ClawCode

Your creative dev tool , AI coding Swiss Army knife

English | 简体中文

Philosophy • Product vision • Development assistance functions • System Architecture • Value at a glance • Get started

ClawCode is a Claude Code‑inspired implementation in Python and Rust, focused on agents and experience‑based evolution. It is also an open‑source coding‑agent CLI for Anthropic, OpenAI, Gemini, DeepSeek, GLM, Kimi, Ollama, Codex, GitHub Models, and 200+ models via OpenAI‑compatible APIs.

Philosophy

We aim to build an open, excellent AI coding Swiss Army knife.

It starts with a coder agent framework (tool use, skills, memory, multi‑agent). On this we built Claw, adding more tools, OpenClaw skill compatibility, and computer‑use abilities.

But that's not enough. Our own projects generate valuable development data through iterations. We shouldn't waste them.

So we built a self‑improving subsystem: it uses that data to continuously enhance the agents. All data stays local, under your control. The system is open source, auditable, with no hidden telemetry.

In short: an experience‑based reinforcement learning framework that turns

Idea → Memory → Plan → Code → Verify → Review → Learned Experience

into an executable, learnable, evolving engineering loop.

---

Product vision

ClawCode is a creative dev tool aimed at real delivery.

Features:

Core motivations:

• Turn ideas into runnable code quickly
  From “I have an idea” to “implemented and verified,” with less context switching and tool friction.

• Stay vendor- and model-agnostic
  Configurable providers/models and open extension paths reduce lock-in.

• Reuse strong UX patterns instead of reinventing habits
  Learn from mature tools (e.g. Claude Code, Cursor) and preserve familiar workflows where possible.

• Remember usage and improve over time
  Session persistence, experience write-back, and closed-loop learning let the system evolve with tasks and team practice.

• Execute “full-stack” engineering tasks end-to-end
  Beyond one-off codegen: planning, delegation, execution, verification, review, and structured learning.

Design

Enable the model to gain a deeper understanding of product design aesthetics.

The demo and baseline now adopt the Google Stitch open-source standard. And The demo provides 55 open-source real-world brand case studies. The coder workflow extensively references them during UI design and coding; you are also free to use them selectively.

Path	Purpose
UI/	Curated product / brand UI references — each brand folder includes DESIGN.md, previews, and notes. See UI/README.md for the full catalog and links.

Development assistance functions

Architecture & review

Command	Function
/architect	Run architecture design/review workflow with trade-off analysis and ADR/checklist options.
/code-review	Review local uncommitted changes with severity-ranked findings and commit gate.
/security-review	Complete a security review of the pending changes on the current branch.
/review	Review a pull request.

Autocomplete / related planning & plugins

Command	Function
/plan	Enable plan mode or view the current session plan (handled in ChatScreen; listed for autocomplete).
/arc-plan	Generate a one-shot alternative implementation plan (ARC planner).
/plugin	Manage clawcode plugins.

Test-driven development

Command	Function
/tdd	Run strict TDD workflow: scaffold, RED, GREEN, refactor, and coverage gate.

Multi-role orchestration & engineering workflows

Command	Function
/clawteam	Run multi-role task orchestration, or target one role via /clawteam:<agent>. Supports --deep_loop for iterative convergence (see docs/CLAWTEAM_SLASH_GUIDE.md).
/clawteam-deeploop-finalize	Parse DEEP_LOOP_WRITEBACK_JSON from pasted or last assistant text using pending deep-loop session metadata.
/multi-plan	Run multi-model collaborative planning workflow (plan-only).
/multi-execute	Run multi-model collaborative execution workflow with traceable artifacts.
/multi-backend	Run backend-focused multi-model workflow (research through review, orchestrator writes code).
/multi-frontend	Run frontend-focused multi-model workflow (UI/UX led, orchestrator writes code).
/multi-workflow	Run full-stack multi-model workflow (backend + UI advisors, orchestrator writes code).
/orchestrate	Run sequential multi-role workflow (HANDOFF between planner/TDD/review/security/architect); /orchestrate show|list.

Learning loop: ECAP, TECAP & instincts

Command	Function
/learn	Learn reusable instincts from recent tool observations.
/learn-orchestrate	Run observe → evolve → import-to-skill-store orchestration in one command.
/experience-create	Create an ECAP experience capsule from recent observations/instincts.
/experience-status	List available ECAP capsules with optional filters.
/experience-export	Export an ECAP capsule as JSON/Markdown for model or human use.
/experience-import	Import an ECAP capsule from local file or URL.
/experience-apply	Apply an ECAP capsule as one-shot agent prompt context.
/experience-feedback	Record success/failure feedback score for an ECAP capsule.
/team-experience-create	Create a TECAP team-experience capsule from collaborative traces.
/team-experience-status	List TECAP capsules with optional team/problem filters.
/team-experience-export	Export a TECAP capsule as JSON/Markdown for agents and humans.
/team-experience-import	Import a TECAP capsule from local file or URL.
/team-experience-apply	Apply a TECAP capsule as collaboration context prompt.
/team-experience-feedback	Record feedback score for a TECAP capsule.
/tecap-create	Alias of /team-experience-create.
/tecap-status	Alias of /team-experience-status.
/tecap-export	Alias of /team-experience-export.
/tecap-import	Alias of /team-experience-import.
/tecap-apply	Alias of /team-experience-apply.
/tecap-feedback	Alias of /team-experience-feedback.
/instinct-status	Show learned instincts grouped by domain and confidence.
/instinct-import	Import instincts from local file or URL into inherited set.
/instinct-export	Export instincts with optional domain/confidence filters.
/evolve	Cluster instincts and optionally generate evolved structures.
/experience-dashboard	Show ECAP-first experience metrics dashboard (add --json or --no-alerts).
/closed-loop-contract	Show closed-loop config contract coverage (consumed vs unconsumed keys).

Observability, diagnostics & diffs

Command	Function
/doctor	Diagnose and verify your clawcode installation and settings.
/diff	View uncommitted changes and per-turn diffs.
/debug	Debug your current clawcode session via logs (bundled viewer).
/insights	Generate a report analyzing your clawcode sessions.

Claw mode & external CLIs

Command	Function
/claw	Enable Claw agent mode or show status (autocomplete entry; see SLASH_AUTOCOMPLETE_EXTRA in builtin_slash.py).
/claude	Enable Claw mode then path A: Anthropic + Claude Code HTTP identity.
/claude-cli	Enable Claw mode then path B: run claude / claude-code CLI in workspace.
/opencode-cli	Enable Claw mode then path B′: run OpenCode opencode CLI in workspace.
/codex-cli	Enable Claw mode then path B″: run OpenAI Codex CLI in workspace.

Session, Git, workspace & engineering context

Command	Function
/checkpoint	Git workflow checkpoints: create, verify, list, clear.
/rewind	Soft-archive chat after a message, or inspect/restore tracked git files.
/tasks	List and manage background tasks (plan build, agent run).
/init	Initialize CLAWCODE.md (or CLAW.md-style) codebase documentation in the project.
/add-dir	Add a new working directory.
/agents	Manage agent configurations.
/skills	List available skills.
/mcp	Manage MCP servers.
/hooks	Manage hook configurations for tool events.
/permissions	Manage allow & deny tool permission rules.
/memory	Edit claw memory files.
/pr-comments	Get comments from a GitHub pull request.

System Architecture

Compatible with Claude Code (config & plugins & skills)

To lower learning and migration cost, ClawCode offers alignable workflows where it matters.

• If you want polished product UX out of the box, Claude Code has strengths.
• If you want deep terminal execution + team orchestration + learning loops + configurable extensions, ClawCode emphasizes that combination.

ClawCode uses alignment as a migration layer and closed-loop engineering evolution as the core value layer.

Alignment	What it means	Extra value in ClawCode
Slash workflows	Organize work with / commands (e.g. /clawteam, /clawteam --deep_loop)	Goes from “command fired” to multi-role orchestration + convergence + write-back
Skills	Reuse and extend skills; lower asset migration cost	Skills can plug into experience loops and improve per project
Terminal-native	TUI/CLI habits and scripting	Analyze, execute, verify, and review in one surface
Extensible tools	plugin / MCP / computer use	Progressive capability expansion under team policy

Compatible with Claude Code (Agent & Subagent)

Agent roles & team (Claude Code–style)

ClawCode aligns with Claude Code’s Agent tool: the main agent spawns subagents with isolated context, custom prompts, and tool allowlists. Subagents cannot nest another Agent / Task (delegation is stripped for the inner run).

1. Built-in subagent roles

Agent id	Purpose (short)
explore	Read-only exploration (Read/Glob/Grep/…)
plan	Read-only research for planning
code-review	Review-focused, read-only tools
general-purpose	Full tool surface (minus delegate tools) when you omit a custom list

2. Built-in clawteam (multi-role “team”)

Use these ids as agent / subagent_type the same way as explore / plan. Examples: clawteam-system-architect, clawteam-rnd-backend, clawteam-qa, clawteam-product-manager, … (all clawteam-* ids ship in the built-in registry).

Collaboration model: the orchestrator is still the main agent—it calls Agent multiple times with different role ids and tasks. There is no separate “team scheduler” UI; “team work” is sequential/parallel tool calls decided by the model.

3. Custom roles (project + user)

Definitions are Markdown files with YAML frontmatter (same idea as .claude/agents/).

User-wide (highest override of project defaults for same name):

• ~/.claude/agents/*.md

Project (read merge order; later roots override earlier for the same name):

• .claw/agents/*.md
• .clawcode/agents/*.md
• .claude/agents/*.md

Frontmatter fields (common):

Field	Meaning
name	Agent id (default: filename stem)
description	Short blurb for routing / docs
tools	Optional allowlist using Claude-style names (Read, Write, Bash, …) — mapped to ClawCode tools
disallowedTools	Block list (same naming style)
model	Optional override (inherit, sonnet, opus, haiku, or full model id)
maxTurns	Cap on ReAct iterations for this subagent
isolation	e.g. none, worktree, fork
permissionMode, background, mcpServers, hooks	Passed through when set

Body markdown becomes the subagent system prompt.

Example — .claw/agents/api-guardian.md:

---
name: api-guardian
description: Reviews public HTTP API changes only.
tools:
  - Read
  - Glob
  - Grep
  - diagnostics
maxTurns: 24
---

You only analyze API routes and OpenAPI/contract files. Report breaking changes as a bullet list.

4. Invoking a subagent (Agent tool)

The model (or harness) calls Agent with at least a task and an agent id:

{
  "agent": "plan",
  "task": "Map how authentication is implemented; list key files."
}

Aliases:

• subagent_type ↔ agent
• prompt ↔ task
• Optional: context, timeout (seconds), max_iterations, isolation, allowed_tools (override allowlist)

Unknown agent → error listing known ids from the merged registry (built-ins + your files).

5. Plan mode (/plan)

Only these subagents are allowed: plan, explore, code-review (plus internal review alias where applicable). Their tools are further restricted to read-only policy (no write/exec tools even if the definition asked for them).

6. Optional: deep-loop / handoff settings (.clawcode.json)

Runtime tuning for multi-round clawteam-style loops lives under settings such as clawteam_deeploop_* (enable flag, max iterations, convergence, handoff target, etc.). See project docs/snippets for a full example—this does not replace defining roles; it shapes how long/how strictly the loop runs.

7. Relation to agents in .clawcode.json

The top-level agents map (coder, task, title, summarizer, …) configures which model/provider backs main flows. Subagent roles (explore, clawteam-*, custom *.md) are selected by the Agent tool and merged from the paths above—not by renaming those slots.

---

Full-stack task execution stack (AI Coding + Claw framework + tools + computer use)

“Full-stack” tasks are not a single codegen step—they chain planning, coding, verification, review, environment actions, and learning into one executable path. ClawCode implements three layers:

Layer	Role	Key components / commands	Typical tasks
Coder agent (default terminal runtime)	The default interactive path: Textual TUI wires sessions/messages, builds a coder runtime bundle from settings, and runs the main Agent loop (or ClawAgent.run_claw_turn when Claw mode is on). It owns send/finalize, HUD, plan state, and optional memory injection—without requiring /claw or workflow slash commands.	ChatScreen, _finalize_send_after_input, _start_agent_run, _process_message, build_coder_runtime, make_claw_agent / make_plain_agent, Agent.run, ClawAgent.run_claw_turn, _handle_agent_event, _rebuild_llm_stack	Everyday coding in the terminal: chat turns, file edits via tools, plan-style flows when /plan is active, model switch (e.g. Ctrl+O stack rebuild), non-Claw and Claw branches from the same screen
Claw framework (agent runtime)	In Claw mode, ClawAgent runs multi-step work aligned with the main agent loop, with iteration budget and sub-agent coordination	/claw, ClawAgent.run_claw_turn, run_agent / run_conversation	Phased complex tasks, cross-turn context, bounded multi-round execution
Tool orchestration (engineering execution)	Slash commands and tools drive plan-to-delivery flows: collaboration, review, diagnostics, learning	/clawteam, /architect, /tdd, /code-review, /orchestrate, /multi-*	Decompose requirements, implement, test, review, converge and write back
Computer use (OS-level)	With desktop.enabled, desktop_* tools provide screenshots, mouse, and keyboard automation; complements browser_*	desktop_screenshot, desktop_click, desktop_type, desktop_key, /doctor	Cross-app actions, desktop checks, GUI-assisted verification

desktop_* is off by default. Enable explicitly and install optional extras (e.g. pip install -e ".[desktop]" or equivalent). Prefer least privilege and a controlled environment.

That is why ClawCode combines terminal execution + team orchestration + experience evolution in one framework: a long-lived engineering partner—not a short Q&A toy.

---

Value at a glance

Dimension	Core capability	User value
Idea to delivery	Terminal-native execution + ReAct tool orchestration	Less switching; ideas become runnable results faster
Long-horizon work	Local persistent sessions + master–slave agents + decomposition	Multi-round complex tasks with handoff and review
Learning loop	deeploop + Experience + ECAP/TECAP	Not one-shot success—the system grows with your team

---

Who it’s for / Get started

Who it’s for

• Developers who live in the terminal and want AI to execute, not only suggest.
• Teams that need multi-role collaboration, governable flows, and reviewable outputs.
• Leads who care about long-term outcomes, not a single answer.

Get started

cd clawcode
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"
python -m clawcode -c "[your project dir]"

Refer to docs/clawcode-configuration.md for further configuration details.

---

Project highlights — beyond “just writing code”

1) Long-horizon projects: durable memory + continuous context

Sessions and messages persist locally—not a throwaway chat. Split complex work across rounds, keep decisions and history, and support handoff and postmortems.

Why it matters: Fits real long-cycle development, not one-off demos.

2) clawteam: a schedulable virtual R&D team

With /clawteam, the system can orchestrate roles and execution:

• Professional role segmentation and extraction of mental models from years of industry experience.
• Intelligent role pick and assignment
• Serial/parallel flow planning
• Per-role outputs and final integration
• 10+ professional roles (product, architecture, backend, frontend, QA, SRE, …)

clawteam roles (overview)

Role ID	Role	Responsibility & typical outputs
clawteam-product-manager	Product manager	Priorities, roadmap, value hypotheses; scope and acceptance criteria
clawteam-business-analyst	Business analyst	Process and rules; requirements, edge cases, business acceptance
clawteam-system-architect	System architect	Architecture and tech choices; modules, APIs, NFRs (performance, security, …)
clawteam-ui-ux-designer	UI/UX	IA and interaction; page/component UX constraints
clawteam-dev-manager	Engineering manager	Rhythm and dependencies; risks, staffing, milestones
clawteam-team-lead	Tech lead	Technical decisions and quality bar; split of work, review, integration
clawteam-rnd-backend	Backend	Services, APIs, data layer; contracts and implementation
clawteam-rnd-frontend	Frontend	UI and front-end engineering; components, state, integration
clawteam-rnd-mobile	Mobile	Mobile/cross-platform; release constraints
clawteam-devops	DevOps	CI/CD and release; pipelines, artifacts, environments
clawteam-qa	QA	Test strategy and gates; cases, regression scope, severity
clawteam-sre	SRE	Availability, capacity, observability; SLOs, alerts, runbooks
clawteam-project-manager	Project manager	Scope, schedule, stakeholders; milestones and change control
clawteam-scrum-master	Scrum Master	Iteration rhythm and blockers; ceremony and collaboration norms

Short aliases (e.g. qa, sre, product-manager) map to the clawteam-* roles above—see docs/CLAWTEAM_SLASH_GUIDE.md. Why it matters: Moves from “one model, one thread” to multi-role collaborative problem solving.

3) clawteam deeploop: convergent closed-loop iteration

Simulates the real‑world iterative development process of a project team, enabling deep collaborative development (this feature is still under improvement).

/clawteam --deep_loop runs multiple converging rounds—not “one pass and done.”

• Structured contract per round (goals, handoffs, gaps, …)
• Parse DEEP_LOOP_WRITEBACK_JSON and write back automatically when configured
• Tunable convergence thresholds, max iterations, rollback, consistency

Why it matters: Turns “feels done” into metric-driven convergence.

4) ECAP Closed-loop learning:

ClawCode treats experience as a first-class artifact—not only conclusions, but portable structure:

• Experience: An experience function representing the gap between a goal and its outcome. It is a learnable function extracted from the process of resolving the goal–outcome gap, using that gap as the driver for improvement. The dimensional experience objects include: model_experience, agent_experience, skill_experience, and team_experience.
• ECAP A personal/task-level experience capsule representing an evolvable triplet knowledge structure: (Instinct, Experience, Skill).
• TECAP (Team Experience Capsule): A team collaboration experience capsule that includes collaboration steps, topology, and handoffs, and associates a role-level ECAP triplet with each team role.
• instinct → experience → skill: A reusable construction chain from (instinct) rules, through experience, to skills.
• Model → Agent → Team: A reusable learning path from a model, to an agent, to team collaboration of agents.

Implementation mapping (concept to code)

Object	Implementation	Commands / surfaces	Storage	Docs
Experience signals	Distill reusable signals from execution traces	/learn, /learn-orchestrate, /instinct-status	Observations under local data directory	docs/ECAP_v2_USER_GUIDE.md
ECAP	ecap-v2 schema: solution_trace.steps, tool_sequence, outcome, transfer, governance, …	/experience-create, /experience-apply, /experience-feedback, /experience-export, /experience-import	<data>/learning/experience/capsules/, exports/, feedback.jsonl	docs/ECAP_v2_USER_GUIDE.md
TECAP	tecap-v1 → tecap-v2 upgrade; fields like team_topology, coordination_metrics, quality_gates, match_explain	/team-experience-create, /team-experience-apply, /team-experience-export, /tecap-*	On-disk capsules + JSON/Markdown export (--v1-compatible optional)	docs/TECAP_v2_UPGRADE.md
Deeploop write-back	Structured rounds + DEEP_LOOP_WRITEBACK_JSON + finalize	/clawteam --deep_loop, /clawteam-deeploop-finalize	Pending session metadata + LearningService path	docs/CLAWTEAM_SLASH_GUIDE.md
Governance & migration	Privacy tiers, redaction, feedback scores, compatibility	--privacy, --v1-compatible, --strategy, --explain	Audit snapshots; export wrappers (schema_meta, quality_score, …)	docs/ECAP_v2_USER_GUIDE.md, docs/TECAP_v2_UPGRADE.md

Implementation view

flowchart LR
  taskExec[Task execution and tool observations] --> expSignal[Experience signal extraction]
  expSignal --> ecapCreate[ECAP create and structured store]
  ecapCreate --> ecapApply[ECAP retrieve and apply before new work]
  ecapApply --> taskOutcome[Outcome and verification]
  taskOutcome --> feedback[experience_feedback scores]
  feedback --> evolveSkill[instinct_experience_skill evolution]
  evolveSkill --> teamCollab[clawteam collaboration]
  teamCollab --> tecapCreate[TECAP create or upgrade to tecap_v2]
  tecapCreate --> teamApply[team_experience_apply context injection]
  teamApply --> loopGate[deep_loop convergence and write-back]
  loopGate --> expSignal

Loading

Why it matters: The system doesn’t only “do it once”—it improves the next run from feedback.

5) Code Awareness: coding perception and trace visibility

In the TUI, Code Awareness helps with:

• Read/write path awareness and behavioral traces
• Clearer context around the working set and file relationships
• Layering and impact scope

Why it matters: Makes what the AI is doing visible and governable—not a black box.

6) Master–slave agents + Plan / Execute

• Master agent: strategy and control
• Sub-agents / tasks: decomposition and execution
• Plan-then-execute for stable progress

Why it matters: Converge on a plan first, then land changes with less churn.

7) Ecosystem alignment (migration-friendly) + extensions

• Aligns with Claude Code / Codex / OpenCode workflow semantics (complementary positioning)
• Reusable plugin and skill systems
• MCP integration
• Optional computer use / desktop (policy- and permission-gated)

Why it matters: Lower migration cost first, then amplify unique capabilities; stay open to your existing toolchain.

---

Capability matrix (definition–problem–value–entry)

Dimension	Definition	Problem solved	User value	Where to look
Personal velocity	Terminal-native loop (TUI + CLI + tools)	Chat vs real execution drift	Analyze, change, verify in one surface	README.md, pyproject.toml, clawcode -p
Team orchestration	clawteam roles (parallel/serial)	One model can’t cover every function	Integrated multi-role output	docs/CLAWTEAM_SLASH_GUIDE.md
Long-term evolution	deeploop + automatic write-back	Lessons lost when the task ends	Reusable structured experience	docs/CLAWTEAM_SLASH_GUIDE.md (deep_loop / write-back)
Learning loop	Experience / ECAP / TECAP	Hard to migrate or audit “tribal knowledge”	Structured, portable, feedback-ready	docs/ECAP_v2_USER_GUIDE.md, docs/TECAP_v2_UPGRADE.md
Observability	Code Awareness	Opaque tool paths	Clearer read/write traces and impact	docs/技术架构详细说明.md, TUI modules
Extensibility	plugin / skill / MCP / computer use	Closed toolchain	Fit existing ecosystem and grow by scenario	docs/plugins.md, CLAW_MODE.md, pyproject.toml extras

---

Full-stack development loop (diagram)

flowchart LR
  idea[Idea] --> plan[Plan mode]
  plan --> team[clawteam roles]
  team --> execute[Execute and tools]
  execute --> deeploop[deeploop convergence]
  deeploop --> writeback[DEEP_LOOP_WRITEBACK_JSON write-back]
  writeback --> ecap[ECAP / TECAP store]
  ecap --> evolve[Experience evolution and reuse]
  evolve --> plan

Loading

---

How ClawCode compares

Dimension	Typical IDE chat	Typical API-only scripts	ClawCode
Primary surface	IDE panel	Custom scripts	Terminal-native TUI + CLI
Execution depth	Often suggestion-first	Deep but DIY	Built-in tool execution loop
Long-horizon continuity	Varies	Custom state	Local persistence + write-back
Team orchestration	Weak / none	Build yourself	clawteam roles and scheduling
Learning loop	Weak / none	Expensive to build	ECAP/TECAP + deep loop
Observability & governance	Varies	DIY	Config-driven, permission-aware, audit-friendly
Ecosystem	Vendor-bound	Flexible but heavy	plugin / skill / MCP / computer-use paths

Scope note: Capability and architecture comparison only—no “X% faster” claims; based on documented, verifiable behavior.

---

Pro development: slash commands & tools & skills

Beyond migration-friendly defaults, ClawCode ships built-in pro workflows: common multi-step flows as /slash commands, with skills to encode team practice.

1) Built-in /slash commands (engineering workflows)

Cluster	Examples	Typical use
Multi-role & convergence	/clawteam, /clawteam --deep_loop, /clawteam-deeploop-finalize	Roles, converging iterations, structured write-back
Architecture & quality gates	/architect, /code-review, /security-review, /review	Design/review, ranked findings, security pass
Execution orchestration	/orchestrate, /multi-plan, /multi-execute, /multi-workflow	Phased plan → execute → deliver
Test-driven dev	/tdd	RED → GREEN → Refactor with gates
ECAP learning	/learn, /learn-orchestrate, /experience-create, /experience-apply	Distill experience and feed the next task
TECAP team learning	/team-experience-create, /team-experience-apply, /tecap-*	Team-level capsules, migration, reuse
Observability & diagnostics	/experience-dashboard, /closed-loop-contract, /instinct-status, /doctor, /diff	Metrics, config contract checks, environment and diff diagnostics

Full list: clawcode/tui/builtin_slash.py. Deep dives: docs/CLAWTEAM_SLASH_GUIDE.md, docs/ARCHITECT_SLASH_GUIDE.md, docs/MULTI_PLAN_SLASH_GUIDE.md.

3) Built-in tools

🔧 Tools (44 max)

Category	Tools	Description
File I/O	view, ls, write, edit, patch, glob, grep, fetch	Read, list, mutate, and search files in the workspace with permission checks
Shell & execution	bash, terminal, process, execute_code	Shell commands, PTY sessions, process control, and controlled code execution
Search & diagnostics	diagnostics, web_search, web_extract, session_search	LSP diagnostics, web search/extract, and in-session message search
Browser	browser_* (×11)	Local browser automation (registered when browser requirements pass)
Agent	Agent	Spawn sub-agents with a filtered tool set (no nested delegation tools)
Task & state	TodoWrite, TodoRead, UpdateProjectState	Persistent todos and project memo injected into future sessions
Schedule	cronjob	Scheduled and deferred job entrypoint
Skills & memory	memory, skills_list, skill_view, skill_manage, experience_evolve_to_skills	Durable memory and skill listing, inspection, management, and evolution
Optional integrations	mcp_call, sourcegraph, desktop_screenshot, desktop_move, desktop_click, desktop_type, desktop_key	MCP proxy, Sourcegraph search, and desktop automation when configured and requirements pass

Notes: 44 is the maximum number of distinct built-in tool registrations from get_builtin_tools() (all optional rows active). A typical setup loads 37 (includes browser_* when requirements pass and always includes web_search / web_extract). MCP servers add more tool names at runtime.

3) Bundled skills (reusable expertise)

Category	Examples	Delivery value
Backend & API	backend-patterns, api-design, django-patterns, springboot-patterns	Consistent API and backend design; less rework
Frontend	frontend-patterns	Shared UI implementation patterns
Languages	python-patterns, golang-patterns	Idiomatic reusable patterns per stack
Data & migrations	database-migrations, clickhouse-io	Safer schema/data changes; verify rollback
Shipping	docker-patterns, deployment-patterns, coding-standards	Build, release, and quality bars
Cross-tool	codex, opencode	Easier multi-tool workflows
Planning	strategic-compact	Dense, actionable plans for complex work

Paths: clawcode/plugin/builtin_plugins/clawcode-skills/skills/.
Suggested flow: frame execution with /clawteam or /multi-plan, then layer domain skills for consistency.

---

📊 Test Results

Suite	Tests	Status
Unit + Integration	833	✅ Agent, tools, and deep-loop regression (max_iters=100 and runtime hard constraints)
CLI Flags	22	✅ CLI and provider cli_bridge paths
Harness Features	6	✅ Multi-step workflows, harness alignment, and closed-loop smoke
Textual TUI	3	✅ Welcome screen, HUD overlay, and status line
TUI Interactions	27	✅ Chat actions, permission dialogs, and Plan / Arc panels
Real Skills + Plugins	53	✅ Built-in skill registration/execution and plugin sandbox

Collected: 944 pytest items (including parametrized cases). Latest full run: 935 passed, 9 skipped, 0 failed.

---

Quick start

Requirements

• Python >=3.12
• At least one configured model provider credential

Install (source / dev)

cd clawcode
python -m venv .venv
# Windows PowerShell
.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"

Run

clawcode
# or
python -m clawcode

Prompt mode

clawcode -p "Summarize this repository’s architecture in five bullets."

JSON output mode

clawcode -p "Summarize recent changes" -f json

---

Configuration & capability switches

ClawCode is configuration-driven. Main entry points:

• pyproject.toml (metadata and dependencies)
• clawcode/config/settings.py (runtime settings model)

Typical knobs:

• Provider / model selection
• /clawteam --deep_loop convergence parameters
• Experience / ECAP / TECAP behavior
• Desktop / computer-use and other optional features

---

Tiered onboarding

~5 minutes (run it)

1. Install and start clawcode
2. Run clawcode -p "..." once
3. In the TUI, try /clawteam <your ask>

~30 minutes (close the loop)

1. Pick a small real task (fix / refactor / tests)
2. Run /clawteam --deep_loop for 2–3 rounds
3. Inspect DEEP_LOOP_WRITEBACK_JSON and write-back results

Team rollout (repeatable)

1. Align model and policy (provider/model)
2. Inventory reusable skills/plugins and minimal conventions
3. Wire feedback into ECAP/TECAP

---

High-value scenarios

• Greenfield complexity: plan first, execute across converging rounds
• Legacy modernization: multi-role risk ordering and sequencing
• Handoffs: sessions and experience that can be reviewed and migrated
• Long-running work: iterate without losing thread
• Automation: CLI and scriptable batches

---

What’s New

• clawteam --deep_loop automatic write-back path and manual /clawteam-deeploop-finalize fallback
• Convergence-related settings such as clawteam_deeploop_consistency_min
• Deeploop event aggregation and test coverage improvements
• Documentation for clawteam_deeploop_* and closed-loop behavior

---

Roadmap

• Richer Code Awareness (read/write traces mapped to architecture layers)
• Team-level experience dashboards (aggregated metrics)
• Slash workflow templates (task type → flow template)
• Stronger computer-use safety policies and extension hooks

---

Contributing

Contributions welcome. Before opening a PR:

pytest
ruff check .
mypy .

For larger design changes, open an issue first to align scope and goals.

---

Security

AI tooling may run commands and modify files. Use ClawCode in a controlled environment, review outputs, and apply least privilege to credentials and capability switches.

---

License

GPL-3.0 license.