sandesh-relay 0.3.4

SQLite-backed multi-project messaging relay for cooperating agent/orchestrator sessions

Sandesh

संदेश (Sanskrit/Hindi: message · dispatch) — a tiny, standalone, multi-project messaging system for cooperating agent/orchestrator sessions. A SQLite-backed maildir + a mailbox watcher, pure Python stdlib (no third-party deps).

Latest release (see all releases) — the global store (one DB for all projects, cross-project messaging behind an admin grant, archive→tombstone lifecycle), inbox filters + FTS5 search (CLI/MCP/Pi), 12 MCP tools, and the Pi extension at full parity.

Built for the "Model-B" parallel-orchestration pattern (a Mainline coordinator + worker Track sessions that can't talk to each other directly), but project-agnostic.

📖 Must-read: User Guide — how to run cooperating sessions (MCP & Pi).

Why

Sessions can't message each other directly, and re-invoking a sleeping agent is only possible via its host's background-task mechanism. Sandesh provides the relay: a durable, queryable mailbox each session reads from, plus a blocking notify watcher a session runs in the background so it wakes when mail addressed to it arrives.

Model

table	holds
address	the addressbook — durable identities, '<Orchestrator> - <Project>'
message	the envelope — subject (required), kind, in_reply_to, body_path
message_recipient	per-message addressees — role (to/cc) + per-recipient read_at
notifier	per-session watcher liveness — pid, token, heartbeat, tombstone

Semantics: To wakes / Cc silent · all-tracks broadcast (minus sender) · per-recipient read (read = being acted on; reply = done) · subject-only ⟷ file-body (full absolute paths) · keep history · reply threading · crash-safe liveness (dead-pid / stale-heartbeat reap) · cooperative tombstone eviction · validated address format.

Layout

Sandesh is a standard Python package (sandesh/) installed via pyproject.toml; the two console scripts (sandesh, sandesh-mcp) land on $PATH. Runtime data lives under the XDG data dir — one global DB for all projects, plus a body folder per project:

$XDG_DATA_HOME/sandesh/                           (default ~/.local/share/sandesh/…)
├── sandesh.db                                    (the ONE global DB, WAL — all projects)
└── projects/<project_id>/
    └── messages/msg-<id>.md

Install

Distribution name: sandesh-relay (the import package + the sandesh / sandesh-mcp commands keep the name sandesh; sandesh was taken on PyPI). The bare install is the stdlib-only CLI + notify; the [mcp] extra adds the MCP server.

Sandesh is CLI + a chosen agent surface (MCP for Claude Code, or the Pi extension). The full per-route × per-surface walkthrough — install → sandesh init (provision) → manage (auto-migrate on update, admin) → uninstall, plus the uninstall matrix — lives in the install guide:

➡ Install & uninstall guide → docs/INSTALL.md

In short: pick a route (uv / pipx / pip / install.sh), install the sandesh-relay distribution with the extras for your surface ([mcp,migrate] for Claude, [migrate] for Pi), then provision with sandesh init (--check to verify). PyPI publishing is automated via OIDC trusted publishing — see RELEASING.md.

Use

sandesh setup --project Atlas
sandesh --project Atlas register --address "Mainline - Atlas" --kind mainline
sandesh --project Atlas register --address "Track 2 - Atlas"  --kind track

# send (subject-only ⇢ no file; --body/--body-file ⇢ md body)
sandesh --project Atlas send --from "Track 2 - Atlas" --to "Mainline - Atlas" \
        --subject "CR-308 started — chain unaffected"

# the watcher (run in the background; exits 0 with ids when 'to' mail lands)
sandesh --project Atlas notify --to "Mainline - Atlas"

# read (consolidates unread to+cc, marks read)
sandesh --project Atlas fetch --to "Mainline - Atlas"

# reply (threads under the parent; the reply IS the completion signal)
sandesh --project Atlas reply --to-msg 1 --from "Mainline - Atlas" --body "ack"

sandesh --project Atlas addressbook

$SANDESH_PROJECT and $SANDESH_ADDRESS default --project and the caller's own address. $SANDESH_POLL_SECONDS sets the watcher cadence (default 10, floor 3).

Cross-project messaging

All projects share the one global DB, but sending across projects is admin-gated: the sender's project needs a one-time grant — sandesh grant --cross-project --project <id> --by <admin> (revoke with sandesh revoke --cross-project …; project-wide, inherited by every participant of the granted project). Without it, the send fails with cross-project sending not approved for project '<id>' — ask the Sandesh admin. The admin is assigned only at install via $SANDESH_ADMIN (install.sh) — there is no CLI or MCP surface to set it. all-tracks broadcasts never cross projects, grant or not. sandesh projects shows each project's state and grant (PROJECT STATE CROSS-PROJECT).

Project lifecycle

A project moves active → archived → tombstoned — strictly two-step, with two-tier authorization:

# read-only freeze — reversible, deletes NOTHING (project's own Mainline only)
sandesh archive   --project Atlas --by "Mainline - Atlas"
sandesh unarchive --project Atlas --by "Mainline - Atlas"

# permanent retirement — ARCHIVED projects only (the install-assigned super-admin only)
sandesh tombstone --project Atlas --by <admin>          # prompts y/N; --yes to skip

• archive evicts the project's live notify watchers (cooperatively; --force reaps stragglers) and freezes it: sends from/to it and new registrations are refused, but every message, body, and thread stays fully readable. unarchive reverses it.
• tombstone is destructive and irreversible: it purges the project's internal messages (sender and all recipients inside the project) and deletes its projects/<id>/ body folder. What survives: cross-project envelopes (rows stay for audit + thread anchoring — their bodies are lost), and the tracker row itself as a permanent tombstoned marker (the project id is retired; setup refuses to reuse it). Afterwards inbox/fetch hide the tombstoned project's traffic, and thread marks holes with incomplete chain — message(s) removed (project tombstoned).
• Who may do what: archive/unarchive take the project's own Mainline as --by; tombstone takes only the super-admin assigned at install via $SANDESH_ADMIN. Without --yes, tombstone asks for interactive confirmation (and refuses when stdin is not a terminal).
• --dry-run (all three verbs) reports what would happen — watchers to evict, the would-be state, and for tombstone the purge counts (internal messages, body files, cross-project messages whose bodies would be lost) — and writes nothing. Guards still apply: a dry-run on the wrong state or with the wrong --by errors exactly like the real command.

MCP server

The verbs are also exposed as an MCP server (stdio) so an agent can call them as tools instead of shelling out. mcp is the only third-party dependency — it ships behind the optional [mcp] extra and is imported only by sandesh.mcp_server, so the CLI above stays pure-stdlib either way. Installing the [mcp] extra (see the install guide) provides the sandesh-mcp command; a base install without the extra still ships sandesh-mcp but it prints a one-line "install the [mcp] extra" hint and exits non-zero rather than crashing.

Register with Claude Code (stdio; bake a default project via env so tools can omit project_id):

claude mcp add sandesh --scope user --env SANDESH_PROJECT=<id> -- sandesh-mcp

That writes an mcpServers entry (~/.claude.json for user scope, or a committed .mcp.json for --scope project). Manage with claude mcp list / claude mcp get sandesh / claude mcp remove sandesh; the in-session /mcp panel shows status + tools.

Twelve tools, each taking project_id (falls back to $SANDESH_PROJECT): sandesh_setup, sandesh_register, sandesh_unregister, sandesh_addressbook, sandesh_send, sandesh_reply, sandesh_inbox, sandesh_fetch, sandesh_thread, sandesh_archive, sandesh_unarchive, sandesh_search. The server also returns usage instructions on initialize and serves the full scenarios doc as the sandesh://usage resource. (Lifecycle: read = being acted on, reply = done — there is no separate status tool.)

Manual smoke — the MCP Inspector (browser UI):

mcp dev -m sandesh.mcp_server     # needs the [mcp] extra; explore tools by hand

The wake path stays the notify watcher — MCP exposes the verbs, not the wake (an MCP server can't re-invoke a sleeping agent). No notify/watch tool is exposed over MCP.

Discover via the MCP Registry

Sandesh is listed on the official MCP Registry as io.github.anthill-tec/sandesh (see server.json), so MCP-aware clients/aggregators (Claude Code, Cursor, Glama, PulseMCP, mcp.so, …) can discover it. The listing points at the PyPI distribution sandesh-relay and the stdio launch uvx --from 'sandesh-relay[mcp]' sandesh-mcp. The listing is the server (the verbs) only — the notify wake is a separate background process, not a tool. Publishing the listing is a maintainer step (after the PyPI release) — see RELEASING.md.

Test

python3 -m unittest -v          # from the repo root (stdlib-only: CLI + library)
.venv/bin/python -m unittest -v # includes the MCP server + E2E tests (needs the venv)

Roadmap

Shipped: the MCP server (stdio, [mcp] extra), packaging + OIDC PyPI publish, the official MCP Registry listing (io.github.anthill-tec/sandesh), and the Pi extension with native wake (integrations/pi/, npm @anthill-tec/sandesh-pi). Registry publishes (PyPI / MCP-registry / npm) are maintainer actions — see RELEASING.md.

License

Copyright © 2026 anthill-tec. Licensed under the GNU General Public License v3.0 — see LICENSE. You may use, study, modify, and redistribute Sandesh under the GPLv3; distributed derivatives must remain GPL-licensed with source available.