The recall-first discipline — and why subagents reach recall via the CLI, not MCP.

Working with AI agents

recall is built to sit at every point where an AI fetches code context — not just the human-in-the-loop session. Most tools help the human; recall plugs into the AI's whole workflow, so the engine you and your team build gets stronger with every commit instead of being re-derived from scratch each session.

How an AI ecosystem plugs in — the five docking points

There are five moments where an AI reaches for code context. recall docks onto each:

#	Point	What recall does
1	Subagents	a spawned review/audit agent starts blank → it runs `recall brief --terse` first, so it inherits the why instead of re-guessing it
2	Pre-edit	`recall brief <file>` before any edit — the gate below
3	Search / locate	`recall resolve <guess>` corrects the hallucinated term before the grep (Search-inversion)
4	Compaction / handoff	`recall handoff` snapshots the in-flight state so the next session rebuilds from recall, not an ad-hoc summary
5	Git hooks	pre-commit risk-warning + post-commit auto-stamp

The full map + status lives in docs/ecosystem-docking.md. Claude Code is the first complete client — the same MCP + CLI surface works for any AI agent (every AI has the same blank-subagent and vocabulary-mismatch gaps).

Wiring recall into your AI client

recall speaks two universal protocols, so it docks into any AI coding tool — not just one. The same engine, two transports:

MCP (Model Context Protocol) — for the interactive session and real users. Every

MCP client (Claude Code, Cursor, VS Code / GitHub Copilot, Windsurf, Zed, Cline, …) registers recall with the same server block; only the config file location differs. Run recall mcp --print-config to print the exact snippet for each:

| Client | Where the config lives | | --- | --- | | Claude Code | claude mcp add recall -- recall mcp, or a checked-in .mcp.json | | Cursor | ~/.cursor/mcp.json (global) or .cursor/mcp.json (project) | | VS Code / Copilot | .vscode/mcp.json (uses a servers key) | | Windsurf / Zed / Cline / other | the same mcpServers block, that client's path |

That exposes recall's 14 tools (recall, brief, explain, resolve, stamp, contested, freshen, dashboard, impact, precedent, callers, dead_code, untested, cycles) as native tools the AI can call.

CLI — for spawned subagents, which can't see MCP (it's session-scoped, see

below). Any agent in any framework reaches recall through the shell: recall brief <file> --terse. No integration needed — if the agent can run a command, it can use recall.

The point: recall isn't a plugin for one assistant. It's a memory layer under the whole ecosystem — whatever AI (or mix of AIs) your team uses, they all read and write the same project brain, so the engine gets stronger no matter who's driving.

The gate: orient before you edit

Before editing any code file, run recall brief <file> first. recall and grep are not alternatives — they run in order:

Orient — recall brief <file> (or recall "<question>"): what must I know

before I touch this? Open tasks, the why, the blast radius. The step grep can't do.

Locate — grep / read: the exact string, the whole file. And before you grep

a name you're guessing, run recall resolve <guess> (Search-inversion).

Edit — now, with both the why and the where in hand.

The agent rule: CLI, not MCP

If you spawn subagents (a review fleet, an audit workflow), each one starts blank — it doesn't inherit your session's memory or MCP connection. **MCP servers are session-scoped, so a spawned subagent cannot see the recall MCP tools** (measured). So an agent reaches recall through the CLI via the shell:

recall brief <file> --terse      # the machine-first pre-edit briefing
recall "<concept>" --terse       # locate by concept
recall resolve <guess> --terse   # correct a guessed term

**Every finder/reviewer/auditor agent you spawn must recall brief <file> --terse before it judges that file.** A blank agent re-derives intent from code alone and raises false alarms about decisions made on purpose. Measured A/B (2026-06-14): a recall-less agent raised ~4–5 false alarms a recall-first agent didn't — while the recall-first agent also found a real bug the other missed. recall-first is less noise AND not blind.

MCP stays the path for the main interactive session and real users in Claude/Cursor.

The 6 dimensions

After every feature and before every push, walk the canonical audit raster in docs/audit-dimensions.md: **Auth-Guard · Audit-Log · Zod/Validation · Error-Handling (res.ok) · State-Updater · No-Client-Secrets/Cleanup**, plus the standing security lenses (business-logic/race/money, injection/takeover). Run it recall-first; verify findings adversarially (most raised findings are false alarms).

The flywheel

Every commit makes the code smarter and every suggestion richer — passively. The search-inversion synonym layer literally learns from what searches landed; the trail/territory work (post-launch) deepens it. Old human code is dumb at first, so the suggestions are dumb at first; they warm with use, and the curve is measurable.