neuralclaw 1.5.7

The Self-Evolving Cognitive Agent Framework

NeuralClaw

Version: 1.5.5

NeuralClaw is a Python agent framework built around a five-cortex runtime: Perception, Memory, Reasoning, Action, and Evolution. The current repository state now covers the full AGENT.md roadmap, including vector memory, persistent identity modeling, browser and desktop control, streaming, structured output, observability, output filtering, audit replay, workspace integrations, and A2A-compatible federation.

Current Capabilities

• Computer use: Take screenshots, analyze screen content with vision, click UI elements, type text, press hotkeys, and launch apps — all controllable remotely via Telegram, Discord, or any channel
• SkillForge: Proactive skill synthesis — give it a URL, API spec, Python library, GitHub repo, MCP server, or plain description and it generates domain-specific tools tailored to your use case. Works from any channel (/forge, !forge, forge:) or the agent forges its own skills mid-conversation
• SkillScout: Discovery layer on top of SkillForge — searches PyPI, GitHub, npm, and MCP registries, ranks results by stars, maintenance, license, and relevance, then auto-forges the best match into a ready-to-use skill. Works from any channel (/scout, !scout, scout:)
• App Builder: Dedicated build_app workflow that provisions new projects under the approved apps workspace root and returns the exact directory for follow-up writes
• Controlled self-improvement: Repeated capability failures are journaled, converted into candidate initiatives, forged or scouted off the live path, and only promoted after probationary tool calls succeed in real use
• Dynamic self-awareness: Agent knows its own capabilities and active tools; never says "I can't" when it has a tool for the job
• Multi-provider routing: openai (GPT-5.4), anthropic (Claude 4.6), openrouter, proxy, local (Ollama) — with GPT-5/o-series API compat
• managed browser-session providers: chatgpt_app, claude_app
• token-backed session auth: chatgpt_token, claude_token
• episodic, semantic, procedural, vector, and identity memory with smart importance scoring
• managed workspaces for app scaffolds and cloned repos
• fast-path, deliberative, reflective, structured, and meta-cognitive reasoning
• vision perception, browser automation, and desktop automation
• streaming responses and Discord voice playback
• Google Workspace and Microsoft 365 built-in skills
• traceline observability, Prompt Armor v2, and audit replay
• swarm, native federation, and A2A interoperability

SkillForge — Teach Your Agent New Tricks

SkillForge turns any input into a deployable NeuralClaw skill. No developer needed.

# CLI
neuralclaw forge create "https://api.stripe.com" --use-case "charge chiro patients"
neuralclaw forge create "twilio" --use-case "send appointment reminders"
neuralclaw forge create "I want to look up drug interactions"

# From Telegram
/forge https://github.com/owner/repo for: analyze X-ray reports

# From Discord
!forge twilio --for send SMS reminders

# The agent can forge its own skills mid-conversation
User: "Can you learn to query our Google Sheets?"
Agent: [forges sheets_query skill] → "Done, I can now read your reports directly."

Supported sources: URLs, OpenAPI/Swagger specs, GraphQL endpoints, Python libraries, GitHub repos, MCP servers, code files, or natural language descriptions.

Skills are saved to ~/.neuralclaw/skills/ and hot-loaded without restart.

SkillScout — Find the Best Tool, Automatically

SkillScout is a discovery layer on top of SkillForge. Instead of handing it a specific URL or library, you describe what you need and SkillScout searches package registries and repositories to find the best candidate, then forges it for you.

Flow:

1. User says: scout: verify patient insurance eligibility
2. SkillScout searches PyPI, GitHub, npm, and MCP registries
3. Results are ranked by stars, maintenance status, license, and relevance
4. The best match is automatically forged into a deployable skill

User: scout: verify patient insurance eligibility

SkillScout searching...
  PyPI:   eligibility-check (★ 340, MIT, maintained)
  GitHub: open-insurance/eligibility-api (★ 1.2k, Apache-2.0)
  npm:    insurance-verify (★ 89, MIT)
  MCP:    mcp-insurance-eligibility (★ 210, MIT)

Best match: open-insurance/eligibility-api (★ 1.2k, Apache-2.0, active)
Forging skill... done.

✓ Skill "insurance_eligibility" forged and loaded.
  Commands: check_eligibility, get_payer_list, verify_member

Channel commands: /scout, !scout, scout:

Install

# Core package
pip install neuralclaw

# Local checkout
pip install -e .

# Development
pip install -e ".[dev]"

# Optional extras
pip install -e ".[voice]"
pip install -e ".[browser]"
pip install -e ".[desktop]"
pip install -e ".[google]"
pip install -e ".[microsoft]"
pip install -e ".[vector]"

# Everything
pip install -e ".[all,dev]"

External runtimes still required for some integrations:

• Playwright browser binaries
• Node.js for the WhatsApp bridge
• signal-cli for Signal
• FFmpeg for Discord voice playback

python -m playwright install chromium

Quick Start

neuralclaw init

neuralclaw session setup chatgpt
neuralclaw session setup claude

neuralclaw session auth chatgpt
neuralclaw session auth claude
neuralclaw session auth google
neuralclaw session auth microsoft

neuralclaw local setup
neuralclaw channels setup

neuralclaw status
neuralclaw session status
neuralclaw doctor

neuralclaw chat
neuralclaw gateway            # foreground terminal session
neuralclaw daemon             # detached background gateway
neuralclaw startup install    # auto-start on login (Windows)
neuralclaw service install    # install managed service
neuralclaw service start      # start managed service

Workspace-Constrained Project Creation

For fresh coding projects, NeuralClaw now uses build_app instead of inventing output directories. By default:

• app scaffolds are created under ~/.neuralclaw/workspace/apps/
• cloned repositories live under ~/.neuralclaw/workspace/repos/

This keeps file writes inside approved roots and gives the agent an exact project path to use for the rest of the task.

Providers

Provider	Purpose	Setup
openai	official OpenAI API	neuralclaw init
anthropic	official Anthropic API	neuralclaw init
openrouter	OpenRouter relay	neuralclaw init
proxy	OpenAI-compatible proxy	neuralclaw proxy setup
local	Ollama or compatible local endpoint	neuralclaw local setup
chatgpt_app	managed ChatGPT browser session	neuralclaw session setup chatgpt
claude_app	managed Claude browser session	neuralclaw session setup claude
chatgpt_token	direct ChatGPT token access	neuralclaw session auth chatgpt
claude_token	direct Claude token access	neuralclaw session auth claude

Channel Trust

Each route can run in one of three modes:

Mode	Behavior
open	always accept inbound messages
pair	require one-time /pair for that route
bound	only trusted bindings may talk

Discord also supports streamed text edits and optional voice playback.

Project Layout

neuralclaw/
  bus/         event bus and telemetry
  channels/    Telegram, Discord, Slack, Signal, WhatsApp, Web, trust layer
  cortex/      perception, memory, reasoning, action, evolution, observability
  providers/   API providers, session providers, router
  session/     managed browser-session runtime
  skills/      manifest model, registry, built-ins
  swarm/       delegation, consensus, mesh, federation
  gateway.py   orchestration entrypoint
  cli.py       command-line interface
  config.py    config, keychain helpers, validation

Docs

• docs/README.md
• docs/configuration.md
• docs/architecture.md
• docs/security.md
• docs/federation.md

Verification

pytest -q
python -m compileall neuralclaw
python -m build
python -m twine check dist/*

PyPI Release

pip install -e ".[dev]"
pytest -q
python -m compileall neuralclaw
python -m build
python -m twine check dist/*

Publish from GitHub Actions by pushing a tag like v1.5.5, or run the manual publish workflow after validating the changelog and built artifacts.