_ _ _ __
/ \ __ _ ___ _ __ | |_(_)/ _|_ _
/ _ \ / _` |/ _ \ '_ \| __| | |_| | | |
/ ___ \ (_| | __/ | | | |_| | _| |_| |
/_/ \_\__, |\___|_| |_|\__|_|_| \__, |
|___/ |___/
Switch agents. Keep the repo's working memory.
Agentify keeps durable working context — decisions, notes, session summaries, command failures, hot files — with the repository instead of inside one agent harness. Install it once and the agent drives it: Claude Code through lifecycle hooks, Codex through installed guidance, everything else over MCP.
It does not replay a provider's hidden conversation state or copy private chain-of-thought. It carries forward explicit, compact project evidence that should survive the switch.
curl -fsSL https://raw.githubusercontent.com/ixigo/agentify/main/install.sh | bash
cd /path/to/your/repo
agentify install # guidance + hooks + MCP registration + index
agentify status # what is wired
agentify uninstall # removes only Agentify's managed blocks and hooksagentify install
|-- CLAUDE.md / AGENTS.md <- agent guidance
|-- .claude/settings.json <- Claude Code lifecycle hooks
`-- .agentify/ <- JSONL context store + optional repo index
- Session starts — the digest is injected (hook on Claude Code, guidance on Codex).
- Agent works — edits, commands, and failures are tracked automatically; a repeat of a previously failed command is flagged before it runs.
- Agent learns something —
agentify ctx note "…"/agentify ctx decision "chose X over Y because Z". - Session ends — a short extractive handoff is written from tracked evidence, at zero model cost.
No daemon, no database server, no per-command wrapping.
| Command | What it does |
|---|---|
agentify ctx load |
Digest of recent activity, notes, hot files |
agentify ctx note "<text>" |
Record a note for future sessions |
agentify ctx decision "<text>" |
Record a decision; query it later with ctx decisions |
agentify risk --since <ref> |
Blast radius + suggested regression tests |
agentify test --since <ref> --run |
Run only the tests a change affects |
agentify delegate <kind> "<task>" |
Shell work out to the model routed for it |
agentify serve |
MCP server for any MCP-capable agent |
Every command takes --json. Full reference and the other 28 commands: docs/README.md.
Measured, with committed receipts under evals/results/ and CI that fails if the code stops reproducing them. Against the same agent with no memory layer the effect is large and consistent (28/0 discordant pairs, p = 7.5e-9 on a 180-trial matrix); against a hand-maintained CLAUDE.md memory bank holding the same knowledge, two campaigns found no significant difference — published rather than buried. No leaderboard claims.
Full tables, caveats, and the corrections log: docs/benchmarks.md.
Node.js 20+, Git, and Claude Code (automatic hooks) or Codex (guidance-driven). Any other agent can call the CLI or connect over MCP.
- docs/README.md — command reference, integrations, and where everything else lives
- docs/usage.md — the full guide
- docs/benchmarks.md — what has actually been measured
git clone https://github.com/ixigo/agentify.git && cd agentify
pnpm install && node --test
pnpm link --globalMIT