MCP tools

The orchestrator-mcp server exposes Spine to assistants like Claude Code and Codex — 34 tools total: 23 are read-only (the 8 comprehension tools plus 15 more), and 11 write, every one gated behind a human approval.

Install

pip install 'synaptixs-spine[all]'

Use [all] — [mcp] alone installs the server with a Python-only graph, so a Java, Go or C# repo would come back empty. Then register the plugin with your assistant and ask about your repo in plain language — the assistant picks the tool.

The registry has 34 tools: 23 read-only, 11 that write — broken down below.

Read-only comprehension (8)

These write nothing and need no credentials:

map_repo · blast_radius · explain_symbol · investigate · localize · regression_gaps · root_cause · docs_for

See Comprehension for the question each one answers.

Further read-only tools (15)

Beyond the 8 comprehension tools, 15 more are also read-only — some need no model (doctor, pkg_joins, read_memory_bank, pkg_grounding, ingest_preview, profile_repo, requirements_check, sdlc_run_status, sdlc_run_result, the operator tools registry_runs, registry_approvals and registry_trace), and a few spend tokens on a model call without writing anything (design_change with use_llm=true, sdlc_baseline, audit_repo).

Gated write tools (11)

The remaining 11 tools perform actions — branch, codegen, tests, PR, tracker and registry writes — and every one of them stops at a governance gate: understand_repo (writes only under episteme/, or nothing with check=true), sdlc_plan, sdlc_approve, sdlc_feature, requirements_answer, sdlc_start_run, sdlc_decide_gate, registry_decide, sdlc_address_review, sdlc_complete and sdlc_remediate. Nothing is written to the outside world without an explicit human approval.

Caution

Comprehension tools are safe to run anywhere — they write nothing. Tools that write state stop at a governance gate and do nothing external until a human approves. (The --safe/--live flags belong to the orchestrator CLI; over MCP the gate is the approval, not a flag.)