uag 0.3.50

A powerful tool execution agent

██╗   ██╗ █████╗  ██████╗ ███████╗███╗   ██╗████████╗ ██████╗██╗     ██╗
██║   ██║██╔══██╗██╔════╝ ██╔════╝████╗  ██║╚══██╔══╝██╔════╝██║     ██║
██║   ██║███████║██║  ███╗█████╗  ██╔██╗ ██║   ██║   ██║     ██║     ██║
██║   ██║██╔══██║██║   ██║██╔══╝  ██║╚██╗██║   ██║   ██║     ██║     ██║
╚██████╔╝██║  ██║╚██████╔╝███████╗██║ ╚████║   ██║   ╚██████╗███████╗██║
 ╚═════╝ ╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝   ╚═╝    ╚═════╝╚══════╝╚═╝

uag (Local AI Agent)

uag is an interactive local agent that can execute commands, manipulate files, and read various data formats (PDF/PPTX/Excel, etc.) on your PC.

• CLI: uag / python -m uagent
• GUI: uagg / python -m uagent.gui
• Web: uagw / python -m uagent.web
• A2A server: uaga / python -m uagent.a2a.server

---

Why uag

• Local-first tool execution with a wide practical tool surface
• Multiple UI entry points: CLI, GUI, and Web
• Multiple providers: Azure OpenAI / OpenAI-compatible, Bedrock, OpenRouter, Ollama, Gemini, Claude, Grok, NVIDIA
• Provider/model switching with session continuity (carry conversation context across LLM changes)
• End-to-end i18n support: localized host UI (UAGENT_LANG) plus optional TO_LLM/FROM_LLM translation for LLM communication
• Strong file/document handling: text, PDF, PPTX, Excel, screenshots, and images
• Session continuity and history controls: :logs / :load, manual :shrink_llm, and optional auto-shrink
• MCP support for discovering and calling external tool servers (mcp_servers_tool.py, mcp_tools_list_tool.py, handle_mcp_v2_tool.py)
• Startup initialization is centralized under src/uagent/runtime_* (workdir/banner/env/memory) and re-exported from src/uagent/runtime_init.py for compatibility
• GPT-5.4+ Responses optimization: lightweight tools prompt, tool_catalog, and narrowed tool exposure per request

---

Minimal Usage

A2A (Agent2Agent) server

uag can also expose an A2A-compatible HTTP server as a separate process (does not affect the existing CLI/GUI/Web behavior).

Start server

# required for authenticated endpoints
export UAGENT_A2A_TOKEN=YOUR_TOKEN

# default: host=0.0.0.0, port=8765
uaga
# or
python -m uagent.a2a.server

Configuration (env)

• UAGENT_A2A_TOKEN: required for authenticated endpoints. If empty, authenticated endpoints are disabled.
• UAGENT_A2A_CONCURRENCY: max concurrent requests (default: 1).
• UAGENT_A2A_ENGINE: backend engine for handling requests (default: uag; tests may use echo).
• UAGENT_A2A_BASE_URL: client-side base URL (default: http://127.0.0.1:8765).
• UAGENT_SIMPLE_PROMPT: defaults to 1. Set 0/false/no/off to disable the readline-based prompt path (useful if the CLI redraws strangely).

Client usage (optional)

You do not need an A2A client for normal uag CLI/GUI/Web usage. Use the client only when you want to call the A2A server from another process or application.

Language / i18n

Host-side UI strings are localized via gettext catalogs.

• Language selection: set UAGENT_LANG (examples: en, ja, zh_CN, zh_TW, es, fr, ko, th).
• If UAGENT_LANG is unset or unsupported, uag falls back to English.

Examples:

Windows (cmd):

set UAGENT_LANG=ko
uag

macOS/Linux:

export UAGENT_LANG=ko
uag

See src/uagent/docs/ADD_LOCALE.md for how to add a new locale.

Developer notes:

• After editing src/uagent/locales/*/LC_MESSAGES/uag.po, regenerate .mo files:

python scripts/compile_locales.py

• Run i18n QC (writes reports under outputs/i18n/):

python scripts/po_qc_summary.py

See also: src/uagent/docs/DEVELOP_I18N.md.

Start

uag
# or
python -m uagent

Specify workdir (working directory)

Usually you do not need to set this explicitly.

• CLI: uag -C ./work
• GUI: uag gui -C ./work (or python -m uagent.gui -C ./work)
• Web: UAGENT_WORKDIR=./work uag web (for Web UI, specify via environment variable)

Exit:

• :exit

Tips (conversation continuity)

• :logs
  Show available session logs.
• :logs 20
  Show up to 20 logs.
• :logs --all
  Show all logs.
• :load 0
  Load the latest session log and continue the conversation.
• :load <index>\
• :skills
  List Agent Skills, choose one, and load its SKILL.md into the session.
• :skills status
  Show the active skill system messages.
• :skills clear
  Remove active skill system messages.
• Loaded skills are stored in the session log and restored on :load.

---

Tool discovery for GPT-5.4+ Responses

When Responses API is enabled and the selected model is gpt-5.4 or later within the GPT-5 line, uag uses a lighter tool-loading path.

• Full tool definitions are not sent up front on every request
• A lightweight tools prompt is used instead of enumerating the whole tool surface
• tool_catalog can be used to discover relevant tools first
• The actual tool specs passed to the model are narrowed based on the user request
• A safe fallback subset is kept when catalog hits are empty

This reduces prompt/tool payload size while preserving the existing full-tool behavior for other models.

---

History compression (manual / auto)

Manual commands:

• :shrink [keep_last] (default keep_last=40): keep the last N non-system messages and drop the rest.
• :shrink_llm [keep_last] (default keep_last=20): summarize older history into one system message and keep the last N non-system messages.

Optional auto shrink (all providers):

• UAGENT_SHRINK_CNT (default: 100)
  • When the number of non-system messages (user/assistant/tool) reaches this count, uag automatically runs the equivalent of :shrink_llm.
  • Set 0 to disable.
• UAGENT_SHRINK_KEEP_LAST (default: 20): how many recent non-system messages to keep after auto summarization.

Log rewrite behavior:

• When shrink runs, the current session log (UAGENT_LOG_FILE / core.LOG_FILE) is rewritten to match the compressed in-memory history.
• A one-generation backup is created under <log_dir>/.backup/.

---

Optional Responses API knobs (reasoning / verbosity)

When using the Responses API (UAGENT_RESPONSES=1) with Azure/OpenAI/Bedrock, you can optionally control reasoning effort and output verbosity.

For Bedrock, uag uses a Bedrock-specific Responses request builder (string input) to match OpenAI-compatible gateway constraints.

If UAGENT_RESPONSES=1 is set with a provider that does not support the Responses API, uag falls back to ChatCompletions at runtime. Ollama is treated as Responses-capable.

• UAGENT_REASONING:
  • auto: automatically choose reasoning.effort per request (Responses API only; streaming is forced off; may retry once on low-quality output)
  • minimal|low|medium|high|xhigh: send reasoning={"effort":...}
  • off / unset / empty: do not send reasoning
• UAGENT_VERBOSITY:
  • low|medium|high: send text={"verbosity":...}
  • off / unset / empty: do not send text.verbosity

In-session commands (CLI/GUI/Web):

• :r [0|1|2|3|auto|minimal|xhigh] (no arg: keep current)
• :v [0|1|2|3] (no arg: keep current)

---

Provider

uag supports multiple LLM providers.

• UAGENT_PROVIDER=openai is treated as OpenAI-compatible (including OpenAI-compatible endpoints).

  • Required: UAGENT_OPENAI_API_KEY
  • Optional: UAGENT_OPENAI_BASE_URL (default: https://api.openai.com/v1)
  • Optional: UAGENT_OPENAI_DEPNAME

• UAGENT_PROVIDER=azure uses Azure OpenAI.

  • Required: UAGENT_AZURE_BASE_URL
  • Required: UAGENT_AZURE_API_KEY
  • Required: UAGENT_AZURE_API_VERSION
  • Optional: UAGENT_AZURE_DEPNAME

• UAGENT_PROVIDER=bedrock uses Bedrock (OpenAI-compatible gateway).

  • Required: UAGENT_BEDROCK_BASE_URL
  • Required: UAGENT_BEDROCK_API_KEY
  • Optional: UAGENT_BEDROCK_DEPNAME

• UAGENT_PROVIDER=openrouter uses OpenRouter (a unified OpenAI-compatible API).

  • Required: UAGENT_OPENROUTER_API_KEY
  • Optional: UAGENT_OPENROUTER_BASE_URL (default: https://openrouter.ai/api/v1)
  • Optional: UAGENT_OPENROUTER_DEPNAME
  • Optional: OpenRouter model fallback (OpenRouter extension)
    • Enabled only when UAGENT_OPENROUTER_DEPNAME="openrouter/auto" (does not affect other providers/models)
    • If you set UAGENT_OPENROUTER_FALLBACK_MODELS (comma-separated), models=[...] is added to the Chat Completions request.
      • Example: UAGENT_OPENROUTER_FALLBACK_MODELS="anthropic/claude-4.5-sonnet,openai/gpt-4o,mistral/mistral-x"

• UAGENT_PROVIDER=nvidia uses NVIDIA (OpenAI-compatible).

  • Required: UAGENT_NVIDIA_API_KEY
  • Optional: UAGENT_NVIDIA_BASE_URL (default: https://integrate.api.nvidia.com/v1)
    • Please specify /v1 (we do not recommend including /v1/chat/completions in the base URL)
  • Optional: UAGENT_NVIDIA_DEPNAME

• UAGENT_PROVIDER=grok uses xAI Grok (OpenAI-compatible).

  • Required: UAGENT_GROK_API_KEY
  • Optional: UAGENT_GROK_BASE_URL (default: https://api.x.ai/v1)
  • Optional: UAGENT_GROK_DEPNAME

• UAGENT_PROVIDER=gemini uses Google Gemini (google-genai).

  • Required: UAGENT_GEMINI_API_KEY
  • Optional: UAGENT_GEMINI_DEPNAME

• UAGENT_PROVIDER=ollama uses Ollama.

  • Optional: UAGENT_OLLAMA_BASE_URL (default: http://localhost:11434/v1)
  • Optional: UAGENT_OLLAMA_API_KEY (default: dummy)
  • Optional: UAGENT_OLLAMA_DEPNAME
  • Optional: Ollama request knobs
    • UAGENT_OLLAMA_TEMPERATURE (default: 0.7)
    • UAGENT_OLLAMA_TOP_P (default: 0.9)
    • UAGENT_OLLAMA_TOP_K (default: 40)
    • UAGENT_OLLAMA_REPEAT_PENALTY (default: 1.1)
    • UAGENT_OLLAMA_KEEP_ALIVE (default: 5m)
    • UAGENT_OLLAMA_NUM_CTX (default: 4096)
    • UAGENT_OLLAMA_NUM_PREDICT (default: 1024)

• UAGENT_PROVIDER=claude uses Anthropic Claude.

  • Required: UAGENT_CLAUDE_API_KEY
  • Optional: UAGENT_CLAUDE_DEPNAME

See samples/env.sample.env for the canonical cross-provider template and samples/provider.*.env.sample for provider-specific templates.

• In this repository: run the interactive wizard python -m uagent.setup_cli to generate samples/env.sample.sh / samples/env.sample.ps1 / samples/env.sample.bat with the intended encoding and newline settings.
• After installing via pip/wheel: run uag_setup to generate your own .env (and optionally env.sh / env.ps1 / env.bat) in the current directory.

For details, see samples/README.md.

Env sample generation

Sample files are available under samples/ (including samples/README.md).

Run the interactive wizard to configure and generate shell-specific variants:

python -m uagent.setup_cli

Generated files and format:

• samples/env.sample.sh : UTF-8, LF
• samples/env.sample.ps1 : UTF-8 with BOM (utf-8-sig), CRLF
• samples/env.sample.bat : CP932, CRLF

The wizard supports numbered selections and back navigation (b). Re-run it any time to update configuration.

---

Optional Translation (TO_LLM / FROM_LLM)

By default, uag does not translate.

Enable translation by setting:

• UAGENT_TRANSLATE_PROVIDER: translation provider (OpenAI-compatible string, e.g. openai / azure / openrouter / nvidia / grok).
• UAGENT_TRANSLATE_TO_LLM: language tag to translate into before sending to the LLM (e.g. en).
• UAGENT_TRANSLATE_FROM_LLM: language tag to translate into for displaying LLM outputs (e.g. ja).

OpenAI-compatible translation settings:

• UAGENT_TRANSLATE_DEPNAME: model name for translation (required when translation is enabled).
• UAGENT_TRANSLATE_API_KEY: optional (falls back to the main provider key).
• UAGENT_TRANSLATE_BASE_URL: optional (falls back to the main provider base URL).

Notes:

• Translation is done per call (stateless).
• When translation is enabled, streaming is forced off to avoid mismatched partial outputs.

---

Image Generation / Analysis

Image Generation (generate_image)

• UAGENT_IMG_GENERATE_PROVIDER: Override the provider for image generation (fallback: UAGENT_PROVIDER).
• UAGENT_IMAGE_OPEN: Whether to automatically open the image after generation.
  • 1: Open (default)
  • 0: Do not open

Model / deployment name (provider-specific):

• UAGENT_<PROVIDER>_IMG_GENERATE_DEPNAME (required)
  • Examples: UAGENT_OPENAI_IMG_GENERATE_DEPNAME, UAGENT_AZURE_IMG_GENERATE_DEPNAME

Provider-specific credentials / endpoints:

• UAGENT_<PROVIDER>_IMG_GENERATE_API_KEY (required)
• UAGENT_<PROVIDER>_IMG_GENERATE_BASE_URL (optional for most providers; default may apply)
• Azure only: UAGENT_AZURE_IMG_GENERATE_API_VERSION (required)

Fallback behavior:

• If a *_IMG_GENERATE_* env var is not set, the tool also tries the main provider env (e.g. UAGENT_OPENAI_API_KEY, UAGENT_OPENAI_BASE_URL).

Notes:

• Depending on the provider SDK/API, supported sizes/features may differ.

Image Analysis (analyze_image)

By default, the analyze_image tool uses the provider specified in UAGENT_PROVIDER. You can override this using specific environment variables.

• UAGENT_RESPONSES=1: If enabled, the analyze_image tool is hidden, and the agent uses multimodal capabilities of the main LLM directly (if supported).
• UAGENT_IMG_ANALYSIS_PROVIDER: Override the provider for image analysis.

Provider-specific overrides:

• UAGENT_<PROVIDER>_IMG_ANALYSIS_DEPNAME
• UAGENT_<PROVIDER>_IMG_ANALYSIS_API_KEY
• UAGENT_<PROVIDER>_IMG_ANALYSIS_BASE_URL

Allowed providers for analyze_image: openai, azure.

---

Release Notes

• Added optional auto shrink_llm (all providers).
  • UAGENT_SHRINK_CNT (default: 100): when the number of non-system messages (user/assistant/tool) reaches this count, uag automatically runs the equivalent of :shrink_llm.
  • UAGENT_SHRINK_CNT=0: disable auto shrink.
  • UAGENT_SHRINK_KEEP_LAST (default: 20): how many recent non-system messages to keep after summarization.
• Added GPT-5.4+ Responses tool narrowing with tool_catalog and a lightweight tools prompt.
• Added smoke tests for MCP server management tools.
• When shrink runs (manual :shrink / :shrink_llm or auto), the current session log is rewritten to match the compressed in-memory history.
  • A one-generation backup is created under <log_dir>/.backup/.

---

For Developers

For developer-focused information, see src/uagent/docs/DEVELOP.md.

Additional docs:

• src/uagent/docs/RUNTIME_INIT.md (startup initialization: workdir/banner/long-term memory)
• src/uagent/docs/WEB_UI_LOGGING.md (Web UI logging/message paths)

---

Web Inspector (playwright_inspector)

You can use Playwright Inspector to capture manual browsing flows (URL transitions / DOM / screenshots / event logs).

Prerequisites:

• playwright is installed
• Browser setup is done (e.g. python -m playwright install)

Docs:

• src/uagent/docs/WEBINSPECTER.md
• uag docs webinspect (available even in wheel environments)

---

Documentation (uag docs)

Even when installed from a wheel (whl), bundled documents are available via uag docs.

uag docs
uag docs webinspect
uag docs develop
uag docs --path webinspect
uag docs --open webinspect

---

Install (distribution: wheel)

See QUICKSTART.md for the Windows-oriented installation steps using a distributed .whl.

• Distribution: GitHub Releases page (Assets: .whl)
• Wheel example: uag-<VERSION>-py3-none-any.whl

Example:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install ./uag-<VERSION>-py3-none-any.whl

Notes:

• uag requires Python 3.11+.
• For development use, python -m pip install -e ..

---