Agent8088 v1.3.0

Open-source agentic harness for reliable, verifiable task completion, even on small, local models. Grounded execution, robust context management, sandboxed tools and transparent control, on your machine.

What's new in 1.3

• File copy and move: new copy_file and move_file tools copy an existing file byte-for-byte, so an attachment such as a PDF can be filed into a project folder (for example Drive/Finance/). They use the same safety checks as write_file and never create a folder you did not name.
• Checked completion: each part of a request is checked against what the run actually delivered, and a run is not marked done while its latest check fails. Smoke tests judge the whole page or script and tell "could not start here" from a real crash.
• Long-run guidance: the model is told when its time or rounds are running out, a run that stops making progress is asked to decide, and your rules are kept word for word through context compaction.
• Credential safety: pasted text and written files are scanned for credential-shaped strings, the web chat warns before sending one, and another session's attachment cannot be read or copied.
• The public channel moves to the AGENT8088-v1.3 branch: installers, agent8088 --update, the Docker image and the wiki follow it. Existing configuration and memory are preserved.

Install

macOS / Linux / WSL2

curl -fsSL --proto '=https' --tlsv1.2 https://raw.githubusercontent.com/palindrome-rl/AGENT8088/AGENT8088-v1.3/install.sh | AGENT8088_BRANCH=AGENT8088-v1.3 bash

Windows (PowerShell)

$env:AGENT8088_BRANCH = "AGENT8088-v1.3"; iex (irm https://raw.githubusercontent.com/palindrome-rl/AGENT8088/AGENT8088-v1.3/install.ps1)

Docker

docker pull ghcr.io/palindrome-rl/agent8088:1.3.0

From the wheel attached below

pip install agent8088-1.3.0-py3-none-any.whl

Checksums for the attached files are in SHA256SUMS.txt.

See the README for the quick start and the CHANGELOG for the full list of changes.

Agent8088 v1.2.0

Open-source agentic harness for reliable, verifiable task completion, even on small, local models. Grounded execution, robust context management, sandboxed tools and transparent control, on your machine.

Highlights

• Small-model-first: efficient prompts, minimized tool schemas and model-aware routing. /local check probes your hardware and /local available lists local Ollama models that fit it.
• Verification-gated completion: tasks are checked against required outputs, tests and observable tool results before completion is reported.
• Robust context management: pre-call budgeting, overflow prevention, compaction, content handles and persistent task state.
• Web UI alongside the CLI, with live diffs; local OCR; one-command install; all settings in a single plain-text config.
• Guardrails: readonly / plan-only / full-auto permission modes, a native OS sandbox for shell commands, credential protection, and an optional audit log (audit_log=1).
• Extensible: MCP client and server, skills, focused sub-agents, and Slack/Discord/WhatsApp/Telegram/email gateways sharing one agent loop and permission layer.
• Works on macOS, Linux and Windows (Python 3.10+).

Install

macOS / Linux / WSL2

curl -fsSL --proto '=https' --tlsv1.2 https://raw.githubusercontent.com/palindrome-rl/AGENT8088/AGENT8088-v1.2/install.sh | AGENT8088_BRANCH=AGENT8088-v1.2 bash

Windows (PowerShell)

$env:AGENT8088_BRANCH = "AGENT8088-v1.2"; iex (irm https://raw.githubusercontent.com/palindrome-rl/AGENT8088/AGENT8088-v1.2/install.ps1)

Docker

docker pull ghcr.io/palindrome-rl/agent8088:latest

From the wheel attached below

pip install agent8088-1.2.0-py3-none-any.whl

Checksums for the attached files are in SHA256SUMS.txt.

See the README for the quick start and the CHANGELOG for the full list of changes.

Agent8088 v1.0.0

Local-first, tool-using agent runtime with a unified session/permission layer shared across CLI and messaging front ends.

• Permission engine: three-tier (readonly / plan-only / full-auto), per-action approval gating, hard-blocked destructive Git ops, credential-path protection, SSRF/egress controls enforced at the tool-call boundary
• Tooling: 21 built-in tools (filesystem, shell, Git, Playwright/Chromium browsing, web search via SearXNG/Tavily/Exa/DDGS, scheduling); natively exposed or re-served over MCP stdio (--mcp-serve); connects external MCP servers
• Execution sandboxing: native OS sandbox for shell commands, Docker fallback, egress disabled by default
• Sub-agent delegation: 5 restricted profiles (explore, research, code, verify, general-purpose) for context-isolated task execution
• Memory: hybrid keyword + semantic retrieval over a local SQLite store, fully on-device
• Model routing: 12 provider profiles + custom OpenAI-compatible endpoints + local Ollama, with configurable fallback chains on retryable failures
• Front ends: one agent loop, session model, and permission layer across the interactive CLI and Slack/Discord/WhatsApp/Telegram/email gateways

---