Harness setup

Causal is experience memory for AI agents — it records what your agents tried, what happened, and why, then serves that experience into the next decision. It runs inside Claude Code, Cursor, and Codex, and inside OpenClaw and Hermes.

Supported harnesses

The first release covers Claude Code, Cursor, Codex, OpenClaw, and Hermes. You keep the agent you already work in. Causal installs into it.

Two pieces land in each one. The MCP server is the connection — it is how the agent reaches the experience graph to read precedent and write outcomes. The always-on rule is the trigger. Each harness has its own name for it, a rule or a project rule or a hook, and the installer writes whichever one that harness expects. Without it, recall is something the agent has to remember to ask for, and it arrives after the agent has already chosen a direction. With it, the relevant precedent is in context when the agent starts deciding, and a prior attempt comes back with the outcome it caused. The longer account is on Recall and capture.

The installer detects the runtime you are in, signs you in, installs the Causal Agent Skills, registers the MCP server, and adds the always-on rule that recalls relevant experience before acting and captures meaningful outcomes afterward. No MCP config to hand-edit, no API keys in project files.

Causal serves precedent and evidence. Your agent still reasons about what to do.

Install commands

One command per agent. Plugin harnesses use --plugin; agent runtimes use --harness. Each section below lists the command and the lines the installer reports when it finishes.

Claude Code

npx causal init --plugin claude-code

The installer reports: MCP server registered, memory skill installed, always-on recall rule added.

Cursor

npx causal init --plugin cursor

The installer reports: MCP server registered, project rule installed, outcome capture wired.

Codex

npx causal init --plugin codex

The installer reports: MCP server registered, AGENTS.md hook written, session briefing enabled.

OpenClaw

npx causal init --harness openclaw

The installer reports: MCP server registered, write firewall active, per-writer signing enabled.

Hermes

npx causal init --harness hermes

The installer reports: MCP server registered, recall hook armed, outcome telemetry wired.

Running npx causal init with no flag detects the harness instead. Install Causal in your agent covers that path end to end.

Handoff between harnesses

The five read and write the same graph, so what one agent recorded is what the next one reads. A long Claude Code session that ends at a usage limit leaves its experience behind, and a fresh Cursor session picks the work up from that experience rather than from a re-paste of the transcript. The handoff runs in either direction across all five.

What the second agent inherits is prior actions and their outcomes, not stored context. A precedent arrives with the outcome it caused, so it can strike a plan today instead of sitting in a search index. An attempt that shipped the wrong part variant last month and cost return freight comes back as that, and the next agent confirms the variant first. That example is written to show the shape; it is not a real record.

Follow-up work stays attached to what caused it, so a change made in Codex after a decision taken in Claude Code traces back to that decision instead of starting a new history. AI agent governance and audit trail covers who may read which entries, how long they are kept, and how a served precedent traces back to the work that produced it.

What if my agent is not on that list?

Python and TypeScript SDKs cover anything outside the five.

npm install causal

pip install causal

Both packages are subject to the same pre-launch caveat as the installer: the names are settled, but nothing is published yet.

Per-harness details

Not yet published. The configuration each harness receives — where the rule or hook is written, what it contains, and how to verify it by hand — will be documented per harness when the installer ships. The commands and reported outcomes above are provisional while Causal is in private beta.