Perseus · live context engine · v1.0.23

Every session starts already oriented.

Perseus resolves your workspace (git, services, tests, tasks) into verified facts before the context window opens. Your agent reads what's true, not instructions to go find out. Fewer tokens, no cold-start tax.

67% smaller tool schema 611× warm-cache renders 33 MCP tools MIT · local-first
$pip install perseus-ctx && perseus quickstart
Resolve before context

Facts, not instructions to go find facts.

Static context files rot the moment you write them. Perseus renders live state at the moment the session opens, so nothing is stale, and the agent never burns turns re-discovering it.

Without Perseus · stale & vague
"Port is 3001 (check .env)"
"47 tests (may be stale)"
"Check docker ps first"
"Where did we leave off?"
With Perseus · resolved & verified
Port: 3001
Tests: all passing (run 8s ago)
mongo-dev: Up 4h 12m
Checkpoint: webhook handler, pending test run
611×

Warm renders, measured

A resolved document is a fraction of the size of "go check these ten things." On the Perseus repo itself, 22,528 varied directives rendered in 1.014s warm vs 619.1s cold.

0

Orientation turns wasted

No "let me check what's running" round-trips. The first turn is productive because the context was already true when it arrived.

1:1

Same answers, fewer tokens

A live judge found 20/20 A/B pairs semantically equivalent: Perseus changes what the agent knows, not what it says.

Under the hood

A context compiler, not another config file.

Directives

Write @query, @services, @read, @waypoint; Perseus resolves them to plain markdown at render time.

Tiered context

Progressive disclosure: lean tier-1 core by default; the agent pulls heavier context on demand. One integer compare per directive.

Prompt-size forensics

perseus prompt-size attributes every byte per directive; @budget fails CI when context bloats.

Recall-first profiles

Per-model context profiles retrieve memory on demand instead of stapling a standing blob into every prompt. Prefix-cache friendly.

Waypoints & multi-agent

Crash-safe checkpoints on disk; namespaced, lock-protected stores let agents hand off without an orchestration platform.

Extensible

Plugins, macros, render hooks, and pipes: drop Python into ~/.perseus/, no source patching.

Measured, reproducible.

Named methodology and rerunnable scripts in the repo.

Works with your assistant

One command. Any MCP host.

Perseus speaks the Model Context Protocol and writes whatever file your assistant reads at session start.

Claude Code Claude Desktop Cursor Codex Hermes Agent Rovo Dev
renders → CLAUDE.md · AGENTS.md · .cursorrules · .hermes.md
2026 paradigm

Context Engineering is the memory layer of the AI harness

Harness engineering — the practice of building the infrastructure around LLMs rather than fine-tuning the models themselves — was named the defining AI paradigm of 2026 by OpenAI, Anthropic, ThoughtWorks, Red Hat, and Mitchell Hashimoto. Perseus implements the Context Engineering dimension: resolving workspace state into verified facts before the model ever sees a token.

Stage 2 → 3
Static → Resolved
Hand-written CLAUDE.md → @query directives that execute shell commands, check services, and embed live state
Stage 3 → 4
Resolved → Memorable
Every session starts fresh → @memory persists decisions, conventions, and corrections across sessions via Perseus Vault
Stage 4 → 5
Memorable → Self-improving
Passive recall → active @focus attention, @research synthesis, and automated coherence maintenance
Read the full harness engineering position →
Pairs with

Perseus Vault · persistent memory

Perseus resolves the present; Vault remembers the past. Add encrypted, durable memory so your agent recalls decisions and conventions across sessions.

Explore Vault →