Skip to content

MCP Setup

Loomtracer’s command-line tool, loomtracer, exposes its full decision-graph surface as an MCP server. Connect your agents to your self-hosted cell over HTTP (a shared, team-wide decision graph), or run the local stdio server yourself for single-user use.


Your self-hosted cell serves MCP at /mcp. Agents connect to that endpoint and write to a shared, team-wide decision graph — no local binary required on the agent’s machine. Authentication is a bearer token: provision a tenant and use the lt_tk_... token it returns. The examples below use http://localhost:8080/mcp; substitute your cell’s address.

Add to .mcp.json in your project root:

{
"mcpServers": {
"loomtracer": {
"type": "http",
"url": "http://localhost:8080/mcp",
"headers": { "Authorization": "Bearer lt_tk_..." }
}
}
}

Or from the CLI:

Terminal window
claude mcp add --transport http loomtracer http://localhost:8080/mcp \
--header "Authorization: Bearer lt_tk_..."

Reload Claude Code and all loomtracer tools are available.

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
"mcpServers": {
"loomtracer": {
"type": "http",
"url": "http://localhost:8080/mcp",
"headers": { "Authorization": "Bearer lt_tk_..." }
}
}
}

Add to ~/.cursor/mcp.json or the project-level .cursor/mcp.json:

{
"mcpServers": {
"loomtracer": {
"url": "http://localhost:8080/mcp",
"headers": { "Authorization": "Bearer lt_tk_..." }
}
}
}

Put the token in an environment variable and register the cell:

Terminal window
export LOOMTRACER_TOKEN=lt_tk_...
codex mcp add loomtracer --url http://localhost:8080/mcp --bearer-token-env-var LOOMTRACER_TOKEN

Codex writes the server to ~/.codex/config.toml and reads the bearer token from that variable.


For single-user use, run the local stdio server. The server is a thin transport over the same commands layer the CLI uses.

Terminal window
loomtracer --loomtracer-dir ./loomtracer/ mcp

The server reads from stdin and writes to stdout in the MCP protocol format.

{
"mcpServers": {
"loomtracer": {
"command": "loomtracer",
"args": ["mcp"],
"env": {
"LOOMTRACER_DIR": "/absolute/path/to/your/loomtracer/dir"
}
}
}
}

Add to .mcp.json in the project root:

{
"mcpServers": {
"loomtracer": {
"command": "loomtracer",
"args": ["--loomtracer-dir", "./loomtracer/", "mcp", "--agent-tool", "claude"]
}
}
}

Or set LOOMTRACER_DIR and omit --loomtracer-dir:

{
"mcpServers": {
"loomtracer": {
"command": "loomtracer",
"args": ["mcp", "--agent-tool", "claude"],
"env": { "LOOMTRACER_DIR": "./loomtracer/" }
}
}
}

--agent-tool claude files captures under Claude Code (see Who a capture is filed under).

{
"mcpServers": {
"loomtracer": {
"command": "loomtracer",
"args": ["mcp", "--agent-tool", "cursor"],
"env": { "LOOMTRACER_DIR": "/path/to/loomtracer/" }
}
}
}
Terminal window
codex mcp add loomtracer -- loomtracer mcp --agent-tool codex

This writes the server to ~/.codex/config.toml; codex mcp list shows it enabled.

[mcp_servers.loomtracer]
command = "loomtracer"
args = ["mcp", "--agent-tool", "codex"]

The loomtracer MCP server exposes 40 tools. See MCP Tools reference for full parameter documentation.

ToolTypeDescription
capture_decisionwriteRecord a decision with rationale and options
capture_evidencewriteRecord supporting evidence for a decision
capture_hypothesiswriteRecord a hypothesis still in flight
disagree_decisionwriteContest a decision as an actor
supersede_decisionwriteSupersede a prior decision with a new one
move_decisionwriteMove a decision to another project, recorded with who, when, from, to and why
retitle_decisionwriteRename a decision; the old title stays readable
link_supersessionwriteRecord that one already-recorded decision replaces another, keeping its own decider
ground_decisionwriteSay what an existing decision rests on, after the fact
reaffirm_decisionwriteSay a decision still holds after something it rests on changed
request_decisionwriteAsk a question before any decision answers it
acknowledge_suggestionwriteRecord that a finding from get_suggestions was looked at
get_decisionreadRetrieve a decision by ID with derived status
get_relevant_decisionsreadSearch by topic, status, actor, or time window
get_situational_decisionsreadDecisions bearing on a situation — matches by topic or evidence overlap, no ID needed
get_supersession_chainreadWalk the full supersession history backward
get_decision_neighborhoodreadOne-hop graph neighborhood around a decision — actors, options, evidence, hypotheses
search_decisionsreadFull-text search across the ledger
recent_decisionsreadList recently proposed decisions
dump_graphreadExport the full projected graph (DOT or JSON)
get_decision_outcomereadWhether a decision still holds — superseded, stale premises, or contested
get_reexamine_itemsreadDecisions resting on something that changed that nobody has looked at again
get_waiting_requestsreadQuestions asked that no decision answers yet, oldest first
get_contested_decisionsreadDecisions in contest: who accepted and who rejected, or conflicting answers
get_changed_decisionsreadDecisions revised or superseded in a window, with what changed and who changed it
get_open_why_decisionsreadBinding decisions whose why is still to answer
loomtracer_compact_viewlayer-3Compact summary of a decision and its context
recall_decisionslayer-3Search + ranked digest — answers “what was decided about X?”
summarize_decisionslayer-3Concise text summary of one or more decisions
decision_quality_candidateslayer-3Bulk quality-signal pull for external scorers
get_decision_contextlayer-3Context record for a decision — authorship, review depth, evidence richness
decision_context_candidateslayer-3Bulk context-feature pull, the independent-variable side of quality scoring
score_decisionlayer-3The quality profile of one decision: seven dimensions, no score, no tier
scan_decision_qualitylayer-3Attention findings: decisions that need a look, such as a bet past its check date
get_suggestionslayer-3The attention findings nobody has acknowledged yet
rank_decisions_by_importancelayer-3Which decisions carry the most impact, each with who decided it
scan_misfiled_decisionslayer-3Flag decisions carrying a caller-named “foreign” topic key — report only, never moves anything
analyze_failure_modeslayer-3Aggregate failure-rate patterns across authorship, review, and context
classify_queue_listreadList pending ingest batches with turn text, plus today’s classification budget (HTTP transport only)
classify_queue_submitwriteSubmit captures for one or more pending batches as one classification event (HTTP transport only)

Who a capture is filed under (local stdio)

Section titled “Who a capture is filed under (local stdio)”

A capture call may name its actor in actor_id; it is optional. Without one, the server records agent:<tool>:mcp-<session-id>, and a capture that names no project lands in that agent’s personal project. <tool> comes from --agent-tool on the server’s command line, so name your client there:

loomtracer mcp --agent-tool claude # Claude Code
loomtracer mcp --agent-tool codex # Codex
loomtracer mcp --agent-tool cursor # Cursor

Without the flag the server usually cannot tell which client started it, and files captures under codex. The server records source=agent and a per-session source_ref for every write.