Lobster
github.com/openclaw/lobster- Category
- AI Agents
- Rank
- No. 613Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- openclaw
- GitHub
- 1.3k stars
- Latest release
- v2026.9.8
- Date
About
OpenClaw-native workflow shell — typed, local-first macro engine that turns skills and tools into composable pipelines and safe automations.
What it does
Lobster runs repeatable agent workflows made from shell commands, data transformations, model calls, conditions, and human checkpoints. Steps exchange structured output, while retries, timeouts, saved state, and resume tokens support interrupted work.
Why it's ranked here
Lobster addresses a real agent reliability problem with concrete controls instead of another free-form planner. Structured data flow, approval identity checks, schema-validated input, dry runs, and workflow graphs make consequential automation easier to inspect and resume.
What's good
Workflows can mix ordinary command-line programs with native model and configured-agent stages. Approval gates can require a specific person or someone other than the initiator. Input responses use JSON Schema validation. Graph output exposes data flow, conditions, and checkpoints before execution.
Tradeoffs
It requires Node 22 or newer and assumes comfort with shell commands, JSON, and workflow definitions. Argument substitution is raw, so unsafe text belongs in environment variables. Commands restart after resumed input and must remain idempotent until the response arrives. Pricing estimates depend on a static table or manual overrides.
How to use it well
Use it for teams turning repeated agent procedures into inspectable workflows, especially operations with model calls, command-line tools, approvals, and recoverable pauses. Prefer environment variables for untrusted arguments and graph workflows before running them. Bring a separate credential system because Lobster deliberately does not own authentication tokens.
Technical notes+
package.json defines an ESM npm package with three executables and public SDK, core, and GitHub recipe exports. src/cli.ts routes human and tool modes, workflow execution, graph rendering, dry runs, and resume handling. src/runtime.ts consumes arrays or iterables stage by stage and tracks suspension state. src/input_request.ts caps replay snapshots at 1,000 items, 1 MiB, and 100 history entries, while src/validation.ts memoizes Ajv validators by stable schema serialization. src/pipeline_resume_state.ts persists resumable approval or input state under UUID-based keys. src/token.ts serializes token payloads as base64url JSON; no signature or encryption appears in that file. src/shell.ts selects cmd.exe on Windows and /bin/sh elsewhere, with an environment override.
Observed
- License
- MIT
- Primary language
- TypeScript, compiled to JavaScript
- Package
- ES module npm package named @clawdbot/lobster
- Runtime requirement
- Node.js 22 or newer
- Interfaces
- Command-line executables plus public SDK and core library exports
- Workflow formats
- Lobster, YAML, YML, and JSON files
- Direct runtime dependencies
- Ajv and YAML
- Shell support
- Windows command shell and POSIX /bin/sh, with an override
Read from README.md, package.json, src/cli.ts, src/shell.ts, src/token.ts, src/parser.ts, src/resume.ts, src/runtime.ts, src/read_line.ts, src/validation.ts, src/input_request.ts, src/pipeline_resume_state.ts, src/sdk/index.ts, src/sdk/token.ts, src/core/index.ts.
What it can do
Create typed automation workflows
User-defined skills and tools → Executable workflow pipelines
Compose modular pipeline components
Individual skills or tools → Combined workflow sequences
Execute local-first macros
Macro definitions and trigger conditions → Automated task execution
Validate workflow safety
Automation scripts and pipelines → Safety verification results
Transform skills into reusable components
Raw skills or capabilities → Composable workflow modules
Process workflows through shell interface
Command-line instructions → Workflow execution results
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.