MCP Tools Reference

MCP Tools Reference

logos mcp serve speaks the Model Context Protocol as newline-delimited JSON-RPC 2.0 over stdio. Every tool reads or writes one folder on your own disk. None of them reaches a network, and each is annotated openWorldHint: false. Read-only tools are annotated readOnlyHint: true, so hosts with a read-only mode (such as Cursor's Ask) keep them available.

Tools marked receipt return a first line the agent should repeat to the user.

Continuity: "where were we"

resume (read-only, receipt)

Pick up where the last agent stopped.

Arg	Type	Notes
project	string	Omit for the most recent project
agent	string	Your name
budget	int	Token ceiling, default 4000
since	enum	day, week, month, quarter, year, all
conversation	string	Just one conversation: its id, a checkpoint, or part of its intent. budget and since don't apply to it

context (read-only, receipt)

Everything bearing on one task: the last checkpoint, goals, notes, memories and open loops.

Arg	Type	Notes
task	string, required	What you are about to do, in a sentence
project	string	A project, file path or topic
budget	int	Default 4000
since	enum	As for resume

before_you_try (read-only)

Call before proposing a fix, refactor or library. It says whether the approach was already tried and failed, searching every project.

Arg	Type
approach	string, required
project	string

why (read-only)

Call before changing code that looks wrong. It returns the decisions and dead ends recorded for a file.

Arg	Type
file	string, required
limit	int, default 5

note_progress (write, receipt)

One line of what you just did or learned. checkpoint folds these in.

Arg	Type
project	string, required
text	string, required
agent	string

checkpoint (write, receipt)

Save where you're stopping. Call it before ending a session.

Arg	Type	Notes
project	string, required
task	string	What you were trying to do
intent	string	Why it matters; later checkpoints with the same task inherit it
state	string	Where things actually stand
decisions	string[]	"X, because Y"
failed	string[]	Prose, or route: … | observation: … | layer: … (see Core Concepts)
verified	string[]	What you demonstrated, with the command
blockers	string[]	What's broken and what it blocks
commands	string[]	Build/test/lint commands actually run
questions	string[]	Still unresolved
files	string[]	Files touched
next	string	The single next step
agent	string	Your name

handoff (write, receipt)

A checkpoint with a named successor. It takes the same fields, plus to (required): the agent or person taking over.

Memory: "what do you know about X"

remember (write, receipt)

Save a durable fact. It is scoped to the project unless global is set. It may go to the review queue; the result says which.

Arg	Type	Notes
text	string, required	A clear, standalone statement
kind	enum	preference, person, context, fact, procedure. A procedure needs route: … | trap: …
project	string	Defaults to the current folder's project
global	bool	Applies to every project

recall (read-only)

Search memory: this project plus global facts.

Arg	Type	Notes
query	string
limit	int	Default 5
project	string	Search another project
all_projects	bool	Only when the user asks
compact	bool	Ids and one line each
ids	string	Comma-separated ids: full text, no query needed

correct (write, receipt)

Replace a wrong memory by id with text, what is true now. The old version stays in history but drops out of recall. Use this instead of forget plus remember.

forget (write, destructive, receipt)

Delete a memory by id.

pin_memory (write, receipt)

Always include a memory (id) in this project's context packs. unpin: true undoes it.

exclude_memory (write, receipt)

Keep a memory (id) on record but never show it. pin_memory with unpin: true reverses this.

list_memories (read-only)

Every memory with its id.

memory_diff (read-only)

What memory learned or dropped recently. Args: subject (optional), days (default 7).

list_projects (read-only)

The user's projects, most recently active first.

Distilling another agent's session

These two let the calling agent act as the distiller for logos import --from transcripts. The result is a candidate that a person still promotes with logos review --candidates.

ingest_harvest (read-only, receipt)

With no args, it lists the queued sessions. With session (an id prefix or auto record path), it returns that session's commands, files and turns. max_turns defaults to 120, and turns cut off by that limit cannot be cited.

ingest_distil (write, receipt)

Return a distillation for session. Every verified and failed entry must cite a turn, and verified needs a command that succeeded, not just a file edit. Other args: blockers, decided, wanted (auto records only), next and model.

Tool sets

Some hosts send every tool definition with every request. The full set costs about 4.5k tokens a turn. To serve fewer:

"args": ["mcp", "serve", "--tools", "continuity"]

or LOGOS_TOOLS=continuity in the server's env. That serves remember, recall, context, resume, before_you_try, why, note_progress, checkpoint, handoff, ingest_harvest and ingest_distil. The curation tools it drops all have CLI equivalents under logos memory and logos show --projects.

Naming the agent, and other transports

• Set LOGOS_AGENT in a server's env to name the agent in the session trail (default cli for the terminal). An orchestrator running several agents can give each its own name this way without the model's help.
• logos mcp serve --http [--port N] serves the same tools over a local WebSocket for the browser extension. It needs LOGOS_BRIDGE_ORIGIN set and only listens on localhost.