Solumbe MCPサーバー

Solumbe

Models generate the change. Solumbe proves whether it is safe to merge.

Formerly Òtítọ́ (@bashbop/otito), renamed in 4.0.0; the CHANGELOG maps every old name to its new one.

Solumbe is a local-first, deterministic, model-agnostic trust layer for AI-assisted development. It builds task-aware repository context before an agent edits, scores how much a change actually touches, and gates merge readiness against the exact staged tree, with no server, no account, and no code leaving the machine.

It does not replace Claude Code, Codex, Cursor, Gemini, or any native agent harness. It runs beside them and keeps working as the models change underneath.

A passing local gate is never an automatic merge approval: hosted CI, GitHub review, CODEOWNERS, and the human release decision remain separate authorities.

Install

npm install -g @bashbop/solumbe
solumbe doctor
solumbe install --host all

The last command connects Solumbe's MCP server to every agent host it finds: Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code, Gemini CLI and Kimi Code CLI. Add --dry-run to see the plan first, or name hosts with --host claude-code,cursor. Restart each host afterwards.

Or without installing: npx -y @bashbop/solumbe doctor.

What it does

Rank what a change actually touches, from the request alone, with no model, no embeddings, and no network:

$ solumbe impact . "add refund handling to checkout" --top 3

  concepts: money flow

   1.  src/payment/checkout.service.ts    score 155
       |- role: required
       |- path matches: checkout
       |- symbol matches: checkout, refund
       |- concept match: money flow (×1.4 + 6.0)
       |- risk: money flow

   2.  src/payment/stripe.webhook.ts    score 6
       |- role: advisory
       |- concept match: money flow (×1.4 + 6.0)

Gate the exact staged tree, and say precisely why:

$ solumbe gate . --staged --base origin/main --request "add refund handling to checkout" --min-convergence 80

  [OK]    Changed files          1 changed file found.
  [OK]    Staged snapshot        Changed-file scope and convergence evidence are captured from the exact staged Git tree.
     |- Tree: 7d0b513dd256877d6bd309a6406468a9b1eaeedc
     |- Base: b776f04fcfbd02c8ce8d63e1a0c2450be58de039
  [OK]    Secret safety          No secret file paths and no credential values found in the changed content.
  [WARN]  Risk review            Risk-sensitive files changed; maintainer review should be explicit.
     |- src/payment/checkout.service.ts
  [OK]    Convergence            Task and diff satisfy the convergence requirement.
     |- Score: 100/100 (aligned)
     |- Receipt handle: rcpt_a8d31efae06e
     |- Subject tree: 7d0b513dd256877d6bd309a6406468a9b1eaeedc

  VERDICT  WARN

The convergence score is the part a model cannot grade for itself: it compares the stated intent against the files the diff actually changed, and produces a receipt bound to the exact base, parent, and staged-tree identity.

Hold the change to what was declared. Record the files a request is expected to touch before the edit, and the gate names any file that was never declared:

$ solumbe declare . "add refund handling to checkout"
$ # ... the agent edits ...
$ solumbe gate . --staged --base HEAD --intent .solumbe/intent.json

  [FAIL]  Scope contract         Touched `src/auth/session.ts`, which was never declared. Amend the intent with the reason it belongs (`solumbe amend <file> --reason "<why>"`), or take it out of the change.
     |- Request: "add refund handling to checkout"
     |- Contract: fe655c2af2b31b7e3edec6aedd248b379235811d1228c829f87ff73796ace487
     |- Changed files: 3 (1 declared, 0 amended, 1 implied, 1 undeclared)
     |- Implied: tests/checkout.service.test.ts (owner-test of src/payment/checkout.service.ts)
     |- Undeclared on a risk-sensitive path: src/auth/session.ts [auth/security]

An amendment adds a file with its reason, hashed onto the contract, so a reviewer reads why the scope grew. Over MCP it is change_impact with declare: true, and review_gate with intent.

The core

Four commands are the product. They call no model, open no socket, and read nothing outside the repository.

Step	Command	What it proves
Before the edit	solumbe context "add refund handling" --path .	Which files, tests and commands the request needs
After the edit	solumbe converge . "add refund handling" --base origin/main --staged	How far the diff is from the request, with a receipt
Before the merge	solumbe gate . --staged --base origin/main --request "add refund handling"	PASS / WARN / FAIL on the exact staged tree
After the merge	solumbe attest . --verdict verdict.json --merge <sha>	A hash-chained record of what shipped, under which verdict

solumbe help shows these and how to set up; solumbe help --all lists every command, including the ones that support them (impact, declare, pr, review, map, calibrate, route, ax, dashboard and the rest). Nothing was removed.

Nothing hosted is required, and nothing hosted can change a verdict. Local Core, Optional Hosted lists exactly what the optional pieces send, and where.

Supporting commands

Goal	Command
Rank change blast radius	solumbe impact . "add refund handling" --top 12
Score intent vs. execution	solumbe converge "add refunds" --path . --base HEAD --staged
Score a committed change exactly	solumbe converge "add refunds" --path . --base HEAD~1 --head HEAD
Gate a product change across repos	solumbe workspace-gate ../web ../api --request "ship change"
Grade risk flags against history	solumbe calibrate . --window 30
Grade route tiers against history	solumbe regret . --window 30 --offline
Regrade a saved run, same answers	solumbe regret --rescore run.json
Score Agent Experience	solumbe ax . "add a new MCP tool"
Recommend a model tier	solumbe route . "add a new MCP tool"
Sharpen context with a model read	solumbe context "add a new MCP tool" --path . --online
Inspect one repo	solumbe repo . --json
Build a code map	solumbe map . --json
Index and search local projects	solumbe index ~/projects --discover then solumbe search "events controller"
Generate an agent harness	solumbe harness . --out .solumbe/harness.md
Run the MCP server	solumbe mcp

Every command takes --json, and solumbe help lists the full set with flags.

MCP

Solumbe ships a stdio MCP server exposing 14 tools: repo_inspect, repo_map, repo_index, repo_search, context_pack, change_impact, agent_experience, model_route, convergence_score, review_context, review_gate, review_verdict, workspace_report, and repo_harness.

{
  "mcpServers": {
    "solumbe": {
      "command": "npx",
      "args": ["-y", "@bashbop/solumbe", "mcp"]
    }
  }
}

solumbe install --host all writes this entry for you, using absolute paths so hosts that do not inherit your shell PATH can still start it. Published in the MCP Registry as io.github.BASHBOP/solumbe. Repo-map lookups use an external per-user cache and never write into the inspected repository. Host-specific setup for Claude Code, Claude Desktop, Codex, Cursor, VS Code, Gemini CLI, and Kimi Code is in MCP and Agent Workflows.

How it compares

Approach	Strengths	Where solumbe differs
Sourcegraph / Cody context	Powerful hosted code search and embedding-based context across an org	Local-first and deterministic: no server, no account, no code leaves the machine, and the same query always yields the same packet
Hand-written CLAUDE.md / rules files	Curated, intent-rich guidance	Hand-written context goes stale; solumbe regenerates context from the actual code (symbols, imports, routes, tests) on every run and complements a short CLAUDE.md
grep / ripgrep	Fast, universal text matching	solumbe ranks whole files by task intent across paths, symbols, exports, and tests, then adds patterns and validation commands, producing a context packet rather than a list of matching lines

Documentation

Full command reference, agent workflows, release process, evaluation method, and the design theses behind the trust layer:

bashbop.github.io/solumbe

Contributing

git clone https://github.com/BASHBOP/solumbe.git
cd solumbe && npm ci && npm run ci

npm run ci is the full gate: format, lint, typecheck, version check, tests, coverage floors (70% lines / 60% branches / 75% functions), three evaluation corpora, dependency audit, and a packaged-tarball smoke test. Run it before requesting review.

Start with CONTRIBUTING.md and the Code of Conduct. All changes need maintainer review; main requires passing gates and resolved conversations. Solumbe follows Semantic Versioning, so say whether a PR is no-impact, patch, minor, or major.

---

Part of the toolchain

solumbe is one of four tools that form a deterministic trust layer for AI-assisted development. Each uses static analysis to answer a question people keep handing to an LLM.

• solumbe (this tool), for context: what does this change actually touch?
• tieline, for contracts: did the front end and back end quietly stop agreeing?
• bouncer, for compliance: could you defend this to Ofcom?
• aiglare, for governance: where can the model do something you can't undo?

More at segunolumbe.com. static analysis, never the model.