Home

🌐 English | 中文

Philont

A self-hosted agent runtime designed to make smaller models dependable on long-running work.

Philont keeps pursuits alive across conversations. Give it a direction in compass.md; it can research while you are away, ask before crossing permission boundaries, resume interrupted work, and compare completion claims with what its tools actually executed.

It is built first for developers who run agents regularly, care about model cost, maintain long-lived projects, and are willing to self-host a developer-preview runtime. See Who Is Philont For?.

The project explores one engineering thesis:

Reliability should live in the runtime, not only in model scale.

Instead of asking one expensive model call to hold an entire job in its head, Philont makes plans, checkpoints, authorization, execution evidence, persistent memory, and skill lifecycle explicit parts of the runtime.

What is worth exploring

If you care about…	Start here
Running persistent agent workflows on lower-cost models	Why Smaller Models · Plan Protocol
Deciding whether Philont fits your work	Who Is Philont For?
Installing and configuring Philont	Quick Start · Configuration · FAQ
Using Philont from another channel	Channels
An agent that advances work between conversations	Compass · Autonomy & Self-Learning
Checking “done” against execution evidence	Honesty Gates
Authorization, hard-deny validators, and auditability	Permission & Security
Persistent memory and governed skill evolution	Memory System · Autonomy & Self-Learning
Formal and evidence-based reasoning trees	Deep Reasoning
Extending the runtime	Architecture · MCP & Plugins
The identity and continuity design behind the runtime	Selfhood Engineering

What is operational today

Area	Status
Persistent memory, pursuits, channels, plan/deep-reasoning loops	Active
Execution ledger and completion-claim checks	Active, heuristic
Permission matrix, scoped grants, validators, audit log	Active, application-level
Autonomous research	Active, bounded
Learning judge	Shadow — records verdicts but does not mint skills
Long-term learning benefit	Not yet established
Lower-cost-model cost per verified success	Not yet benchmarked
OS-level containment	Not a shipped default boundary

Philont is a developer preview for research, dogfooding, and trusted-operator self-hosting—not unattended high-risk production.

How to read this wiki

• Pages under Use Philont explain the product surface.
• Pages under Runtime mechanisms explain current code and control flow.
• Permission & Security separates enforced controls from assumptions and known limits.
• Evaluation Status separates implementation evidence from product and longitudinal evidence.
• Design documents in the repository may describe proposed or superseded work; the wiki describes the current public contract.

How Philont differs

OpenClaw, Hermes, and Philont overlap and evolve quickly. Philont's emphasis is not “features the others lack”; it is making persistent pursuits, authorization, execution evidence, recovery, and learning governance first-class runtime objects. See Design Centers.

---

中文

English | 中文

Philont

一个自托管 Agent runtime，目标是让较小、低成本模型也能可靠承担持续性工作。

Philont 让任务目标跨对话持续存在。你在 compass.md 里给出方向；它可以在你离开后继续研究，跨越权限边界前请求批准，从中断处恢复，并把“已经完成”的声明与工具真实执行记录对照。

它首先面向每天使用 Agent、关心模型成本、维护长期项目，并愿意自托管 developer-preview runtime 的开发者。参见谁适合使用 Philont？。

项目探索一个工程命题：

可靠性应存在于 runtime，而不应只依赖模型规模。

Philont 不要求一次昂贵的模型调用把整个任务都装在脑中，而是把计划、checkpoint、授权、执行证据、长期记忆和技能生命周期做成显式运行时机制。

值得从哪里开始

你关心…	从这里开始
用低成本模型运行长期 Agent 工作流	为何选择较小模型 · 计划协议
判断 Philont 是否适合你的工作	谁适合使用 Philont？
安装与配置 Philont	快速开始 · 配置 · 常见问题
从其他渠道使用 Philont	渠道
Agent 在对话之间继续推进工作	指南针 · 自主与自学习
用执行证据检查“完成”声明	诚实门
授权、硬拒校验和审计	权限与安全
长期记忆和受治理的技能演化	记忆系统 · 自主与自学习
形式化与证据型推理树	深度推理
扩展 runtime	架构 · MCP 与插件
runtime 背后的身份与连续性设计	自我工程

当前成熟度

领域	状态
长期记忆、pursuit、渠道、计划与深度推理	已启用
执行账本与完成声明检查	已启用，启发式
权限矩阵、作用域授权、校验链、审计日志	已启用，应用层控制
自主研究	已启用，有边界
学习裁判	Shadow——记录判定，但不据此铸造技能
长期学习收益	尚未证明
低成本模型的每次验证成功成本	尚无基准
OS 级隔离	尚不是默认安全边界

Philont 当前是 developer preview，适合研究、日常 dogfood 和可信操作者自托管；不适合无人监管的高风险生产工作。

差异在哪里

OpenClaw、Hermes 与 Philont 的能力大量重叠，也都在快速变化。Philont 的重点不是宣称“别人没有某项功能”，而是把长期 pursuit、授权、执行证据、恢复和学习治理变成一等 runtime 对象。参见设计中心。