature.slang`) β€” an *Architect* decomposes your request and writes the design doc, but is scoped so it **physically cannot write code** (`write_paths: ["**/*.md"]`, `deny: [Bash]`) and must delegate. You approve the design (`escalate @Human`), then a *Developer* implements it slice-by-slice while a *Reviewer* signs off each round, with a final review gate before it commits. - **Ship a feature as separate, verified deliverables** (`implement-feature-complex.slang`) β€” a linear Design β†’ Implement β†’ Test β†’ Review β†’ Document pipeline where five specialists each produce **one** artifact (the implementation, a *passing* vitest spec, usage docs) and can't do another's job β€” so code, tests, and docs actually match the design. Converges only when every stage has committed. - **Troubleshoot a bug with two independent investigators** (`debug.slang`) β€” paste the symptom/repro; two developers root-cause it **in parallel, strictly read-only** (no accidental edits), an *Orchestrator* consolidates their independent findings into one fix plan, one developer implements it, and the other **peer-reviews the fix in a loop** until satisfied. - **A reviewer that can inspect but never edit** β€” give the review agent `read`/`execute` and no write scope, so a "check my work" run can run tests and read code but cannot alter it β€” enforced, not just prompted. ## Features - **Reproducible multi-agent pipelines** β€” the collaboration is codified in a `.slang` file and driven by a deterministic (non-LLM) executor, so a run unfolds the same way every time β€” no improvised, unrepeatable subagent coordination. - **Enforced per-agent tool-scoping** β€” `write_paths` restricts each agent's Write/Edit to path globs and `deny` removes tools (e.g. `Bash`), enforced via the SDK's `canUseTool` β€” not merely requested in a prompt. - **Static analysis before running** β€” `validate_workflow` detects deadlocks, unknown references, and orphaned outputs at parse time, before any tokens are spent. - **Typed output contracts** β€” each stake must return a structurally *and* semantically valid result (`output: {…} where `); invalid results retry instead of silently propagating downstream. - **Provable termination** β€” round budgets (`budget: rounds(N)`) + per-stake timeouts guarantee every run finishes. - **Auto-generated diagrams** β€” every run renders a Mermaid topology (`get_topology`) and a sequence-diagram trace (`get_trace`), for live or post-mortem inspection. - **Convergence-driven collaboration** β€” agents route via mailboxes and iterate until a declared convergence condition or budget is reached, with session resume so an agent keeps its context across rounds. ## How this differs from Claude Code's native *dynamic workflows* Claude Code already has a built-in **dynamic workflows** feature: when a task needs orchestration, Claude writes a **JavaScript script** (via the Agent SDK / `Workflow` tool) that spawns and coordinates subagents. That's a real step up from improvised, one-off subagent calls β€” the script codifies the orchestration and, once written, runs deterministically and coordinates for ~0 extra LLM cost. slang shares those goals; our [benchmark](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/benchmark/) shows **both** approaches reach real, working implementations with near-zero coordination-LLM cost. **The difference is what the orchestration *is*, and what's *guaranteed* about it:** | | Native dynamic workflows | slang-workflows | |---|---|---| | The orchestration is… | an **LLM-authored JavaScript** script | a **typed, declarative `.slang`** file run by a fixed non-LLM executor | | Who wrote the coordination logic | Claude, per task, in a general-purpose language | you (or an LLM, once) in a domain-specific language the runtime understands | | Output contracts between stages | whatever the script happens to check | **enforced** by the runtime β€” structural + semantic (`output: {…} where `); invalid β†’ retry | | Per-agent tool scoping | up to the script | **enforced** β€” `write_paths` / `deny` via the SDK's `canUseTool` | | Correctness of the structure | nothing checks the JS | **static analysis before running** β€” deadlock / unknown-ref / orphan-output | | Termination | up to the script | **provable** β€” `budget: rounds(N)` + per-stake timeouts | | Observability | instrument it yourself | **auto-generated** Mermaid topology + sequence-diagram trace | | The reusable artifact | a script the model regenerates each time | a **versioned `.slang` file** + a fixed interpreter | In short: native dynamic workflows put the orchestration in **LLM-written code you have to trust**; slang puts it in a **typed declaration the runtime validates and enforces** β€” analyzable *before* it runs, scoped and contract-checked *while* it runs, and rendered as diagrams *after*. Reach for slang when you want **guarantees and auditability** (safety-scoped agents, provable termination, contract-valid hand-offs, a reusable versioned workflow), not just "the model coordinated some subagents this time." Both are far better than unstructured subagents β€” slang trades a bit of up-front declaration for enforcement and repeatability. ## What works today - Discover / validate / run `.slang` workflows β€” **authored or LLM-generated inline** (MCP tools below). - Deterministic executor: stake β†’ **output-contract** validation + retry β†’ mailbox routing β†’ convergence; **multi-agent** flows; **session resume** (one agent = one session across stakes); `escalate @Human`. - **Output contracts**: structural (`output: {...}`, via SDK `outputFormat`) + semantic (`where `). - **Enforced tool scoping**: `write_paths` (Write/Edit restricted to path globs, via a PreToolUse command hook) and `deny` (remove native or MCP tools). - **Static analysis** (`validate_workflow`): deadlock / unknown-ref / orphan-output detection before running. - **Provable termination**: `budget: rounds(N)` + per-stake timeouts β€” the run always finishes. - **Diagrams**: `get_topology` (Mermaid flowchart) + `get_trace` (Mermaid sequence + event log). - **Synchronous or background** runs (`background:true`) with live polling of state/topology/trace. See [`DESIGN.md` Β§ Implementation Status](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/DESIGN.md#implementation-status) for the full matrix. ## Requirements - **Node.js 22+** and **pnpm**. - **Claude Code** installed and authenticated (the `claude` CLI on your `PATH`) β€” the Agent SDK spawns it to run agents. ## Install ```bash cd server pnpm install ``` > **Agent SDK note.** `@anthropic-ai/claude-agent-sdk` is declared as an *optional* dependency > because some internal registries don't mirror it. If `pnpm install` skips it, add it from the > public registry: > > ```bash > pnpm add @anthropic-ai/claude-agent-sdk --registry=https://registry.npmjs.org/ > ``` > > The server *parses/validates* workflows without it; *running* agents requires it. Verify the install: ```bash pnpm typecheck # type-checks the whole server pnpm test # runs the unit suite (mock-based, no model calls) ``` ## Use it At runtime the server discovers `.slang` files in **your project's** `.claude/workflows/` (and `~/.claude/workflows/`) β€” that's the user's space for their own workflows. The plugin ships showcase workflows in [`server/test/fixtures/`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/server/test/fixtures); copy them in to try them: ```bash mkdir -p .claude/workflows cp /PATH/TO/slang-orchestrator/server/test/fixtures/*.slang .claude/workflows/ ``` ### Quickest: register the MCP server with Claude Code From the project whose `.claude/workflows/` you want to run: ```bash claude mcp add slang-workflows -- npx tsx /ABSOLUTE/PATH/TO/slang-orchestrator/server/src/main.ts ``` Then in Claude Code, ask it to use the tools β€” e.g. *"list the slang workflows"*, *"run the where-clause workflow"*. ### As a Claude Code plugin This directory is a self-contained plugin: [`.claude-plugin/plugin.json`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/.claude-plugin/plugin.json) + [`.mcp.json`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/.mcp.json) (a stdio server launched as `npx tsx ${CLAUDE_PLUGIN_ROOT}/server/src/main.ts`). Install it through Claude Code's plugin mechanism to expose the tools automatically. ### Slash commands Once the plugin is loaded, these commands drive it directly (Claude Code namespaces them by the plugin, so your `/` menu shows `slang-workflows:…`): | Command | Does | |---------|------| | `/slang-workflows:slang-run [k=v …]` | Run a workflow to completion β€” synchronous, so `@Human` gates prompt you. | | `/slang-workflows:slang-list` | List the `.slang` workflows it can see. | | `/slang-workflows:slang-new ` | Generate a workflow from a description, validate it, and run it. | | `/slang-workflows:slang-trace [workflow_id]` | Render the topology + sequence-diagram trace of a run. | They're thin wrappers over the MCP tools below β€” you can always just ask Claude in natural language instead (e.g. *"run the where-clause workflow"*). ## MCP tools | Tool | Purpose | |------|---------| | `list_workflows` | Discover `.slang` files (name, title, params, agent count). | | `get_slang_grammar` | Concise grammar cheatsheet + example, so an LLM can **generate** a workflow to run inline. | | `validate_workflow` | Parse + static analysis (deadlocks, unknown refs, orphan outputs) without running. Accepts a `name`/`path` **or** inline `source`. | | `run_workflow` | Run a workflow by `name`/`path` **or** inline `source` (rejects parse/static-analysis errors first). Synchronous by default; `background:true` returns a `workflow_id` immediately to poll live. | | `get_workflow_state` | Serialized `FlowState` (per-agent status, round, budget) by `workflow_id` β€” live during a `background` run. | | `get_topology` | Run topology as a Mermaid **flowchart** (status-colored snapshot) by `workflow_id`. | | `get_trace` | Execution **trace** as a Mermaid **sequenceDiagram** + raw event log (who staked/routed to whom, commits, escalations, terminal) by `workflow_id`. | **Generate-and-run loop:** `get_slang_grammar` β†’ author slang β†’ `validate_workflow{source}` β†’ `run_workflow{source, background:true}` β†’ poll `get_topology` / `get_trace` while it runs. The workflow is authored by an LLM but **executed deterministically** (contracts enforced, always terminates). `@Human` escalation works in synchronous runs only (interactive elicitation needs the tool call to stay open). ## Develop ```bash pnpm dev # run the server over stdio (logs to stderr) pnpm typecheck # tsc --noEmit pnpm test # node:test suite ``` The executor depends only on a `Dispatcher` interface; `FakeDispatcher` makes the whole VM testable without the Agent SDK or any model calls (see [`server/test/`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/server/test)). ## Layout ``` .claude-plugin/plugin.json plugin manifest .mcp.json stdio MCP server declaration server/ src/ main.ts MCP server + tool surface constants.ts tunable defaults in one place (round cap, retry budget, stake timeout, loop guard) executor.ts deterministic round loop + output contracts dispatcher.ts Dispatcher interface + FakeDispatcher (the agent-runtime seam) agent-sdk-dispatcher.ts production backend (Claude Agent SDK) tool-group-map.ts slang tool-groups β†’ Claude Code tools workflows.ts .slang discovery + validation slang/ vendored, framework-agnostic slang VM (lexer/parser/interpreter/…) test/ node:test unit + conformance suite fixtures/ showcase .slang workflows (also the conformance fixtures) ``` ## License Apache-2.0. See [`LICENSE`](https://github.com/shofer-dev/claude-code-slang-orchestrator/blob/HEAD/LICENSE).