Outpost

You walk into a meeting cold because your context sits scattered across email, Slack, and last week's notes. Outpost assembles and maintains that context. You then arrive already oriented.

What becomes possible

For Empowered Engineers

Keep track of people, projects, and threads. You do not depend on memory. Walk into every meeting already oriented. Set the team up once. A team of agents then keeps your awareness current in the background.

• A team of scheduled agents that sync mail, calendar, and Teams chat while you work
• A shared knowledge graph of people, organizations, projects, and topics, kept current as messages arrive
• A daily briefing that synthesizes what changed into your priorities
• Meeting briefs assembled before every call from attendee history and open threads
• Drafted email replies, chat messages, documents, and slide decks grounded in your context, when you opt into drafting
• Optional recruitment agents that screen candidates against your engineering standard

---

How Outpost Works

Outpost runs a team of agents rather than a single chat assistant. Each agent has a job and wakes on its own schedule. On each wake, an agent observes what changed, decides the most useful action, and does it. The agent then goes back to sleep. You configure the team once. The team then works without further input from you.

Your agent team

A fresh install ships six agents. Each carries a small set of skills and a cron schedule you can edit or disable in ~/.fit/outpost/scheduler.json.

Agent	What it does	Runs	Access
postman	Syncs mail and Teams, triages messages, drafts replies	every 15 min, work hours	full
concierge	Syncs calendar, prepares meeting briefs, files transcripts	every 30 min, work hours	full
librarian	Extracts entities into the knowledge graph, organizes files	four times a day	restricted
chief-of-staff	Synthesizes every agent's notes into a daily briefing	morning and evening	restricted
recruiter	Screens CVs and assesses interviews against your standard	three times a day	restricted
head-hunter	Scans public sources for open candidates, never contacts them	each morning	restricted

Each agent writes a short note per wake. The chief-of-staff reads all of them to assemble the daily briefing. You do not assemble it by hand. full and restricted refer to macOS access. See macOS Privacy and Security below. The recruitment agents (recruiter, head-hunter) ground their judgments in your Pathway engineering standard. Leave them disabled if you do not hire.

What your agents can do

The team's abilities come from skills. Skills are self-contained capabilities the agents load as needed. The default install ships these skills, grouped by purpose. Skills that compose content (marked draft) run only after you opt into the brief+draft posture.

Area	Skills
Sync sources	Apple Mail, Apple Calendar, Microsoft Teams chat
Build the knowledge graph	Extract entities, organize files (draft), record a changelog
Prepare for meetings	Meeting prep, process and trim meeting-notes sessions, follow up
Compose and send (draft)	Draft email replies, send chat messages
Documents and decks	Create and collaborate on documents (create is draft), create, review, and summarize slide decks
Look people up	Identify yourself and look up anyone in the corporate directory
Recruit engineers	Scan, track, screen, assess, decide on, and forget candidates

Outpost auto-discovers skills from the knowledge base. You can add your own or pull updates with npx fit-outpost update.

One vault, tiered for its audiences

The knowledge base root is an Obsidian vault. Numbered tier directories at the root hold the knowledge graph. Each tier is one unit of sharing. A lower tier number means a narrower audience. A fresh install ships five tiers:

Tier	Directory	Who reads it
0	0-Draft/	You only. Never shared. Drafts, agent reports, and owner-only notes.
1	1-Management/	You and fellow senior managers.
2	2-Confidential/	Managers with people or hiring duties. Recruitment records live here.
3	3-Team/	The whole team. People, projects, and topics live here.
4	4-Public/	Anyone, inside or outside the team.

Every other root entry is personal and never shared. Personal entries include the instruction files, Briefings/, and registry.yaml. A personal folder never starts with one digit and a dash. You can rename a tier, add one, or remove one.

Agents place each note in the widest tier that excludes everyone who must not read it. A note links only to its own tier or to a wider tier. Links in shared tiers are tier-prefixed, for example [[3-Team/People/Sarah Chen]]. The rule keeps every share complete. A recipient's links always resolve inside the tiers they received.

Sharing is cumulative. A member who receives tier 2 also receives tiers 3 and 4. Shares travel over any folder-syncing mount, such as OneDrive or Git. A shared tier directory is often a symlink into a separately synced folder. The tier rank comes from the symlink's own name at the vault root. Keep the vault root outside every cloud-synced folder. Tier 0 then never leaves your machine.

Validate any vault or received share:

npx fit-outpost validate ~/.local/share/fit/outpost/Team

The validator checks tier ranks, link direction, link resolution, link format, literal path strings, legacy layouts, and note frontmatter and tags. A validation-baseline.json file at the vault root downgrades known findings to warnings. New findings exit with a non-zero code. Add --json for machine-readable findings.

Prerequisites

Outpost spawns claude as a subprocess. It does not load your shell profile. Install Claude Code through Homebrew (brew install claude) rather than the native binary. The Homebrew install runs on Node.js, which supports NODE_EXTRA_CA_CERTS for enterprise CA certificates.

If your network requires a custom CA bundle, add an env block to ~/.fit/outpost/scheduler.json:

{
  "env": {
    "NODE_EXTRA_CA_CERTS": "~/.config/ssl/ca-bundle.pem"
  }
}

Where your data lives

Outpost runs on your Mac and keeps your data on it. This section covers where your context lives, where AI calls go, and Forward Impact's role. The enterprise-CA note above covers a different topic.

On-device storage. All content that Outpost handles lands on your device:

• The knowledge base at the path you pass to npx fit-outpost init. Outpost writes drafted emails into its 0-Draft/ tier.
• Outpost's cache directory (~/.cache/fit/outpost/), which holds all synced source content and each agent's per-wake output.
• Apple Mail's local store, which Outpost reads from.
• Apple Calendar's local store, which Outpost reads from. See Getting Started for the accounts Outpost picks up.
• Outpost's scheduler home (~/.fit/outpost/), which holds config, runtime state, logs, and a local socket. The log and state files retain bounded excerpts of agent output.

Where AI calls go. Outpost delegates every AI call to the Claude Code CLI already installed on your Mac. It does not select or override the endpoint. The endpoint is therefore whichever provider you configured Claude Code to reach. By default that is the Anthropic API. Each call's prompt carries the user content the agent assembled for that wake (knowledge-graph excerpts, synced mail and calendar content).

The model endpoint is not the only egress. Agents in the default install templates also make outbound calls beyond it. Those calls include scheduled scans of public sources. They also include browser automation that sends messages through your chat web apps.

Forward Impact's role. The Outpost product runs no Forward Impact-operated server that processes your content. It is a local scheduler around your own Claude Code installation. AI calls reach the provider you configured (Anthropic by default). They do not reach Forward Impact.

Regulated workloads. No BAA, SOC 2 attestation, or enterprise data-processing agreement exists for Outpost today. If your data is under a regulated gate, run your own approval process before you adopt it.

---

Choosing your posture

Before you turn Outpost on, decide how much it acts on your behalf. You record this choice at init. The scheduler honours it on every wake. Outpost has two postures. The CLI and fit-outpost status name them the same way:

• brief — the default. Outpost runs only skills whose every output stays inside its own knowledge base or cache. These skills sync your mail and calendar, build the knowledge graph, and prepare briefings. Outpost never composes a reply, message, or document for someone else. It never moves files outside the knowledge base. In this posture, Outpost reads and briefs only.
• brief+draft — everything brief does, plus the skills that compose content as you: email replies, chat messages, and documents. Outpost drafts these items and stages them for your review. Nothing leaves until you give explicit approval. Choose this posture when you also want help to write.

A fresh init defaults to brief. You opt into drafting deliberately. The trust contract never turns on without your action.

Getting Started

Outpost currently requires macOS. Email and calendar sync read from Apple Mail and Apple Calendar. A transitive dependency (@forwardimpact/libmacos) declares "os": ["darwin"]. npm install @forwardimpact/outpost fails on Linux and Windows with EBADPLATFORM. No degraded mode exists today. A cross-platform degraded install (without Apple sync) is on the roadmap. Until it ships, install Outpost on a Mac.

Mail and calendar account coverage. Outpost reads Mail.app's and Calendar.app's local stores. It picks up any account you sync inside those apps. That includes a Gmail account synced over IMAP in Mail.app and a Google Calendar synced over CalDAV in Calendar.app. Outpost does not see mail or calendar data that lives only outside those apps, for example the Gmail web app or a separate Outlook client.

brew install claude                     # Runtime: Outpost spawns claude as a subprocess
npm install @forwardimpact/outpost      # macOS only
npx fit-outpost init                    # Initialize the default "Team" knowledge base
npx fit-outpost daemon                  # Start the scheduler
npx fit-outpost status                  # Check what each agent is doing

After the scheduler runs, these commands drive the team day to day:

Command	What it does
fit-outpost status	Show each agent's schedule, last wake, and last action
fit-outpost wake <agent>	Wake one agent now and do not wait for its schedule
fit-outpost posture [mode]	Show or set the adoption posture (brief, brief+draft)
fit-outpost update	Pull the latest instructions, agents, and skills into a KB
fit-outpost validate	Validate agent definitions and knowledge bases
fit-outpost stop	Gracefully stop the daemon and any active agents

macOS Privacy and Security

Outpost needs access to the live Mail and Calendar stores it reads. Grant every permission to a single app, fit-outpost.app. That grant covers the whole scheduler and the agents it runs. You never grant access to node, claude, or any other helper process.

Outpost runs two kinds of agent. They need different access:

• full agents sync the live Mail and Calendar stores or send mail. They read those stores and drive Mail under the one fit-outpost.app grant.
• restricted agents only process already-synced content and your knowledge base, which lives outside every protected folder (~/.local/share/fit/outpost/). They need no macOS grant. A restricted agent cannot reach protected files, even if an attacker compromises it.

When macOS prompts for the Mail and Calendar stores, grant Full Disk Access to fit-outpost.app. If a draft-side skill sends mail, macOS also prompts once under Automation to let fit-outpost.app control Mail. Click Allow. Your knowledge base needs no grant.