PHILONT developer preview GitHub

Self-hosted agent runtime

Work that outlives
a chat window.

Philont gives smaller models the runtime around intelligence: persistent pursuits, scoped authorization, execution evidence, recovery, memory, and governed learning.

Run it locally Read the security model

Developer preview. Dogfooded on lower-cost models; task-level cost-per-success benchmarks are not published yet.

RUNTIME / PURSUIT-17 ACTIVE
OWNER DIRECTION

Ship the release only after tests and evidence agree.

  1. 01Persistgoal survives the session✓
  2. 02Plandecompose · review · checkpoint
  3. 03Authorizegrant stays with its purpose
  4. 04Executetools write an evidence ledger
  5. 05Verifyclaim and execution must agree
EXECUTION LEDGER awaiting authorized step

Architecture illustration — not a product recording.

The bet

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

The runtime around the model

Explicit mechanisms.
Inspectable consequences.

The model proposes. The runtime remembers what matters, controls what may happen, and records what actually did.

01 / CONTINUITY

Persistent pursuits

Long-running goals survive sessions. A plain-text compass.md declares what the owner wants the agent to pursue—and where it must stop.

02 / AUTHORITY

Scoped authorization

A permission stays attached to the pursuit that requested it. Ambiguous replies resolve nothing; consequential commands remain separately gated.

03 / EVIDENCE

Execution evidence

Tool calls create a ledger. Recognized completion claims are checked against it and can be intercepted when the claimed work did not occur.

04 / RECOVERY

Checkpointed recovery

Plans revise around failure and resume from useful work instead of silently restarting. Deferred channel messages survive until delivery or visible expiry.

05 / ADAPTATION

Governed learning

Rules and skills carry maturity, reuse checks, demotion, revision history, and repair paths. The learning judge remains in shadow while its signal earns trust.

CURRENT BOUNDARY

Defense in depth,
not a sandbox.

Philont is suitable for dogfooding under a trusted operator—not unattended high-risk production changes.

See evaluation status →

Why smaller models

Spend intelligence
where it matters.

Model scale is valuable. It should not have to carry the plan, permissions, memory, retries, and proof of execution inside one transient context.

MODELreason · propose · judge
+
RUNTIMEpersist · constrain · verify

Philont makes Anthropic- and OpenAI-compatible endpoints interchangeable and treats lower-cost models as a first-class target. The claim is practical—not magical: task-level economics still need reproducible benchmarks.

Designed for

Developers who keep
agents running.

01

Multi-step, interruptible work

software · research · automation
02

Model-cost-aware deployments

hosted · compatible · local
03

Inspectable autonomy

traces · grants · checkpoints
04

Long-lived collaboration

memory · pursuits · skills

Run it yourself

Point a second mind
at something real.

No administrator access or preinstalled Node.js required. Bring an Anthropic- or OpenAI-compatible endpoint.

# macOS / Linux
$ curl -fsSL https://philont.ai/install.sh | bash

# Windows PowerShell
> irm https://philont.ai/install.ps1 | iex

Read the code. Inspect the boundaries. Shape the runtime.

View on GitHub Explore the wiki