Agents

Any coding agent that can run shell commands or speak MCP can use Codifix. Codex, Claude Code, OpenCode, and custom agents all follow the same pattern.

How agents connect

CLI

The fastest path. The agent calls the codifix binary directly:

codifix query  --project <slug> --question "What context matters for this task?"
codifix resume --project <slug>
codifix remember --project <slug> --title "…" --summary "…" --note "…"

MCP

For agents that speak MCP rather than shell. Use the stdio server for local clients, or Streamable HTTP for local, token-bearing clients. See MCP for server setup, tools, and security.

For Codex Desktop, use the Codex plugin. It packages the supported stdio MCP connection and workflow skill; keep one Codifix MCP connection enabled so tools are not registered twice.

Repo-local instructions

The project wizard writes .agents/ files that agents pick up automatically. These contain CLAUDE.md, AGENTS.md, and Codex instructions with evidence-first rules already embedded.

codifix wizard          # writes .agents/ alongside .codifix/
ls .agents/            # inspect what was generated

Agent prompts

Copy-paste prompts for common setup tasks. The agent handles the rest.

Install Codifix

You are installing Codifix for me. Work in the terminal, explain before
using sudo, and stop before destructive changes.

Find the correct install path for this OS, configure PostgreSQL with pgvector,
run `codifix wizard --global`, run health checks, and report where config files
live. Do not delete data or overwrite existing config without asking first.

The agent should finish with codifix doctor, codifix health, installed version, service status, and next commands.

Configure a project

You are configuring this repository for Codifix. Do not install the
package. Inspect existing `.codifix` and `.agents` files, run
`codifix wizard --dry-run`, explain planned writes, then run `codifix wizard`
only after the preview is safe. Verify with
`codifix status --project <project-slug>` and a simple query.

The agent should report changed files, project slug, health/status output, and any follow-up checks.

Briefings

Bring a new or returning agent up to speed without re-discovering context:

codifix resume       --project <slug>                # recent work + next steps
codifix up-to-speed  --project <slug>                # broader project context
codifix activities   --project <slug> --limit 20     # raw activity feed

Use briefings at task start, after interruptions, and before continuing work that another agent began.

Recommended behavior

1. Query before claiming — ask project memory before making project-specific assertions.
2. Cite evidence — reference memory IDs or source evidence when using remembered facts.
3. Verify before editing — inspect code and docs in the working tree, not stale training data.
4. Store discoveries — run codifix remember after meaningful work so the next session benefits.
5. Keep secrets out — never commit database URLs, API keys, or provider tokens to repo files.

Next

Read MCP, Watchers, or How it works.