Suasor

Gathers, remembers, advises - you decide.

Suasor is a local-first AI secretary. It gathers your scattered work context - chat, email, calendar, documents, code, the web - into private memory on your own machine, so you and your AI agents can search and summarize it over MCP. It advises you, and proposes replies, tasks, and decisions - and, once you approve, can act on your behalf (publish a task, transition an issue). Ingest is read-only, and nothing egresses without your approval.

日本語 / Japanese →

What it does

• Gathers — pulls your scattered work context from across your tools into one local, private store. Ingest is read-only — gathering never writes back to your sources.

• Remembers — keeps it as searchable, queryable memory on your own machine.

• Advises — surfaces, summarizes, and proposes replies, tasks, and decisions over MCP. You and your AI agents query it; you approve every action. Once approved, Suasor can also carry it out for you — publish a task to GitHub / Jira, transition an issue — but nothing egresses without your say (ADR-0036).

Related MCP server: OmniHub

What it is not (Boundaries)

These boundaries keep Suasor a local-first, human-in-the-loop advisor (see docs/requirements/scope.md):

• No unapproved egress, no auto-send — ingest is read-only, and nothing leaves your machine without your explicit approval. Approved actions (e.g. publishing a task or transitioning an issue via task.publish / task.act) are then carried out by Suasor on your behalf — only after you approve them, never automatically (ADR-0004 / ADR-0036).

• No daemon, no unsolicited notifications — nothing runs always-on. Proactive digests exist (suasor digest), but only as an OS-scheduled cron one-shot that sends a preconfigured, named job (standing consent) — with no configured job it sends nothing, and per-event write approval is unchanged (ADR-0040 / ADR-0004).

• No heavy in-process ML — model training/inference is delegated, not run in-process (ADR-0006).

• Single-user, local-only — no multi-user, team sharing, or server-side aggregation.

• No web / mobile UI — the boundary is the CLI and MCP.

Status

Early development — published on npm / standalone binaries / Docker. Built spec-first.

Install

Suasor is an MCP server — an application, not a library — so it runs on its own runtime, Bun. Pick a channel by whether you already use Bun; the binary and Docker image need no runtime at all (Bun is bundled). Details: docs/guide/install.md.

• Standalone binary (no runtime needed) — download per OS/arch from Releases. Bun is compiled in. Core + a few native bits; the heavier connector SDKs are external (use npm/Docker for the full connector set).

• Docker (batteries-included + Ollama) (no runtime needed) — docker run ghcr.io/ozzy-labs/suasor. Local embedding with no external egress.

• npm — for Bun users — bunx @ozzylabs/suasor mcp serve (or bun add -g @ozzylabs/suasor). Requires Bun ≥ 1.2 (install Bun) — uses bun:sqlite, so npx/Node won't run it; pnpm/npm can fetch it but Bun runs it. OIDC-published with provenance.

• MCP registry — discoverable via server.json.

Published on npm / binaries / Docker. Contributors can also run from source.

Quickstart (provisional)

Early development, but every CLI command below is implemented (ingest, retrieval, MCP server, and skills all work), and the MCP surface — including brief and graph.related / graph.expand — is shipped. See docs/design/mcp-surface.md.

These commands assume Suasor is installed via one of the channels above, so suasor is on your PATH. Pick the form that matches your install:

The examples below use the suasor <cmd> form. Working from a clone instead? See From source.

suasor --version

# First-run setup: writes ~/.config/suasor/config.toml and the local SQLite store.
# Prints a multi-step next-steps guide (doctor -> onboard -> sync -> skills).
suasor init

# Guided setup: pick connector(s), store tokens, wire the [connectors.X] config
# slice (enabled = true), run the first sync, and print the scheduler + MCP
# snippets — all in the correct order (ADR-0029).
suasor onboard --connector github,slack   # interactive on a TTY; --json for a summary
# slack now completes here too: onboard bridges its flat/single-workspace setup
# (multi-workspace still uses `suasor slack auth set --workspace <alias>`).

# Verify config / DB / connector readiness (diagnostic only; creates nothing).
suasor doctor

# Ingest read-only from a connector (github / slack / ms-graph / google / box / web / local / notion / jira).
suasor github sync

# Or ingest from every enabled connector in one read-only pass (one-shot).
suasor sync                   # --connector a,b / --json available

# Full-text search over ingested sources (FTS5; --json / --limit available).
suasor search "<query>"

# Install the bundled assistant skills into your agent host(s).
suasor skills install        # .claude/skills/ + .agents/skills/
suasor skills list           # installed / missing / modified

# Maintenance.
suasor db migrate            # apply the projection schema (idempotent)
suasor projections rebuild   # replay the event log into projections
suasor export backup         # consistent store backup (--format sqlite|tgz)
suasor config edit           # edit config.toml in $EDITOR, validate on save
suasor validate-config       # check config.toml (--fix applies safe repairs)

Config lives in ~/.config/suasor/ (override with SUASOR_CONFIG_DIR). Edit it with suasor config edit (validates on save, rolls back a bad edit) and check it with suasor validate-config [--fix]. <connector> sync ingests read-only from github / slack / ms-graph / google / box / web / local / notion / jira — see docs/guide/connectors.md for per-connector setup. Back up your local store with suasor export backup and audit / purge ingested data with suasor source list / suasor source forget — see docs/guide/data-audit.md. Diagnose common failure modes (empty sync, recall returning nothing, dimension mismatch, rate limits) with docs/guide/troubleshooting.md. See docs/design/cli.md for the full command/flag reference and docs/skills/README.md for the assistant skills.

From source

Contributors and anyone running from a clone use Bun directly — bun run src/index.ts replaces suasor in every command above. Requires Bun 1.2+.

git clone https://github.com/ozzy-labs/suasor.git
cd suasor
bun install                          # install dependencies
bun run src/index.ts --version

bun run src/index.ts init            # same first-run setup as `suasor init`
bun run src/index.ts doctor          # same diagnostics as `suasor doctor`
bun run src/index.ts sync            # same bulk ingest as `suasor sync`

bun run dev is a shorthand for bun run src/index.ts. See AGENTS.md for the development and verification workflow (bun test / bun run typecheck / lint). CI (.github/workflows/ci.yaml) is the source of truth for quality gates: typecheck + test (with coverage) + build, Biome + markdownlint, and security scans (gitleaks / Trivy / actionlint) — so PRs that bypass local hooks are still guarded.

Periodic sync

suasor sync ingests from every enabled connector in one short-lived, idempotent pass (read-only, continue-on-error, exit 1 if any connector failed). Suasor runs no daemon — schedule it with your OS scheduler (cron / launchd / systemd timer):

# Hourly bulk sync via cron; gate on the exit code, log the JSON output.
15 * * * * suasor sync --json >> "$HOME/.local/state/suasor/sync.log" 2>&1

See docs/guide/scheduling.md for launchd / systemd timer examples and failure monitoring (ADR-0027).

The same cron model drives the proactive push lane (ADR-0040): once you configure a named [digest.jobs] entry (standing consent), suasor digest bundles your top priorities (overdue / demand / due-soon, ADR-0041) and pushes them to a channel — OS notification, a file in the export sandbox, or a Slack DM-to-self. No job configured means nothing is sent.

# Every morning, write a file digest; hourly on weekdays, DM the urgent items.
0 8 * * *  suasor digest --job morning >> "$HOME/.local/state/suasor/digest.log" 2>&1

Connect an agent host (MCP)

Suasor exposes its memory to AI agents over the Model Context Protocol (stdio transport). The server is the agent boundary. Read tools are side-effect-free and annotated read-only so hosts may auto-approve them; write tools stay behind human-in-the-loop approval (ADR-0004) — nothing is applied or egressed without your say. The full surface, generated from src/mcp/tool-catalog.ts so this list can't drift:

Read tools — side-effect-free (readOnlyHint: true), so hosts may auto-approve them:

• search — FTS5 full-text search over ingested sources.

• recall.search — Semantic (embedding) search; degrades to FTS when no backend is enabled.

• search.hybrid — Hybrid search: RRF fusion of FTS + semantic hits; degrades to FTS-only.

• source.list — List ingested sources newest-first.

• source.get — Fetch one source (with body) by id.

• source.get.full — Bundle a source's body + outgoing provenance links + extraction_meta in one call.

• source.history — List a source's body versions from the event log (newest first).

• task.list — List tasks, most-recently-updated first.

• decision.list — List recorded decisions, newest-recorded first.

• demand.list — List connector-neutral demand (Slack @mentions/DMs + github notifications); un-acked only by default (ADR-0041).

• priority.list — Deterministic cross-entity next-actions ranking (tasks + commitments + un-acked demand, ADR-0041).

• brief — Bundle the period's tasks/decisions/sources/inbox for the host to summarize.

• graph.related — Provenance neighbours of an entity (1 hop) over the links projection.

• graph.expand — Breadth-first provenance expansion from an entity (N hops); direction in/out/both for backward trace.

• activity.timeline — Entity-axis merged source/task/decision timeline (newest-first) for one entity.

• inbox.list — List inbox items, most-recently-updated first.

• propose.list — List proposal candidates by state (pending/applied/rejected).

• commitment.list — List commitments by state (open/resolved/dismissed) and direction.

• person.list — List resolved persons with their connector author identities (ADR-0022).

Write tools — every one is HITL: a host must gate it behind human approval, and there is no auto-apply path (ADR-0004). The set includes actuators that carry out an approved action on your behalf — task.publish / task.act / task.update egress to your GitHub / Jira / Slack task home (ADR-0036) — and source.forget, which irreversibly purges an ingested source. Suasor never triggers any of these on its own; you approve each one first:

• connector.sync — Run a read-only connector ingest pass into the local store.

• propose.generate — Frame reply/task/decision/triage candidates and record them as pending.

• propose.apply — Persist approved candidates as domain events (idempotent).

• propose.reject — Reject a pending candidate with a reason (idempotent).

• proposal.feedback — Record a regeneration hint on a pending candidate without applying/rejecting it.

• propose.batch — Apply and/or reject candidates in one atomic RPC (single transaction).

• task.create — Create a task directly (TaskProposed).

• task.update — Transition a task's lifecycle state (TaskApplied).

• task.publish — Publish a task to its external home (TaskPublished, egress).

• task.act — Act on a published task: complete/reopen/comment (TaskActionIssued).

• decision.record — Record a decision directly (DecisionRecorded).

• inbox.add — Capture an inbox item referencing a source (InboxItemTriaged, state open).

• inbox.triage — Resolve an open inbox item (task / decision / discard).

• link.add — Create a manual provenance link between two entities (LinkAdded, manual_link).

• link.remove — Remove a manual link by id (LinkRemoved).

• person.merge — Merge two persons into one (PersonsMerged); reversible via person.split.

• person.split — Split one identity off a person into another (PersonSplit).

• commitment.resolve — Mark an open commitment fulfilled (CommitmentResolved).

• commitment.dismiss — Dismiss an open commitment as a false-positive (CommitmentDismissed).

• commitment.reopen — Reopen a resolved/dismissed commitment back to open (CommitmentReopened).

• demand.ack — Mark a demand row handled — drops it from the un-acked demand.list (DemandAcknowledged).

• demand.dismiss — Mark a demand row not relevant — drops it from the un-acked demand.list (DemandDismissed).

• draft.export — Write a draft to a local file in the export sandbox (DraftExported).

• source.forget — Purge an ingested source locally — redact + delete + tombstone (SourceForgotten).

• source.unforget — Lift a forget tombstone so the source can be re-ingested (SourceUnforgotten).

suasor mcp serve                 # start the MCP server over stdio
# from source: bun run src/index.ts mcp serve

Register it with an MCP host (Claude Code, Claude Desktop, Codex CLI, …). For Claude Desktop, add to claude_desktop_config.json. With a global install (bun add -g @ozzylabs/suasor, so suasor is on PATH and resolves to Bun):

{
  "mcpServers": {
    "suasor": {
      "command": "suasor",
      "args": ["mcp", "serve"]
    }
  }
}

No Bun on the host? Point it at the Docker image instead (no runtime needed):

{
  "mcpServers": {
    "suasor": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "-v", "suasor-data:/data", "ghcr.io/ozzy-labs/suasor:latest"]
    }
  }
}

Semantic search (recall.search) returns an embedding_disabled signal until you enable an embedding backend, so hosts gracefully fall back to FTS search (ADR-0005). See docs/design/mcp-surface.md for the tool schemas.

License

MIT