MCP Agent Mail

"It's like gmail for your coding agents!"

A mail-like coordination layer for coding agents, exposed as an HTTP-only FastMCP server. It gives agents memorable identities, an inbox/outbox, searchable message history, and voluntary file reservation "leases" to avoid stepping on each other.

Think of it as asynchronous email + directory + change-intent signaling for your agents, backed by Git (for human-auditable artifacts) and SQLite (for indexing and queries).

Status: Under active development. The design is captured in detail in project_idea_and_guide.md (start with the original prompt at the top of that file).

Why this exists

Modern projects often run multiple coding agents at once (backend, frontend, scripts, infra). Without a shared coordination fabric, agents:

• Overwrite each other's edits or panic on unexpected diffs

• Miss critical context from parallel workstreams

• Require humans to "liaison" messages across tools and teams

This project provides a lightweight, interoperable layer so agents can:

• Register a temporary-but-persistent identity (e.g., GreenCastle)

• Send/receive GitHub-Flavored Markdown messages with images

• Search, summarize, and thread conversations

• Declare advisory file reservations (leases) on files/globs to signal intent

• Inspect a directory of active agents, programs/models, and activity

It's designed for: FastMCP clients and CLI tools (Claude Code, Codex, Gemini CLI, etc.) coordinating across one or more codebases.

TLDR Quickstart

One-line installer

What this does:

• Installs uv if missing and updates your PATH for this session

• Creates a Python 3.14 virtual environment and installs dependencies with uv

• Runs the auto-detect integration to wire up supported agent tools

• Starts the MCP HTTP server on port 8765 and prints a masked bearer token

• Creates helper scripts under scripts/ (including run_server_with_token.sh)

Prefer a specific location or options? Add flags like --dir <path>, --project-dir <path>, --no-start, --start-only, or --token <hex>.

If you want to do it yourself

Clone the repo, set up and install with uv in a python 3.14 venv (install uv if you don't have it already), and then run scripts/automatically_detect_all_installed_coding_agents_and_install_mcp_agent_mail_in_all.sh. This will automatically set things up for your various installed coding agent tools and start the MCP server on port 8765. If you want to run the MCP server again in the future, simply run scripts/run_server_with_token.sh:

Ready-Made Blurb to Add to Your AGENTS.md or CLAUDE.md Files:

Integrating with Beads (dependency‑aware task planning)

Beads is a lightweight task planner (bd CLI) that complements Agent Mail by keeping status and dependencies in one place while Mail handles messaging, file reservations, and audit trails. Project: steveyegge/beads

Highlights:

• Beads owns task prioritization; Agent Mail carries the conversations and artifacts.

• Shared identifiers (e.g., bd-123) keep Beads issues, Mail threads, and commits aligned.

• Install the bd CLI via prebuilt release or Go build; see the repository for platform specifics.

Copy/paste blurb for agent-facing docs (leave as-is for reuse):

Core ideas (at a glance)

• HTTP-only FastMCP server (Streamable HTTP). No SSE, no STDIO.

• Dual persistence model:

  • Human-readable markdown in a per-project Git repo for every canonical message and per-recipient inbox/outbox copy

  • SQLite with FTS5 for fast search, directory queries, and file reservations/leases

• "Directory/LDAP" style queries for agents; memorable adjective+noun names

• Advisory file reservations for editing surfaces; optional pre-commit guard

• Resource layer for convenient reads (e.g., resource://inbox/{agent})

Typical use cases

• Multiple agents splitting a large refactor across services while staying in sync

• Frontend and backend teams of agents coordinating thread-by-thread

• Protecting critical migrations with exclusive file reservations and a pre-commit guard

• Searching and summarizing long technical discussions as threads evolve

• Discovering and linking related projects (e.g., frontend/backend) through AI-powered suggestions

Architecture

Web UI (human-facing mail viewer)

The server ships a lightweight, server‑rendered Web UI for humans. It lets you browse projects, agents, inboxes, single messages, attachments, file reservations, and perform full‑text search with FTS5 when available (with an automatic LIKE fallback).

• Where it lives: built into the HTTP server in mcp_agent_mail.http under the /mail path.

• Who it's for: humans reviewing activity; agents should continue to use the MCP tools/resources API.

Launching the Web UI

Recommended (simple):

Advanced (manual commands):

Auth notes:

• GET pages in the UI are not gated by the RBAC middleware (it classifies POSTed MCP calls only), but if you set a bearer token the separate BearerAuth middleware protects all routes by default.

• For local dev, set HTTP_ALLOW_LOCALHOST_UNAUTHENTICATED=true (and optionally HTTP_BEARER_TOKEN), so localhost can load the UI without headers.

• Health endpoints are always open at /health/*.

Routes and what you can do

• /mail (Unified inbox + Projects + Related Projects Discovery)

  • Shows a unified, reverse‑chronological inbox of recent messages across all projects with excerpts, relative timestamps, sender/recipients, and project badges.

  • Below the inbox, lists all projects (slug, human name, created time) with sibling suggestions.

  • Suggests likely sibling projects when two slugs appear to be parts of the same product (e.g., backend vs. frontend). Suggestions are ranked with heuristics and, when LLM_ENABLED=true, an LLM pass across key docs (README.md, AGENTS.md, etc.).

  • Humans can Confirm Link or Dismiss suggestions from the dashboard. Confirmed siblings become highlighted badges but do not automatically authorize cross‑project messaging—agents must still establish AgentLink approvals via request_contact/respond_contact.

• /mail/projects (Projects index)

  • Dedicated projects list view; click a project to drill in.

• /mail/{project} (Project overview + search + agents)

  • Rich search form with filters:

    • Scope: subject/body/both, Order: relevance or time, optional "boost subject".

    • Query tokens: supports subject:foo, body:"multi word", quoted phrases, and bare terms.

    • Uses FTS5 bm25 scoring when available; otherwise falls back to SQL LIKE on subject/body with your chosen scope.

  • Results show subject, sender, created time, thread id, and a highlighted snippet when using FTS.

  • Agents panel shows registered agents for the project with a link to each inbox.

  • Quick links to File Reservations and Attachments for the project header.

• /mail/{project}/inbox/{agent} (Inbox for one agent)

  • Reverse‑chronological list with subject, sender, created time, importance badge, thread id.

  • Pagination (?page=N&limit=M).

• /mail/{project}/message/{id} (Message detail)

  • Shows subject, sender, created time, importance, recipients (To/Cc/Bcc), thread messages.

  • Body rendering:

    • If the server pre‑converted markdown to HTML, it's sanitized with Bleach (limited tags/attributes, safe CSS via CSSSanitizer) and then displayed.

    • Otherwise markdown is rendered client‑side with Marked + Prism for code highlighting.

  • Attachments are referenced from the message frontmatter (WebP files or inline data URIs).

• /mail/{project}/search?q=... (Dedicated search page)

  • Same query syntax as the project overview search, with a token "pill" UI for assembling/removing filters.

• /mail/{project}/file_reservations (File Reservations list)

  • Displays active and historical file reservations (exclusive/shared, path pattern, timestamps, released/expired state).

• /mail/{project}/attachments (Messages with attachments)

  • Lists messages that contain any attachments, with subject and created time.

• /mail/unified-inbox (Cross-project activity)

  • Shows recent messages across all projects with thread counts and sender/recipients.

Human Overseer: Sending Messages to Agents

Sometimes a human operator needs to guide or redirect agents directly—whether to handle an urgent issue, provide clarification, or adjust priorities. The Human Overseer feature provides a web-based message composer that lets humans send high-priority messages to any combination of agents in a project.

Access: Click the prominent "Send Message" button (with the Overseer badge) in the header of any project view (/mail/{project}), or navigate directly to /mail/{project}/overseer/compose.

What Makes Overseer Messages Special

1. Automatic Preamble: Every message includes a formatted preamble that clearly identifies it as coming from a human operator and instructs agents to:

   • Pause current work temporarily

   • Prioritize the human's request over existing tasks

   • Resume original plans afterward (unless modified by the instructions)

2. High Priority: All overseer messages are automatically marked as high importance, ensuring they stand out in agent inboxes.

3. Policy Bypass: Overseer messages bypass normal contact policies, so humans can always reach any agent regardless of their contact settings.

4. Special Sender Identity: Messages come from a special agent named "HumanOverseer" (automatically created per project) with:

   • Program: WebUI

   • Model: Human

   • Contact Policy: open

The Message Preamble

Every overseer message begins with this preamble (automatically prepended):

Using the Composer

The composer interface provides:

• Recipient Selection: Checkbox grid of all registered agents (with "Select All" / "Clear" shortcuts)

• Subject Line: Required, shown in agent inboxes

• Message Body: GitHub-flavored Markdown editor with preview

• Thread ID (optional): Continue an existing conversation or start a new one

• Preamble Preview: See exactly how your message will appear to agents

Example Use Cases

Urgent Issue:

Priority Adjustment:

Clarification:

How Agents See Overseer Messages

When agents check their inbox (via fetch_inbox or resource://inbox/{name}), overseer messages appear like any other message but with:

• Sender: HumanOverseer

• Importance: high (displayed prominently)

• Body: Starts with the overseer preamble, followed by the human's message

• Visual cues: In the Web UI, these messages may have special highlighting (future enhancement)

Agents can reply to overseer messages just like any other message, continuing the conversation thread.

Technical Details

• Storage: Overseer messages are stored identically to agent-to-agent messages (Git + SQLite)

• Git History: Fully auditable—message appears in messages/YYYY/MM/{id}.md with commit history

• Thread Continuity: Can be part of existing threads or start new ones

• No Authentication Bypass: The overseer compose form still requires proper HTTP server authentication (if enabled)

Design Philosophy

The Human Overseer feature is designed to be:

• Explicit: Agents clearly know when guidance comes from a human vs. another agent

• Respectful: Instructions acknowledge agents have existing work and shouldn't just "drop everything" permanently

• Temporary: Agents are told to resume original plans once the human's request is complete

• Flexible: Humans can override this guidance directly in their message body

This creates a clear hierarchy (human → agents) while maintaining the collaborative, respectful tone of the agent communication system.

Related Projects Discovery

The Projects index (/mail) features an AI-powered discovery system that intelligently suggests which projects should be linked together—think frontend + backend, or related microservices.

How Discovery Works

1. Smart Analysis The system uses multiple signals to identify relationships:

• Pattern matching: Compares project names and paths (e.g., "my-app-frontend" ↔ "my-app-backend")

• AI understanding (when LLM_ENABLED=true): Reads README.md, AGENTS.md, and other docs to understand each project's purpose and detect natural relationships

• Confidence scoring: Ranks suggestions from 0-100% with clear rationales

2. Beautiful Suggestions Related projects appear as polished cards on your dashboard with:

• 🎯 Visual confidence indicators showing match strength

• 💬 AI-generated rationales explaining the relationship

• ✅ Confirm Link — accept the suggestion

• ✖️ Dismiss — hide irrelevant matches

3. Quick Navigation Once confirmed, both projects display interactive badges for instant navigation between related codebases.

Why Suggestions, Not Auto-Linking?

TL;DR: We keep you in control. Discovery helps you find relationships; explicit approvals control who can actually communicate.

Agent Mail uses agent-centric messaging — every message follows explicit permission chains:

This design ensures:

• Security: No accidental cross-project message delivery

• Transparency: You always know who can talk to whom

• Audit trails: All communication paths are explicitly approved

Why not auto-link with AI? If we let an LLM automatically authorize messaging between projects, we'd be:

• ❌ Bypassing contact policies without human oversight

• ❌ Risking message misdelivery to unintended recipients

• ❌ Creating invisible routing paths that are hard to audit

• ❌ Potentially linking ambiguously-named projects incorrectly

Instead, we give you discovery + control:

• ✅ AI suggests likely relationships (safe, read-only analysis)

• ✅ You confirm what makes sense (one click)

• ✅ Agents still use request_contact / respond_contact for actual messaging permissions

• ✅ Clear separation: discovery ≠ authorization

The Complete Workflow

Think of it like LinkedIn: The system suggests connections, but only you decide who gets to send messages.

Search syntax (UI)

The UI shares the same parsing as the API's _parse_fts_query:

• Field filters: subject:login, body:"api key"

• Phrase search: "build plan"

• Combine terms: login AND security (FTS)

• Fallback LIKE: scope determines whether subject, body, or both are searched

Prerequisites to see data

The UI reads from the same SQLite + Git artifacts as the MCP tools. To populate content:

1. Ensure a project exists (via tool call or CLI):

   • Ensure/create project: ensure_project(human_key)

2. Register one or more agents: register_agent(project_key, program, model, name?)

3. Send messages: send_message(...) (attachments and inline images are supported; images may be converted to WebP).

Once messages exist, visit /mail, click your project, then open an agent inbox or search.

Implementation and dependencies

• Templates live in src/mcp_agent_mail/templates/ and are rendered by Jinja2.

• Markdown is converted with markdown2 on the server where possible; HTML is sanitized with Bleach (with CSS sanitizer when available).

• Tailwind CSS, Lucide icons, Alpine.js, Marked, and Prism are loaded via CDN in base.html for a modern look without a frontend build step.

• All rendering is server‑side; there's no SPA router. Pages degrade cleanly without JavaScript.

Security considerations

• HTML sanitization: Only a conservative set of tags/attributes are allowed; CSS is filtered. Links are limited to http/https/mailto/data.

• Auth: Use bearer token or JWT when exposing beyond localhost. For local dev, enable localhost bypass as noted above.

• Rate limiting (optional): Token‑bucket limiter can be enabled; UI GET requests are light and unaffected by POST limits.

Troubleshooting the UI

• Blank page or 401 on localhost: Either unset HTTP_BEARER_TOKEN or set HTTP_ALLOW_LOCALHOST_UNAUTHENTICATED=true.

• No projects listed: Create one with ensure_project.

• Empty inbox: Verify recipient names match exactly and messages were sent to that agent.

• Search returns nothing: Try simpler terms or the LIKE fallback (toggle scope/body).

On-disk layout (per project)

Message file format

Messages are GitHub-Flavored Markdown with JSON frontmatter (fenced by ---json). Attachments are either WebP files referenced by relative path or inline base64 WebP data URIs.

Data model (SQLite)

• projects(id, human_key, slug, created_at)

• agents(id, project_id, name, program, model, task_description, inception_ts, last_active_ts, attachments_policy, contact_policy)

• messages(id, project_id, sender_id, thread_id, subject, body_md, created_ts, importance, ack_required, attachments)

• message_recipients(message_id, agent_id, kind, read_ts, ack_ts)

• file_reservations(id, project_id, agent_id, path_pattern, exclusive, reason, created_ts, expires_ts, released_ts)

• agent_links(id, a_project_id, a_agent_id, b_project_id, b_agent_id, status, reason, created_ts, updated_ts, expires_ts)

• project_sibling_suggestions(id, project_a_id, project_b_id, score, status, rationale, created_ts, evaluated_ts, confirmed_ts, dismissed_ts)

• fts_messages(message_id UNINDEXED, subject, body) + triggers for incremental updates

Concurrency and lifecycle

• One request/task = one isolated operation

• Archive writes are guarded by a per-project .archive.lock under projects/<slug>/

• Git index/commit operations are serialized across the shared archive repo by a repo-level .commit.lock

• DB operations are short-lived and scoped to each tool call; FTS triggers keep the search index current

• Artifacts are written first, then committed as a cohesive unit with a descriptive message

• Attachments are content-addressed (sha1) to avoid duplication

How it works (key flows)

1. Create an identity

• register_agent(project_key, program, model, name?, task_description?) → creates/updates a named identity, persists profile to Git, and commits.

2. Send a message

• send_message(project_key, sender_name, to[], subject, body_md, cc?, bcc?, attachment_paths?, convert_images?, importance?, ack_required?, thread_id?, auto_contact_if_blocked?)

• Writes a canonical message under messages/YYYY/MM/, an outbox copy for the sender, and inbox copies for each recipient; commits all artifacts.

• Optionally converts images (local paths or data URIs) to WebP and embeds small ones inline.

3. Check inbox

• fetch_inbox(project_key, agent_name, since_ts?, urgent_only?, include_bodies?, limit?) returns recent messages, preserving thread_id where available.

• acknowledge_message(project_key, agent_name, message_id) marks acknowledgements.

4. Avoid conflicts with file reservations (leases)

• file_reservation_paths(project_key, agent_name, paths[], ttl_seconds, exclusive, reason) records an advisory lease in DB and writes JSON reservation artifacts in Git; conflicts are reported if overlapping active exclusives exist (reservations are still granted; conflicts are returned alongside grants).

• release_file_reservations(project_key, agent_name, paths? | file_reservation_ids?) releases active leases (all if none specified). JSON artifacts remain for audit history.

• Optional: install a pre-commit hook in your code repo that blocks commits conflicting with other agents' active exclusive file reservations.

5. Search & summarize

• search_messages(project_key, query, limit?) uses FTS5 over subject and body.

• summarize_thread(project_key, thread_id, include_examples?) extracts key points, actions, and participants from the thread.

• reply_message(project_key, message_id, sender_name, body_md, ...) creates a subject-prefixed reply, preserving or creating a thread.

Semantics & invariants

• Identity

  • Names are memorable adjective+noun and unique per project; name_hint is sanitized (alnum) and used if available

  • whois returns the stored profile; list_agents can filter by recent activity

  • last_active_ts is bumped on relevant interactions (messages, inbox checks, etc.)

• Threads

  • Replies inherit thread_id from the original; if missing, the reply sets thread_id to the original message id

  • Subject lines are prefixed (e.g., Re:) for readability in mailboxes

• Attachments

  • Image references (file path or data URI) are converted to WebP; small images embed inline when policy allows

  • Non-absolute paths resolve relative to the project repo root

  • Stored under attachments/<xx>/<sha1>.webp and referenced by relative path in frontmatter

• File Reservations

  • TTL-based; exclusive means "please don't modify overlapping surfaces" for others until expiry or release

  • Conflict detection is per exact path pattern; shared reservations can coexist, exclusive conflicts are surfaced

  • JSON artifacts remain in Git for audit even after release (DB tracks release_ts)

• Search

  • External-content FTS virtual table and triggers index subject/body on insert/update/delete

  • Queries are constrained to the project id and ordered by created_ts DESC

Contact model and "consent-lite" messaging

Goal: make coordination "just work" without spam across unrelated agents. The server enforces per-project isolation by default and adds an optional consent layer within a project so agents only contact relevant peers.

Isolation by project

• All tools require a project_key. Agents only see messages addressed to them within that project.

• An agent working in Project A is invisible to agents in Project B unless explicit cross-project contact is established (see below). This avoids distraction between unrelated repositories.

Policies (per agent)

• open: accept any targeted messages in the project.

• auto (default): allow messages when there is obvious shared context (e.g., same thread participants; recent overlapping active file reservations; recent prior direct contact within a TTL); otherwise requires a contact request.

• contacts_only: require an approved contact link first.

• block_all: reject all new contacts (errors with CONTACT_BLOCKED).

Use set_contact_policy(project_key, agent_name, policy) to update.

Request/approve contact

• request_contact(project_key, from_agent, to_agent, reason?, ttl_seconds?) creates or refreshes a pending link and sends a small ack_required "intro" message to the recipient.

• respond_contact(project_key, to_agent, from_agent, accept, ttl_seconds?) approves or denies; approval grants messaging until expiry.

• list_contacts(project_key, agent_name) surfaces current links.

Auto-allow heuristics (no explicit request required)

• Same thread: replies or messages to thread participants are allowed.

• Recent overlapping file reservations: if sender and recipient hold active file reservations in the project, messaging is allowed.

• Recent prior contact: a sliding TTL allows follow-ups between the same pair.

These heuristics minimize friction while preventing cold spam.

Cross-project coordination (frontend vs backend repos)

When two repos represent the same underlying project (e.g., frontend and backend), you have two options:

1. Use the same project_key across both workspaces. Agents in both repos operate under one project namespace and benefit from full inbox/outbox coordination automatically.

2. Keep separate project_keys and establish explicit contact:

   • In backend, agent GreenCastle calls:

     • request_contact(project_key="/abs/path/backend", from_agent="GreenCastle", to_agent="BlueLake", reason="API contract changes")

   • In frontend, BlueLake calls:

     • respond_contact(project_key="/abs/path/backend", to_agent="BlueLake", from_agent="GreenCastle", accept=true)

   • After approval, messages can be exchanged; in default auto policy the server allows follow-up threads/reservation-based coordination without re-requesting.

Important: You can also create reciprocal links or set open policy for trusted pairs. The consent layer is on by default (CONTACT_ENFORCEMENT_ENABLED=true) but is designed to be non-blocking in obvious collaboration contexts.

Resource layer (read-only URIs)

Expose common reads as resources that clients can fetch. See API Quick Reference → Resources for the full list and parameters.

Example (conceptual) resource read:

File Reservations and the optional pre-commit guard

Exclusive file reservations are advisory but visible and auditable:

• A reservation JSON is written to file_reservations/<sha1(path)>.json capturing holder, pattern, exclusivity, created/expires

• The pre-commit guard scans active exclusive reservations and blocks commits that touch conflicting paths held by another agent

• Agents must set AGENT_NAME so the guard knows who "owns" the commit

Install the guard into a code repo (conceptual tool call):

Configuration

Configuration is loaded from an existing .env via python-decouple. Do not use os.getenv or auto-dotenv loaders.

Common variables you may set:

Configuration reference

Development quick start

Prerequisite: complete the setup above (Python 3.14 + uv venv + uv sync).

Dev helpers:

Database schema (automatic):

Run the server (HTTP-only). Use the Typer CLI or module entry:

Connect with your MCP client using the HTTP (Streamable HTTP) transport on the configured host/port. The endpoint tolerates both /mcp and /mcp/.

Search syntax tips (SQLite FTS5)

• Basic terms: plan users

• Phrase search: "build plan"

• Prefix search: mig*

• Boolean operators: plan AND users NOT legacy

• Field boosting is not enabled by default; subject and body are indexed. Keep queries concise. When FTS is unavailable, the UI/API automatically falls back to SQL LIKE on subject/body.

Design choices and rationale

• HTTP-only FastMCP: Streamable HTTP is the modern remote transport; STDIO is not exposed here by design

• Git + Markdown: Human-auditable, diffable artifacts that fit developer workflows (inbox/outbox mental model)

• SQLite + FTS5: Efficient indexing/search with minimal ops footprint

• Advisory file reservations: Make intent explicit and reviewable; optional guard enforces reservations at commit time

• WebP attachments: Compact images by default; inline embedding keeps small diagrams in context

  • Optional: keep original binaries and dedup manifest under attachments/ for audit and reuse

Examples (conceptual tool calls)

This section has been removed to keep the README focused. See API Quick Reference below for canonical method signatures.

Operational notes

• One async session per request/task; don't share across concurrent coroutines

• Use explicit loads in async code; avoid implicit lazy loads

• Use async-friendly file operations when needed; Git operations are serialized with a file lock

• Clean shutdown should dispose any async engines/resources (if introduced later)

Security and ops

• Transport

  • HTTP-only (Streamable HTTP). Place behind a reverse proxy (e.g., NGINX) with TLS termination for production

• Auth

  • Optional JWT (HS*/JWKS) via HTTP middleware; enable with HTTP_JWT_ENABLED=true

  • Static bearer token (HTTP_BEARER_TOKEN) is independent of JWT; when set, BearerAuth protects all routes (including UI). You may use it alone or together with JWT.

  • When JWKS is configured (HTTP_JWT_JWKS_URL), incoming JWTs must include a matching kid header; tokens without kid or with unknown kid are rejected

  • Starter RBAC (reader vs writer) using role configuration; see HTTP_RBAC_* settings

• Reverse proxy + TLS (minimal example)

  • NGINX location block:

    upstream mcp_mail { server 127.0.0.1:8765; } server { listen 443 ssl; server_name mcp.example.com; ssl_certificate /etc/letsencrypt/live/mcp.example.com/fullchain.pem; ssl_certificate_key /etc/letsencrypt/live/mcp.example.com/privkey.pem; location /mcp/ { proxy_pass http://mcp_mail; proxy_set_header Host $host; proxy_set_header X-Forwarded-Proto https; } }

• Backups and retention

  • The Git repos and SQLite DB live under STORAGE_ROOT; back them up together for consistency

• Observability

  • Add logging and metrics at the ASGI layer returned by mcp.http_app() (Prometheus, OpenTelemetry)

• Concurrency

  • Archive writes: per-project .archive.lock prevents cross-project head-of-line blocking

  • Commits: repo-level .commit.lock serializes Git index/commit to avoid races across projects

Python client example (HTTP JSON-RPC)

This section has been removed to keep the README focused. Client code samples belong in examples/.

Troubleshooting

• "sender_name not registered"

  • Create the agent first with register_agent or create_agent_identity, or check the project_key you're using matches the sender's project

• Pre-commit hook blocks commits

  • Set AGENT_NAME to your agent identity; release or wait for conflicting exclusive file reservations; inspect .git/hooks/pre-commit

• Inline images didn't embed

  • Ensure convert_images=true; images are automatically inlined if the compressed WebP size is below the server's INLINE_IMAGE_MAX_BYTES threshold (default 64KB). Larger images are stored as attachments instead.

• Message not found

  • Confirm the project disambiguation when using resource://message/{id}; ids are unique per project

• Inbox empty but messages exist

  • Check since_ts, urgent_only, and limit; verify recipient names match exactly (case-sensitive)

FAQ

• Why Git and SQLite together?

  • Git provides human-auditable artifacts and history; SQLite provides fast queries and FTS search. Each is great at what the other isn't.

• Are file reservations enforced?

  • Yes, optionally. The server can block message writes when a conflicting active exclusive reservation exists (FILE_RESERVATIONS_ENFORCEMENT_ENABLED=true, default). Reservations themselves are advisory and always return both granted and conflicts. The optional pre-commit hook adds local enforcement at commit time in your code repo.

• Why HTTP-only?

  • Streamable HTTP is the modern remote transport for MCP; avoiding extra transports reduces complexity and encourages a uniform integration path.

• Why JSON‑RPC instead of REST or gRPC?

  • MCP defines a tool/resource method call model that maps naturally to JSON‑RPC over a single endpoint. It keeps clients simple (one URL, method name + params), plays well with proxies, and avoids SDK lock‑in while remaining language‑agnostic.

• Why separate "resources" (reads) from "tools" (mutations)?

  • Clear semantics enable aggressive caching and safe prefetch for resources, while tools remain explicit, auditable mutations. This split also powers RBAC (read‑only vs writer) without guesswork.

• Why canonical message storage in Git, not only in the database?

  • Git gives durable, diffable, human‑reviewable artifacts you can clone, branch, and audit. SQLite provides fast indexing and FTS. The combo preserves governance and operability without a heavyweight message bus.

• Why advisory file reservations instead of global locks?

  • Agents coordinate asynchronously; hard locks create head‑of‑line blocking and brittle failures. Advisory reservations surface intent and conflicts while the optional pre‑commit guard enforces locally where it matters.

• Why are agent names adjective+noun?

  • Memorable identities reduce confusion in inboxes, commit logs, and UI. The scheme yields low collision risk while staying human‑friendly (vs GUIDs) and predictable for directory listings.

• Why is project_key an absolute path?

  • Using the workspace’s absolute path creates a stable, collision‑resistant project identity across shells and agents. Slugs are derived deterministically from it, avoiding accidental forks of the same project.

• Why WebP attachments and optional inlining?

  • WebP provides compact, high‑quality images. Small images can be inlined for readability; larger ones are stored as attachments. You can keep originals when needed (KEEP_ORIGINAL_IMAGES=true).

• Why both static bearer and JWT/JWKS support?

  • Local development should be zero‑friction (single bearer). Production benefits from verifiable JWTs with role claims, rotating keys via JWKS, and layered RBAC.

• Why SQLite FTS5 instead of an external search service?

  • FTS5 delivers fast, relevant search with minimal ops. It’s embedded, portable, and easy to back up with the Git archive. If FTS isn’t available, we degrade to SQL LIKE automatically.

• Why is LLM usage optional?

  • Summaries and discovery should enhance—not gate—core functionality. Keeping LLM usage optional controls cost and latency while allowing richer UX when enabled.

API Quick Reference

Tools

Tip: to see tools grouped by workflow with recommended playbooks, fetch resource://tooling/directory.

Resources

Client Integration Guide

1. Fetch onboarding metadata first. Issue resources/read for resource://tooling/directory (and optionally resource://tooling/metrics) before exposing tools to an agent. Use the returned clusters and playbooks to render a narrow tool palette for the current workflow rather than dumping every verb into the UI.

2. Scope tools per workflow. When the agent enters a new phase (e.g., "Messaging Lifecycle"), remount only the cluster's tools in your MCP client. This mirrors the workflow macros already provided and prevents "tool overload."

3. Monitor real usage. Periodically pull or subscribe to log streams containing the tool_metrics_snapshot events emitted by the server (or query resource://tooling/metrics) so you can detect high-error-rate tools and decide whether to expose macros or extra guidance.

4. Fallback to macros for smaller models. If you're routing work to a lightweight model, prefer the macro helpers (macro_start_session, macro_prepare_thread, macro_file_reservation_cycle, macro_contact_handshake) and hide the granular verbs until the agent explicitly asks for them.

5. Show recent actions. Read resource://tooling/recent/60?agent=<name>&project=<slug> (adjust window as needed) to display the last few successful tool invocations relevant to the agent/project.

See examples/client_bootstrap.py for a runnable reference implementation that applies the guidance above.

Monitoring & Alerts

1. Enable metric emission. Set TOOL_METRICS_EMIT_ENABLED=true and choose an interval (TOOL_METRICS_EMIT_INTERVAL_SECONDS=120 is a good starting point). The server will periodically emit a structured log entry such as:

2. Ship the logs. Forward the structured stream (stderr/stdout or JSON log files) into your observability stack (e.g., Loki, Datadog, Elastic) and parse the tools[] array.

3. Alert on anomalies. Create a rule that raises when errors / calls exceeds a threshold for any tool (for example 5% over a 5‑minute window) so you can decide whether to expose a macro or improve documentation.

4. Dashboard the clusters. Group by cluster to see where agents are spending time and which workflows might warrant additional macros or guard-rails.

See docs/observability.md for a step-by-step cookbook (Loki/Prometheus example pipelines included), and docs/GUIDE_TO_OPTIMAL_MCP_SERVER_DESIGN.md for a comprehensive design guide covering tool curation, capability gating, security, and observability best practices.

Operations teams can follow docs/operations_alignment_checklist.md, which links to the capability templates in deploy/capabilities/ and the sample Prometheus alert rules in deploy/observability/.

Deployment quick notes

• Direct uvicorn: uvicorn mcp_agent_mail.http:build_http_app --factory --host 0.0.0.0 --port 8765

• Python module: python -m mcp_agent_mail.http --host 0.0.0.0 --port 8765

• Gunicorn: gunicorn -c deploy/gunicorn.conf.py mcp_agent_mail.http:build_http_app --factory

• Docker: docker compose up --build

CI/CD

• Lint and Typecheck CI: GitHub Actions workflow runs Ruff and Ty on pushes/PRs to main/develop.

• Release: Pushing a tag like v0.1.0 builds and pushes a multi-arch Docker image to GHCR under ghcr.io/<owner>/<repo> with latest and version tags.

• Nightly: A scheduled workflow runs migrations and lists projects daily for lightweight maintenance visibility.

Log rotation (optional)

If not using journald, a sample logrotate config is provided at deploy/logrotate/mcp-agent-mail to rotate /var/log/mcp-agent-mail/*.log weekly, keeping 7 rotations.

Logging (journald vs file)

• Default systemd unit (deploy/systemd/mcp-agent-mail.service) is configured to send logs to journald (StandardOutput/StandardError=journal).

• For file logging, configure your process manager to write to files under /var/log/mcp-agent-mail/*.log and install the provided logrotate config.

• Environment file path for systemd is /etc/mcp-agent-mail.env (see deploy/systemd/mcp-agent-mail.service).

Container build and multi-arch push

Use Docker Buildx for multi-arch images. Example flow:

Recommended tags: a moving latest and immutable version tags per release. Ensure your registry login is configured (docker login).

Systemd manual deployment steps

1. Copy project files to /opt/mcp-agent-mail and ensure permissions (owner appuser).

2. Place environment file at /etc/mcp-agent-mail.env based on deploy/env/production.env.

3. Install service file deploy/systemd/mcp-agent-mail.service to /etc/systemd/system/.

4. Reload systemd and start:

Optional (non-journald log rotation): install deploy/logrotate/mcp-agent-mail into /etc/logrotate.d/ and write logs to /var/log/mcp-agent-mail/*.log via your process manager or app config.

See deploy/gunicorn.conf.py for a starter configuration. For project direction and planned areas, read project_idea_and_guide.md.

CLI Commands

The project exposes a developer CLI for common operations:

• serve-http: run the HTTP transport (Streamable HTTP only)

• migrate: ensure schema and FTS structures exist

• lint / typecheck: developer helpers

• list-projects [--include-agents]: enumerate projects

• guard install <project_key> <code_repo_path>: install the pre-commit guard into a repo

• guard uninstall <code_repo_path>: remove the guard from a repo

• clear-and-reset-everything [--force]: DELETE the SQLite database (incl. WAL/SHM) and WIPE all contents under STORAGE_ROOT (including per-project Git archives). Use only when you intentionally want a clean slate.

• list-acks --project <key> --agent <name> [--limit N]: list messages requiring acknowledgement for an agent where ack is missing

• acks pending <project> <agent> [--limit N]: show pending acknowledgements for an agent

• acks remind <project> <agent> [--min-age-minutes N] [--limit N]: highlight pending ACKs older than a threshold

• acks overdue <project> <agent> [--ttl-minutes N] [--limit N]: list overdue ACKs beyond TTL

• file_reservations list <project> [--active-only/--no-active-only]: list file reservations

• file_reservations active <project> [--limit N]: list active file reservations

• file_reservations soon <project> [--minutes N]: show file reservations expiring soon

Examples:

Client integrations

Use the automated installer to wire up supported tools automatically (e.g., Claude Code, Cline, Windsurf, OpenCode). Run scripts/automatically_detect_all_installed_coding_agents_and_install_mcp_agent_mail_in_all.sh or the one‑liner in the Quickstart above.