>_ scream / terminal-native

Scream
Code

Scream Code combines unlimited concurrency, per-agent model orchestration, memory-stable goal loops, an independent referee, and multi-hop knowledge graphs in an enterprise-secure terminal workspace.

Install Scream Code Explore capabilities
Open sourcemacOS / Linux / WindowsA model per AgentUnlimited concurrency
npm install -g scream-code
✓ Local-first
✓ Private
✓ Ready to run
Goal Loop / referee loop
  • understand
  • execute
  • judge
  • iterate
Agent DAG / orchestrate
  • explore → plan
  • coder → verify
  • reviewer → done
SAG / retrieval
  • query → vector
  • rank → memory
Wolfpack / parallel
  • agent_01 RUN
  • agent_02 DONE
  • agent_N ∞
Let the
Agent Run.
Live terminal previewLOCAL / READY
scream-code / session
WOLFPACK / RUNNING0%
tool belt / capability legend

A complete terminal Agent tool belt

From natural-language goals to multi-model orchestration, parallel execution, long-loop judging, knowledge graphs, persistent sessions, Skills, and MCP extensions.

Natural language
Built-in tools
7 agent profiles
Per-agent models
Wolfpack concurrency
Goal Loop referee
Persistent sessions
SAG multi-hop graph
Skills & plugins
MCP extensions
01Unlimited concurrent execution

Wolfpack

Split one goal into work units that move at the same time. Each Agent runs independently, then converges on one execution line.

∞
Concurrency ceiling
  • Parallel launch with no preset limit
  • Dispatch any of 7 profiles
  • Centralized status and result reporting
Task matrix● Live completion status
Scan dependenciesDONE
agent_01 / complete
Refactor authDONE
agent_02 / complete
Run testsDONE
agent_03 / complete
Draft docsRUN
agent_04 / working
Security reviewRUN
agent_05 / working
Aggregate resultsRUN
agent_06 / working
Observable Profile × Model routingDAG / depth 3
Observable Profile × Model routingChoose a model independently for every profile. Dependencies and context pass automatically. Specialized models handle the work they do bestinputexploreplancoderverifyreview
Profile × ModelEvery sub-agent binds an independent model alias

Specialized models handle the work they do best

02Multi-agent · multi-model orchestration

Agent Orchestration

Express dependencies as a DAG and configure a model for every sub-agent. Let coding models implement, reasoning models plan and review, and long-context models organize knowledge—each model does the work it is best at.

7
Specialist profiles
  • Choose a model independently for every profile
  • Dependencies and context pass automatically
  • Specialized models handle the work they do best
goal://activeIndependent referee: not complete; inject notes and continue
1
SET GOAL
2
INJECT MEMORY
3
EXECUTE
4
REFEREE
14:02:11memory injection: goal, constraints, 10 working notes
14:02:16executor: continue tests, security audit, and result collection
14:02:24independent referee: evidence incomplete, continue with injected notes
14:02:38result converged: acceptance criteria passed, deliver verified outcome
03Memory injection · independent referee

Goal Loop

Every round injects key notes back into memory so long loops do not drift. Execution and judgment are separated: an independent referee decides whether to iterate or converge until a verifiable result is delivered.

Stable
Long-loop context
  • Inject working notes and critical facts every round
  • Keep goals, constraints, and context stable in long loops
  • An independent referee accepts, continues, or stops
  • Deliver a verifiable result instead of an intermediate answer
04Multi-hop enhanced knowledge graph

SAG Knowledge

A next-generation graph knowledge system with enhanced multi-hop retrieval. It follows entities, relationships, and prior experience to assemble actionable context—a knowledge base designed for Agents, not just document search.

  • Entity and relationship modeling
  • Multi-hop association and context assembly
  • Cross-session memory and experience reuse
  • Built for Agent decisions and actions
sag://queryMulti-hop Retrieval mode
›Why did the auth system trigger a permission error?READY
queryentity graphmulti-hopagent context

The query expands across multiple graph hops, then assembles context the Agent can use directly.

Multi-hop related knowledge

TOP 3
Selected: Auth → Session → Policy. Continue across graph relationships, then assemble actionable Agent context.
Scream Code / CLI Agent

A terminal Agent built to deliver results

Scream Code—also searched as scream-code, screamcode, or screamAgent—is an open-source CLI Agent for developers and enterprise teams. It combines task decomposition, multi-agent and multi-model orchestration, unlimited concurrency, long-loop memory injection, an independent referee, knowledge graphs, and enterprise permission controls into one execution runtime.

Like Claude Code and Codex, Scream Code works in the terminal developers already use. Its focus is different: coordinated specialist Agents, an independently selected model for every sub-agent, persistent goal loops, observable execution, and a multi-hop knowledge system built for Agent action.

About Scream Code

FAQ / EN
01What is Scream Code?

Scream Code is an open-source, cross-platform CLI terminal AI Agent with Wolfpack concurrency, multi-agent and multi-model orchestration, an independent Goal Loop referee, SAG multi-hop knowledge graphs, MCP, Skills, and enterprise permission modes.

02Can sub-agents use different models?

Yes. Profiles such as coder, explore, plan, verify, reviewer, oracle, and writer can bind different models or aliases so coding, reasoning, retrieval, and long-context models each handle the work they do best.

03Why is Goal Loop suitable for long-running tasks?

Every round injects working notes and critical facts back into memory to reduce context drift. An independent referee evaluates the result and decides whether to continue or converge until the result is verifiable.

04How is SAG different from a basic vector database?

SAG combines vector retrieval with graph entities and relationships, enabling multi-hop associations and returning assembled context an Agent can directly use for decisions and actions.

05Is Scream Code suitable for enterprise security?

Scream Code supports local-first execution, permission modes, tool approvals, and observable task state so teams can apply their own models, services, and security policies.

Start with one goal

Put the next complex task in your terminal.

Launch Scream Code in three steps and keep the shell workflow you already know.

npm install -g scream-code

Start in 3 steps ↓

  1. 01Install the CLI

    Install on this machine

    npm install -g scream-code
  2. 02Launch the TUI

    Open the interactive terminal

    scream
  3. 03Enter a goal

    Start execution and collaboration

    > /goal your goal