research / agent

FrontierAgent

An open-source terminal agent for research and file-based work, with a single-agent ReAct mode and an Agent Team mode that delegates to parallel sub-agents.

Repository owner ApodexAI

Not rated yet
0 comments

Like 0 Save

Published 11 Oct 2026Listing updated 12 Oct 2026

At a glance

Purpose: An open-source terminal agent for research and file-based work, with a single-agent ReAct mode and an Agent Team mode that delegates to parallel sub-agents.
Runs in: Local agent with shell
Operating systems: macOS, Linux, Windows
Needs: Runtime: Git, Python 3.12 and uv; Needs internet: Yes; Needs accounts or API keys beyond your agent: Not stated
Files and data: Reads files: Yes; Writes files: Yes; Sends data to other services: Not stated
Tool type: Agent
Delivery: From GitHub
Version: Latest on the main branch
Access: Code on GitHub
Compatibility: According to the README: macOS (native or Docker Desktop), Linux hosts and VMs (native by default, bubblewrap or Docker), managed Linux GPU containers, and Windows through WSL2. The README's quick start requires Python 3.12, uv and an OpenAI-compatible model endpoint. For local models the README describes SGLang on Linux with NVIDIA GPUs. The README lists container images for linux/amd64 and linux/arm64.
External costs: According to the README, you need an OpenAI-compatible model endpoint with an API key, base URL and model name. The README promotes a limited-time free offer for the Apodex API platform and gives no prices. The README describes optional web research through a Serper-compatible search endpoint with a provider-issued key, with Serper as the default, or through Parallel Search MCP, which it calls free and keyless. The README's .env block also lists an optional JINA_API_KEY. Your coding agent or model provider may charge for usage.
Permissions: The README describes a task sandbox with read-only /inputs, read-write /workspace and controlled read-write /outputs. The README says interactive writes, deletion, package installation and risky shell commands have approval gates, with some operations denied even under --yes. The README says mutations are journaled for /revert and native task packages are installed under the project's .apodex/runtime/native directory. The README shows uv dependency installation and a user-level uv tool install, with model credentials and endpoint settings in a project .env or a user configuration file under .config/apodex. The README describes model API calls and web research through Serper by default or Parallel Search MCP. The README gives SWARM_NO_WEB=1 and REACT_NO_WEB=1 to disable web tools for their respective workflows.
Usage terms: Apache License 2.0, as stated in the repository. See its LICENSE file for the exact terms.

Code on GitHub

ApodexAI/FrontierAgent

Apache-2.0

Python5,243 stars284 forksLast pushed 13 Oct 2026 UTC

🧩 FrontierAgent, our agent framework, open-sourced alongside it — native command-line TUI, ReAct and Agent Team modes, one command on macOS and Linux, no preinstall, no hard Docker dependency.

Topics: agent-orchestration agentic-ai agentic-framework ai-agents harness multi-agent terminal-agent tui

Public GitHub data fetched 13 Oct 2026 UTC · 0 days old

View on GitHub ↗

What it does

The README describes FrontierAgent as an open-source agent runtime, terminal product and evaluation suite for research and file-based work. The README says ReAct uses one stateful agent for research, file work and commands, while Agent Team uses a coordinator, task board and parallel subagents.

The README describes a shared shell and file sandbox with read-only /inputs, read-write /workspace and controlled read-write /outputs. The README says interactive mutations require approval unless --yes is enabled, some operations remain denied, and file changes are journaled for /revert. The README describes local checkpoints and traces and saved-run continuation with --resume.

The README says framework, tools, workflows and evaluation code are separate layers. The README describes an isolated-subprocess benchmark runner and names BrowseComp and Humanity's Last Exam among its benchmarks.

Who it is for

People who want a command-line agent for research and file-based tasks, and people who run the listed benchmarks against a model.

Example tasks

Research a broad question in Agent Team mode and have it write a report.
Ask ReAct mode to explain a file in a project with a one-shot prompt.
Attach read-only documents to a session before the interface starts.
Resume a saved session.
Run a one-question BrowseComp smoke test with the evaluation harness after downloading the datasets.

Getting started

The README's quick start needs Git, Python 3.12, uv and an OpenAI-compatible model endpoint. Clone the repository, then run uv sync --python 3.12 --extra dev and cp .env.example .env. Put OPENAI_API_KEY, OPENAI_BASE_URL and OPENAI_MODEL in .env. Start a mode with uv run frontier-agent --mode react --cwd /path/to/project or uv run frontier-agent --mode agent_team --cwd /path/to/project. The README also describes a one-time tool install with uv, helper scripts for macOS and Linux, a Docker route and local SGLang serving. The README has the exact commands.

Limitations

The README requires Git, Python 3.12, uv and an OpenAI-compatible model endpoint; Docker is optional. The README says native-mode scientific and document packages are optional and task dependencies are installed inside the project. The README says prebuilt container images are private and require an authorized GitHub Container Registry login, or users can build the checkout. The README lists macOS, Linux hosts and VMs, managed Linux GPU containers, and Windows through WSL2. The README describes local SGLang serving for Linux NVIDIA GPUs and tells users to check its driver compatibility matrix. The README says Parallel Search MCP lacks custom region, language and time filters and requests above 10 results per query, with Serper used when those are needed. The README warns that evaluation model concurrency can approach runner concurrency multiplied by the team spawn limit. The repository can change at any time; check its README before you install.

Data practices

According to the README, the agent uses the OpenAI-compatible endpoint you set in .env. It says every action is traced locally and that a run directory holds the checkpoint, trace, engine log and trajectories. The README says that on macOS/Docker, /outputs maps to .apodex/runs/<session-id>/outputs on the host. The README describes Serper as the default web-search provider and Parallel Search MCP as an alternative selected with WEB_SEARCH_PROVIDER=parallel. The README gives SWARM_NO_WEB=1 to disable Agent Team web tools and REACT_NO_WEB=1 for closed-book ReAct tasks. The README does not mention telemetry.