Glaux

Configuration

Models

Glaux runs offline by default. Switch backends at runtime — no restart:

glaux models                                   # current + available
glaux models --set ollama:qwen2.5:7b-instruct  # fully local
glaux models --set anthropic                   # needs ANTHROPIC_API_KEY
glaux models --set openai                      # needs OPENAI_API_KEY

30+ providers are supported. API keys are read from the environment and are never written to the ledger or logs.

Model tiering. Routing and reflection can use a cheaper model than the specialists that do the real work — set a fast model via POST /api/llm/fast or the dashboard.

Environment variables

Variable Default Purpose
GLAUX_HOME ~/.glaux State directory: ledger, memory, skills, workspace, logs
GLAUX_API_SECRET — Require signed requests on the HTTP API
GLAUX_WORKSPACE off Seed the jailed file workspace as enabled
GLAUX_PLAYBOOKS off Seed authored playbooks as exposed to the swarm
GLAUX_ANSWER_CACHE 0 (off) Enable the semantic answer cache
GLAUX_CACHE_THRESHOLD 0.92 Similarity required for a cache hit
GLAUX_CACHE_MAX 500 Max cached answers
GLAUX_CONTEXT_MAX_CHARS 6000 History budget per prompt on long runs
GLAUX_CONTEXT_KEEP_RECENT 3 Steps kept verbatim before digesting
GLAUX_SANDBOX_BACKEND auto wasm or subprocess
GLAUX_PYTHON_WASM — Path to python.wasm for the WASM sandbox
GLAUX_ANTHROPIC_CACHE_TTL — Set 1h for the extended prompt cache

Each variable also answers to its pre-rename ATHENA_* name (the GLAUX_* one wins when both are set), and an existing ~/.athena keeps being used while there is no ~/.glaux.

Runtime toggles (glaux workspace on, glaux cache on, …) override the env seed and persist, so the env vars are only about the starting state.

Capabilities

Glaux ships conservative — each capability is opt-in:

Capability Enable What it grants
Workspace glaux workspace on Real files in one jailed, size-capped directory
Playbooks glaux playbooks on The swarm can list and follow authored procedures
Answer cache glaux cache on Near-identical repeats answered with zero model calls
Auto-evolve glaux evolution auto on The harness improves itself on a cadence

Safety controls

glaux safety live       # normal
glaux safety dry-run    # skip side effects
glaux safety halted     # stop all execution
glaux approval gated    # a human approves each synthesized skill before it can run

Budgets

Each task runs under a budget (model calls, tokens, spend). The swarm halts cleanly when a limit is hit rather than running away — the halt reason is returned with the result and recorded.

Context management

Long runs compact their own history: recent steps stay verbatim, older ones collapse to one-line digests, and the deepest history is elided — all under GLAUX_CONTEXT_MAX_CHARS. Every compaction is signed to the ledger, so nothing silently disappears from the record.