MORFEUS

LOCAL-FIRST · OPEN SOURCE · APACHE-2.0

Ask for an outcome.
Keep control of the work.

Morfeus keeps one conversation while it plans, routes, reviews, and records AI-assisted work as durable Goals and Runs. Models are workers; policy, evidence, and terminal state belong to the control plane.

Install Morfeus View source

Current release v0.2.2 · Python 3.11+ · macOS and Linux

morfeus

MORFEUS v0.2.2 · outcome-driven AI assistant

session 82b3672b · new chat

guidance AGENTS.md · RULES.md

boundary YOLO · no sandbox

model local · Qwen3.5-9B · medium

START Describe the result, constraints, and proof of done.

VERIFY Trust tests and evidence, not model confidence.

›

One conversationInternal agents and providers stay out of your way.
Durable workConversation → Goal → immutable Run → Work Package.
Evidence firstA model cannot declare its own work successful.

QUICK START

Install without sudo.

The installer is user-scoped and network-free. Local chat requires llama.cpp.

Terminal
git clone https://github.com/ta2jam/morfeus.git
cd morfeus
./install.sh
morfeus

Requirements

  • Python 3.11 or newer
  • Git
  • llama.cpp for local Qwen chat

macOS: brew install llama.cpp

Full installation notes →

EVERYDAY USE

Describe what done means.

A useful request contains the outcome, constraints, and acceptance evidence.

Example

“Add state migration recovery; preserve existing databases; done when migration, interruption, and rollback tests pass.”

  1. 01

    Discuss

    Plan and refine the outcome with local Qwen in one stable session.

  2. 02

    Run

    Morfeus can split substantial work into bounded background packages.

  3. 03

    Inspect

    Use /runs for progress and /context for active policy files.

  4. 04

    Verify

    Trust tests, command results, and recorded evidence—not model confidence.

CONTROL PLANE

Complexity stays behind the interface.

Local by default

Qwen3.5-9B runs through a managed, loopback-only llama.cpp server with a 16K context cap.

Explainable routing

Advisor records risk, provider, model, effort, cost, and alternatives. Deterministic policy remains final.

Bounded review

Answers can receive isolated critique and at most one revision. Critique is never called verification.

Verified patches

Changes stage in a temporary workspace, pass bounded checks, and apply only if the original workspace has not drifted.

Guidance you can see

AGENTS.md, RULES.md, provider guidance, and known issues are visible, bounded, and unable to grant authority.

Measured resources

CPU time, peak RSS, tokens, latency, and energy proxies are measured before optimization claims are made.

SECURITY BOUNDARY

No vague safety language.

Morfeus deliberately uses a host/unrestricted YOLO runtime. There is no sandbox and no execution approval gate. Coding runs have the invoking user's OS permissions. Cloud chat, delegation, automation, Advisor, and review require separate user-owned permissions.

Read the full security model →

BUILD WITH EVIDENCE

Use the model. Keep the authority.

Get the latest release Report an issue