Shrimp Task Manager
Persistent planning, decomposition, execution, and verification workflows for AI-assisted software development.
least_privilege=0: The README explicitly recommends launching the client with --dangerously-skip-permissions and supplies no lower-privilege path. user_confirmation=1: A confirm mock appears in viewer test setup, but the shown server APIs for deletion and template replacement, plus continuous execution, do not demonstrate confirmation gates. data_flow_transparency=2: DATA_DIR, a user settings file, task files, and template environment overrides are identified, and tests expose local read/write paths; a complete flow, retention policy, and network-exposure description are absent. sensitive_data_handling=0: No guidance is shown for secrets, private data, log redaction, or access control. dependency_security=1: Dependencies and an npm-ci release workflow are disclosed, but ranges are permissive and the supplied evidence includes no lockfile, audit process, update policy, or vulnerability mitigation. external_effects=1: File creation, overwrite, deletion, HTTP APIs, and continuous task execution are visible, but effect previews, authorization boundaries, and authentication are weak or absent; the API also permits wildcard CORS. rollback=1: Automatic task-history backup and restoration are asserted in the README but not supported by the supplied implementation or tests; template overwrite and deletion have no demonstrated recovery path. source_attribution=1: Repository, issue channel, and creator links are present, but package author is blank, README attribution to cjo4m06 differs from the LICENSE copyright name siage, and community maintenance ownership is vague.
self_consistency=2: README, package configuration, and API tests generally agree on task management, templates, and viewer functions; strong claims such as never losing context and automatic backup/restore lack corresponding evidence. dependency_availability=2: Node 18+, package managers, build commands, and dependencies are stated, while publishing uses Node 20 and npm ci; no supplied lockfile or offline/compatibility policy earns a deduction. failure_messages=3: Tests cover missing fields, malformed JSON, wrong content type, missing resources, file permission/read/write failures, absent build output, and startup errors, asserting specific status codes and useful messages throughout the shown surface.
audience_and_scenarios=3: The project clearly targets MCP-based AI development clients and thoroughly illustrates feature development, debugging, research, project rules, task viewing, and multilingual audiences. capability_boundaries=2: The command table, prerequisites, optional GUI, and template modes outline the principal capabilities, but continuous-mode stopping rules, safety boundaries, and unsupported cases are unclear. trigger_precision=2: init, plan, execute, continuous, list, research, and reflect phrases are explicit and illustrated; natural-language ambiguity and full parameter semantics are not defined. environment_fit=2: Windows/WSL2, macOS/Linux, Claude Code, Cline, Claude Desktop, and environment-variable setup are covered; examples remain Claude-centric, and port binding, containers, permissions, and production deployment are incomplete.
information_architecture=3: The README is clearly organized around quick start, features, interfaces, documentation, scenarios, configuration, and contribution, with useful navigation and language links. install_notes=2: Prerequisites, cloning, installation, build, and MCP configuration are concrete; the permission-bypass example, manual path substitution, and lack of uninstall, upgrade, and troubleshooting guidance prevent full credit. naming_stability=2: Package, server, command, and template-function names are mostly consistent; PROMPT_LANGUAGE versus TEMPLATES_USE and the broader environment-variable scheme shown in tests introduce inconsistency. examples_and_faq=2: Installation configurations and three practical scenarios are provided with further documentation links, but there is no actual FAQ and few edge-case configuration examples. known_limitations=0: No consolidated limitations, risks, or unsupported scenarios are disclosed. license=3: Package metadata and the complete MIT license agree and include the relevant grant and disclaimer. versioning_changelog=1: Version 1.0.21 and a v*-tag publishing workflow exist, but no changelog, compatibility policy, or migration notes are supplied. maintenance_responsibility=1: Issues, Discussions, and contribution entry points exist and community maintenance is asserted, but accountable maintainers, response expectations, and a security-reporting path are unspecified.
output_usability=2: Structured tasks, dependencies, status, persisted files, a viewer, and template export are designed for direct development-workflow use, while tests show consumable JSON interfaces; no representative core-agent outputs or quality acceptance evidence are supplied. marginal_value=2: Planning, decomposition, dependency tracking, reflection, research, and cross-session state are combined into a coherent MCP workflow that adds value beyond a plain prompt, though the automation benefits remain largely descriptive. cost_benefit=1: Installation and local-operation requirements are understandable, but users must maintain Node, client configuration, a data directory, and optional separate web tooling; no performance, token, time, or maintenance measurements substantiate the acceleration claim.
claim_traceability=2: Package metadata, README configuration, release workflow, and API tests make several installation and interface claims traceable; core planning, reflection, dependency execution, and memory implementation files are not supplied. cross_source_corroboration=2: README, package.json, publishing configuration, and viewer tests partially corroborate build, server, and template capabilities; core agent behavior lacks both implementation and test corroboration. fact_inference_separation=1: Commands, configurations, and tested behavior are generally identifiable, but terms such as intelligent, automatic, never lose, and accelerate are presented without evidence labels or limitations, weakening separation of fact from inference.
- The README recommends --dangerously-skip-permissions; do not use that configuration on a real codebase without isolation, constrained filesystem permissions, and human approval gates.
- The shown web/API behavior uses wildcard CORS and demonstrates no authentication. If exposed beyond localhost, task-file and template read, overwrite, and deletion surfaces may be accessible.
- Do not treat automatic backup/restore or never losing context as verified guarantees; the supplied files show neither recovery implementation nor recovery tests.
- Templates and task content are written to local files and a user-directory settings file; verify path permissions, retention, and cleanup before handling confidential code or credentials.
- Dependency-security evidence is limited. Pin and inspect resolved dependencies, run vulnerability auditing, and review the actual MCP SDK, Express, and build-tool versions before deployment.
What does this agent do, and when should you use it?
Shrimp Task Manager is a self-hosted MCP server for persistent task management in AI-assisted software development. It turns natural-language requests into structured development tasks, tracks dependencies and progress, and supports iterative planning, execution, verification, and reflection. The core service runs on Node.js 18+ and connects to compatible clients through MCP, while task history is stored in a filesystem directory selected with DATA_DIR. The repository also includes a modern React Task Viewer and an optional lightweight Web GUI controlled by ENABLE_GUI. It is a practical fit for developers who need continuity across agent sessions and explicit workflows, provided they are willing to build, configure, and operate the server and its local data storage.
A user starts with “init project rules” to establish project standards, then submits a requirement through “plan task [description].” The server analyzes that requirement, decomposes it into atomic, testable tasks, and records relationships between them. “execute task [id]” runs a selected task, while “continuous mode” processes the task list sequentially; “list tasks” exposes current status and “reflect task [id]” reviews and improves an existing task. “research [topic]” starts Research Mode for systematic technology or solution exploration, whose findings can feed a later plan. Tasks, progress, and history persist under DATA_DIR across sessions. Users can interact through an MCP client, manage tasks with search, drag-and-drop, and multiple profiles in the React Task Viewer, or inspect them through the optional Web GUI.
- A Claude Code user building JWT authentication can create a dependency-aware plan and execute the implementation one task at a time.
- A maintainer investigating a memory leak can split diagnosis and remediation into tracked tasks, then process them sequentially with continuous mode.
- A team carrying a large migration across multiple sessions can persist task history and progress in DATA_DIR instead of rebuilding context each time.
- A developer choosing between React and Vue can run a research workflow first and turn the resulting analysis into a migration plan.
- An engineering lead seeking consistent AI-generated code can initialize Project Rules before planning and execution begin.
- A team that wants human oversight of agent work can deploy the Task Viewer for drag-and-drop management, live search, and multiple profiles.
What are this agent's strengths and limitations?
- Combines planning, decomposition, dependency tracking, execution, verification, and reflection in an explicit development workflow instead of a flat task list.
- Persists tasks and progress under DATA_DIR, directly addressing lost context and repeated work across client sessions.
- Provides an MCP interface, a React Task Viewer, and an optional lightweight Web GUI for both agent-driven and human task management.
- Research Mode, Project Rules, and the specialized Agent System bring investigation, coding standards, and role assignment into the same task framework.
- Requires Node.js 18+, a local build, absolute path configuration, and an operated data directory, making adoption heavier than a hosted task service.
- The complete setup and first-run path centers on Claude Code; Cline and Claude Desktop are mentioned with configuration locations but without equally detailed verification steps.
- Filesystem persistence under DATA_DIR leaves permissions, backup, migration, and concurrent access practices to the adopter.
- The supplied Claude Code launch command uses --dangerously-skip-permissions, which bypasses permission checks and warrants review before use on sensitive codebases.
- The supplied material gives no benchmarks or guarantees for performance, concurrency, recovery behavior, or consistency across MCP clients.
How do you install or deploy this agent?
Prerequisites are Node.js 18+, npm or yarn, and an MCP-compatible AI client. Clone, install, and build the server:
git clone https://github.com/cjo4m06/mcp-shrimp-task-manager.git
cd mcp-shrimp-task-manager
npm install
npm run buildTo install Claude Code, run:
npm install -g @anthropic-ai/claude-codeCreate .mcp.json in the consuming project. Set command to node, point args at the absolute path of dist/index.js, and provide at least a DATA_DIR through env:
{
"mcpServers": {
"shrimp-task-manager": {
"command": "node",
"args": ["/path/to/mcp-shrimp-task-manager/dist/index.js"],
"env": {
"DATA_DIR": "/path/to/your/shrimp_data",
"TEMPLATES_USE": "en",
"ENABLE_GUI": "false"
}
}
}
}
No API credential is documented for Shrimp itself. The provided first launch command is:
claude --dangerously-skip-permissions --mcp-config .mcp.jsonBecause this flag bypasses Claude Code permission checks, adopters should assess its security implications before using it with sensitive repositories.
How do you use this agent?
After connecting the server to an MCP client, send “init project rules” to initialize project standards. Send “plan task: implement user authentication” to create a plan, inspect it with “list tasks,” and run work through “execute task [id]” or “continuous mode.” Use “research: compare React vs Vue for this project” for a structured investigation and “reflect task [id]” to review and refine a task. To enable the lightweight Web GUI, set ENABLE_GUI=true and optionally set WEB_PORT=3000. To run the React Task Viewer:
cd tools/task-viewer
npm install
npm run start:allThen open http://localhost:5173.