# Vouch > A git-native, review-gated knowledge base for AI agents: they propose writes, you approve them. Every claim cites a source, every change is a diff in your repo. MCP + CLI. - **Type:** MCP server - **Install:** `agentstack add mcp-vouchdev-vouch` - **Verified:** Pending review - **Seller:** [vouchdev](https://agentstack.voostack.com/s/vouchdev) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [vouchdev](https://github.com/vouchdev) - **Source:** https://github.com/vouchdev/vouch - **Website:** https://vouchai.dev ## Install ```sh agentstack add mcp-vouchdev-vouch ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # vouch **Git-native, review-gated knowledge base for LLM agents. MCP server + JSONL tool server + CLI.** > Agents should not start every session with amnesia — but they shouldn't get to write whatever they want either. `vouch` is a knowledge base for LLM agents with an explicit **review gate**: agents *propose* writes; humans *approve* them with `vouch approve`. Approved artifacts are plain files on disk — YAML for claims, markdown for pages — so the KB lives in your repo, is reviewed in PRs, diffs cleanly, and can be exported as a portable bundle. Still alpha — surface is small on purpose; expect breaking changes pre-1.0. > **Featured for Gittensor (SN74).** vouch ships a one-command starter pack for **Gittensor** — Bittensor subnet 74. `vouch init --template gittensor` seeds a cited, already-approved decision-memory of SN74's scoring model: merged-PR rewards, PAT verification, sybil-resistance, the repo allow-list, the issue multiplier, and the emission split. It's the durable *why* behind each rule — reviewed, cited, and committed alongside your code. → **[docs/gittensor.md](docs/gittensor.md)** ## Why this exists Three opinionated choices distinguish vouch from the neighbours: 1. **The KB is a folder in your repo.** Git is your audit log, your backup, and your sync mechanism. PRs are your review UI. 2. **Writes require approval.** Agents file *proposals*; a human (or trusted approving agent) explicitly accepts them. `proposed/` is gitignored, so rejected drafts never pollute history. 3. **Claims must cite sources.** A claim without at least one Source / Evidence id is a validation error, not a warning. Sources are content-hashed; the same evidence registered twice de-duplicates. ## When to use vouch Worth it when: - **You run or contribute to a Gittensor (SN74) repo.** Scoring weights, the repo allow-list, anti-sybil thresholds, and emission splits get debated across PRs, Discord, and validator changes — then settle into nobody's notes. `vouch init --template gittensor` gives you a cited, reviewed record of *why* each rule exists and what it superseded. See [docs/gittensor.md](docs/gittensor.md). - **Multiple agents share a repo** (Claude Code + Cursor + a CI bot). Per-agent attribution in the audit log makes "which agent claimed what" answerable. - **Sessions keep re-explaining the same context.** Curated, cited claims let new sessions start from your team's agreed answer instead of re-grepping. - **You want decision records without the ADR ceremony.** `vouch crystallize` promotes a session's durable parts into proposals; one approve and they're permanent. - **You'd review agent-stated facts the way you review agent-written code.** Vouch is a PR queue for claimed knowledge. - **Compliance / audit trails matter.** Required citations + append-only audit log give you "who decided X, citing what, when" for free. Skip it if: - Solo hobby project where context fits in your head. - Short-lived branches with no compounding context. - Your team won't actually review proposals — without the gate, vouch is a worse note app. ## Install ```bash # one-liner (Linux + macOS) — picks a Python, ensures pipx, installs vouch-kb curl -fsSL https://raw.githubusercontent.com/vouchdev/vouch/main/install.sh | sh # …or directly via pipx (vouch-kb on PyPI; the command stays `vouch`) pipx install vouch-kb # …or from the cloned repo, in a venv pip install -e '.[dev]' ``` The one-liner is POSIX `sh`, never needs `sudo`, and detects an existing Claude Code install to point you at the next step (`vouch install-mcp claude-code`). Inspect it first if you'd like — it's [`install.sh`](install.sh) at the repo root. ## Quick start ```bash # 1. set up a KB at your project root vouch init # 2. as agent, register evidence + propose claims (via MCP/JSONL server) vouch serve # MCP over stdio (Claude Code, Cursor, Codex) vouch serve --transport jsonl # newline-delimited JSON over stdin/stdout # 3. as human, review and decide vouch status # one-line summary vouch pending # list pending proposals vouch show # full details vouch approve # → durable artifact vouch reject --reason "..." vouch expire --apply # optional: clear stale pending proposals # 4. commit git add .vouch/ && git commit -m "kb: approve auth-uses-jwt" ``` ## 30-second tour `vouch init` now creates a starter config plus one cited example claim, so you can try the loop before wiring an agent: ```bash mkdir vouch-demo && cd vouch-demo vouch init vouch status vouch search agent vouch cite vouch-starter-reviewed-knowledge ``` The starter claim is already durable and cites the starter source. Replace it with your project's first real source and claim when you are ready. The full captured walkthrough lives at [docs/example-session.md](docs/example-session.md); re-render the GIF from [docs/demo.tape](docs/demo.tape) with `vhs docs/demo.tape`. ## Gittensor (SN74) vouch's first domain template targets **Gittensor** — Bittensor subnet 74, which rewards open-source contribution by rule. Its scoring model evolves across PRs, Discord, and validator changes, and the rationale usually lives in people's heads. vouch is the durable, cited memory for it: ```bash cd your-gittensor-repo vouch init --template gittensor # seeds 1 source, 1 entity, 7 cited claims about SN74 scoring vouch status # durable: 7 claims · 1 source · 1 entity vouch search "emission split" git add .vouch && git commit -m "chore: add vouch decision-memory KB" ``` The seeded pack covers merged-PR rewards, PAT verification, scoring factors, sybil-resistance, the repo allow-list, the issue-solving multiplier, and the emission split — each a cited, approved, supersede-able claim. When a rule changes, `vouch supersede` the old claim with the new one so the history of *what changed* stays queryable. vouch stores **no** live signals — it is not a validator or miner client and never reads on-chain scores. It is the institutional memory that sits beside the live layer (Gittensory). The seeded claims are starter-grade; `vouch supersede` them with the real spec or PR once you confirm the live rule. Full adoption guide — install, seed, wire the MCP server, capture decisions as cited claims: **[docs/gittensor.md](docs/gittensor.md)**. ## Object model ``` Source immutable input material (file, URL, transcript, commit, …) content-addressed by sha256 Evidence span pointer into a Source (line range, timestamp, quote) Claim atomic durable assertion type: fact | decision | preference | workflow | observation | question | warning status: working | actionable | stable | contested | superseded | archived | redacted confidence: 0.0–1.0 cites: list of Source / Evidence ids Entity typed thing (person | project | repo | concept | decision | system | …) Relation typed edge (uses | depends_on | supersedes | contradicts | implements | …) Page maintained markdown — entity write-up, decision record, session summary Session work block opened by an agent; bundles its proposals AuditEvent append-only log record for every mutation ``` ## File layout After `vouch init`, your repo contains: ``` .vouch/ ├── config.yaml # KB settings ├── .gitignore # ignores proposed/, state.db ├── audit.log.jsonl # append-only audit (committed) ├── state.db # SQLite FTS5 index (derived; not committed) ├── claims/.yaml # reviewed claims (init seeds vouch-starter-*.yaml) ├── pages/.md # markdown pages with YAML frontmatter ├── sources//{meta.yaml,content} ├── entities/.yaml # graph nodes ├── relations/.yaml # graph edges ├── evidence/.yaml # span pointers into sources ├── sessions/.yaml # agent session records ├── proposed/.yaml # pending (gitignored, local-only) └── decided/.yaml # approved/rejected (committed for audit) ``` The files are the source of truth; `state.db` is a derivable cache (`vouch index` rebuilds it). ## CLI surface ``` vouch init # set up .vouch/ at PATH vouch discover [--path P] # find the nearest .vouch/ root vouch capabilities # emit the JSON capabilities descriptor vouch status [--json] # KB counts + pending proposals vouch stats [--days N] [--json] # observability: queue, review rates, citations vouch lint [--stale-days N] # user-actionable problems vouch doctor # full sweep incl. source verification vouch fsck # deep consistency: indexes, lifecycle, decided vouch migrate [--check] [--dry-run] # upgrade .vouch/ format safely vouch pending # list pending proposals vouch review [--limit N] [--type KIND] # guided proposal review queue vouch show vouch approve [--reason ...] vouch reject --reason "..." vouch expire [--apply] [--days N] [--json] # GC stale pending proposals vouch propose-claim --text ... --source ... [--type ...] [--confidence X] vouch propose-page --title ... [--body -] [--claim ID ...] vouch propose-entity --name ... --type ... vouch propose-relation --from ID --rel uses --to ID vouch source add PATH [--title ...] [--url ...] vouch source verify [--fail-on-issue] vouch supersede OLD_ID NEW_ID vouch contradict CLAIM_A CLAIM_B vouch archive CLAIM_ID vouch confirm CLAIM_ID vouch cite CLAIM_ID vouch session start [--task ...] [--note ...] vouch session end SESSION_ID vouch crystallize SESSION_ID [--no-page] vouch search QUERY [--limit N] vouch context TASK [--max-chars N] [--min-items N] [--require-citations] vouch index vouch audit [--tail N] [--json] vouch export --out path.tar.gz vouch export-check path.tar.gz vouch import-check path.tar.gz vouch import-apply path.tar.gz [--on-conflict skip|overwrite|fail] vouch sync-check PATH_OR_BUNDLE vouch sync-apply PATH_OR_BUNDLE [--on-conflict fail|skip|propose] vouch serve [--transport stdio|jsonl] ``` ## MCP tools / JSONL methods (same surface, two transports) Read (unrestricted): `kb.capabilities`, `kb.status`, `kb.search`, `kb.context`, `kb.read_{page,claim,entity,relation}`, `kb.list_{pages,claims,entities,relations,sources,pending}` Source intake (not gated — evidence is harmless and de-duplicates): `kb.register_source`, `kb.register_source_from_path`, `kb.source_verify` Write (gated → produce proposals): `kb.propose_{claim,page,entity,relation}` (with `dry_run:true` for preview-only) Decisions: `kb.approve`, `kb.reject` (host trust required) Lifecycle (metadata about reviewed knowledge — direct mutation, audited): `kb.supersede`, `kb.contradict`, `kb.archive`, `kb.confirm`, `kb.cite` Sessions: `kb.session_start`, `kb.session_end`, `kb.crystallize` Maintenance: `kb.index_rebuild`, `kb.lint`, `kb.doctor`, `kb.audit`, `kb.export`, `kb.export_check`, `kb.import_check`, `kb.import_apply` ## Wiring into Claude Code In your project's `.mcp.json`: ```json { "mcpServers": { "vouch": { "command": "vouch", "args": ["serve"], "env": { "VOUCH_AGENT": "claude-code" } } } } ``` `VOUCH_AGENT` is recorded as `proposed_by` and as the actor on every audit event, so multi-agent setups can attribute writes correctly. ## Running vouch as an OpenClaw plugin Vouch ships an [OpenClaw](https://github.com/dripsmvcp/openclaw) plugin manifest at the repo root — [`openclaw.plugin.json`](openclaw.plugin.json). Drop the vouch repo into an OpenClaw deployment and the plugin loader picks it up automatically: the MCP server, the four slash commands (`/vouch-recall`, `/vouch-status`, `/vouch-resolve-issue`, `/vouch-propose-from-pr`), and the CLAUDE.md fenced snippet become available as one bundle. The manifest declares vouch's trust boundary explicitly — remote callers' filesystem access is confined, every write tool routes through the review gate, every lifecycle op is audit-logged. The `configSchema` exposes only `kb_path`, `agent`, and `transport` — no API keys, no secrets; vouch is local-first. ```bash # Inside an OpenClaw deployment that vendors plugin repos: openclaw plugin add vouchdev/vouch openclaw plugin enable vouch ``` The plugin's `mcpServers.vouch` block matches the same `.mcp.json` shape Claude Code uses — both platforms drive the same `vouch serve` process, so the kb.* surface is identical regardless of host. ## JSONL request/response shape The JSONL transport reads one envelope per line on stdin, writes one per line on stdout: ```jsonl {"id":"r1","method":"kb.search","params":{"query":"jwt","limit":5}} {"id":"r1","ok":true,"result":[{"kind":"claim","id":"auth-uses-jwt","snippet":"…","score":1.2,"backend":"fts5"}]} ``` Errors come back with `ok:false` and a structured `error.code` (`method_not_found`, `missing_param`, `invalid_request`, `internal_error`). Every successful `kb.*` result that is object-shaped carries read-only trust metadata so clients can detect remote confinement: ```json { "id": "r1", "ok": true, "result": { "backend": "fts5", "hits": [], "_meta": { "vouch_trust": { "remote": false, "caller_kind": "jsonl", "auth_subject": null } } } } ``` | Transport | `remote` | `caller_kind` | `auth_subject` | |-----------|----------|---------------|----------------| | JSONL stdio | `false` | `jsonl` | `null` | | HTTP `/rpc` | `true` | `jsonl_http` | bearer fingerprint when authenticated | | MCP stdio | `false` | `mcp_stdio` | `null` | | HTTP `/mcp` | `true` | `mcp_http` | bearer fingerprint when authenticated | | CLI `--json` | `false` | `cli` | `null` | The block is server-attached metadata — client mutations are ignored. Array-shaped read results (e.g. `kb.list_claims`) pass through unchanged; trust rides on dict-shaped responses only (#233). ## Portable bundles ```bash vouch export --out kb.tar.gz # tar.gz + manifest.json with per-file sha256 vouch export-check kb.tar.gz # verify every file against the manifest vouch import-check kb.tar.gz # diff against destination — new / conflict / identical vouch import-apply kb.tar.gz --on-conflict skip # apply (default skip; never destructive without overwrite) ``` `proposed/`, `state.db`, and `audit.log.jsonl` are excluded from bundles — only committable artifacts travel. ## Compared to neighbours | | mem0 / Letta | LLM-Wiki tools | **vouch** | |---|---|---|---| | Knowledge lives in | a service | filesystem | your **repo** | | Review of writes | none | none | **explicit `approve`** | | Evidence required | no | optional | **enforced** | | Per-agent attribution | partial | none | **yes** (audit log) | | Graph (entities + relations) | no | no | **yes** | | FTS5 search | no | varies | **yes** | | Portable bundle | no | no | **yes** | | Transports | SDK / HTTP | none | **MCP + JSONL** | ## What ships today | Area | Current support | |------|-----------------| | Knowledge base | `.vouch/` folder, YAML claims/entities/relations/evidence/sessions, markdown pages with frontmatter, JSONL audit log, content-addressed sources | | CLI | `init`, `discover`, `capabilities`, `status`, `lint`, `doctor`, `fsck`, `pending`, `show`, `approve`, `reject`, `propose-{claim,page,entity,relation}`, `source add`, `source verify`, `supersede`, `contradict`, `archive`, `confirm`, `cite`, `session {start,end}`, `crystallize`, `search`, `context`, `index`, `audit`, `export`, `export-check`, `import-check`, `import-apply`, `serve` | | Tool servers | MCP over stdio + JSONL over stdin/stdout, same `kb.*` surface across both transports, capabilities + knowledge-capability descriptor | | Schemas | 13 JSON Schemas (Draft 2020-12) generated from pydantic … ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [vouchdev](https://github.com/vouchdev) - **Source:** [vouchdev/vouch](https://github.com/vouchdev/vouch) - **License:** MIT - **Homepage:** https://vouchai.dev Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** yes - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: flagged — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-vouchdev-vouch - Seller: https://agentstack.voostack.com/s/vouchdev - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.