Roadmap

What’s shipped, what’s actively being built, what’s committed for next quarter, and what’s on the horizon. Living doc; updated on every release.

Full source at ROADMAP.md. The roadmap has four tiers, each with a different honesty contract:

• Shipped is in the tree. You can git log it.
• In flight has actual work behind it this quarter. IC assigned, scope locked.
• Next is committed for the following quarter. Design first, code second.
• Horizon is candid about being speculative. No quarter, no IC.

Effort sizing is uniform across tiers: S is roughly a week of focused work, M is two to four weeks, L is a quarter.

Shipped

The changelog is the detailed, version-by-version ledger. This tier is the short version, so the roadmap stays a forward-looking document.

v0.5.3 (2026-06-24)

The signed macOS desktop app and Linux packages published with a working brew install --cask alfred-os, a conversational Ask surface, a full-width workflow canvas with a dismissible agent drawer, an honest run timeline with an errors-only filter, a self-healing reliability core (classify, retry, break, loop-detect), a code-structure memory layer over MCP, single-source scheduled auth, and a stronger memory store.

v0.5.2 (2026-06-22)

Design parity between the internal and OSS builds (the refreshed dark UI across Inbox / Ask / Work / Agents / Setup), Redis Agent Memory as the default local memory layer with FleetBrain as the review and reliability layer, smoother Slack planning threads, repo-graph support in the code-map flow, desktop visual-QA tooling, and a fresher public site.

v0.5.1 (2026-06-17)

First-run trust polish and the public download path for the signed native client: a public /download/ page, alfred serve and the desktop app aligned on port 7010, and the docs/site brought in line with the shipped client.

v0.5.0 (2026-06-15)

The native desktop app and the trust features around it: a signed Mac/Linux app over alfred serve (Inbox, Ask, Work, Agents, Setup), live Claude and Codex subscription usage, a single-repo approval gate, a disk guardian, an approved-Slack-plan-to-issue bridge, and review-first fleet memory with an optional Redis Agent Memory Server.

v0.4.0 (2026-05-23)

Substrate, observability, planning, approval, memory, and connector primitives: the agent_runner package decomposition, alfred metrics / alfred logs, the issue-claim state machine and multi-repo coordination, Damian and Batman planning/execution, the FleetBrain ledger, the connector protocol, alfred serve v1, and the slop detector.

v0.3.0 and earlier

Linux support, dry-run mode, fleet control verbs, and the solo-builder setup wizard. See the changelog for the full ledger.

In flight (this quarter)

Items with active work and a committed IC.

• Plan-review gate as a runtime feature. Promote plan() -> review_plan() -> execute() -> review_diff() from an architecture note to the default lifecycle for codenames that opt in. Today the review step exists in prose; the runtime makes it enforceable. IC: core. Effort: M. Issue: TBD.
• Public unattended-SLA emit format. Extend alfred-shipped-public with a 30-day window covering firings, success rate, and unattended hours. People who want a public usage page can render this on their own site. IC: core. Effort: S. Issue: TBD.
• Alfred Desktop v2. Keep Slack as the collaboration surface and build on the packaged Tauri shell with guided setup repair, release/update status, lock recovery, safer command previews, and a first-class Goals inbox with evidence. No extra gateway, no local mirror, no second source of truth. Keep alfred serve JSON APIs stable so the Tauri shell stays thin. IC: core. Effort: M. Issue: TBD.
• Memory quality loop v2. Improve duplicate collapse, evidence ranking, stale lesson retirement, and approved follow-up execution before a lesson can shape future runs. IC: core. Effort: M. Issue: TBD.

Next (next quarter)

Committed for the following quarter. Design first, then code.

• Multi-engine routing v2. Add Gemini and Ollama adapters alongside the current Claude and Codex engines. Per-codename engine selection stays the existing surface; the work is the adapter contract plus auth probes plus billing posture docs. Effort: M. Issue: TBD.
• Better Batman v2. Bundle-completion tracking after approval: Batman keeps watch on child issues and PRs, reports per-repo progress, and closes the bundle once every child has landed or been explicitly dropped. Effort: M. Issue: TBD.
• Native lifecycle dry-run for every shipped runner. alfred dry-run <codename> now resolves every codename safely; next step is making every individual runner support the full synthetic lifecycle, not just the safe simulation. Effort: S. Issue: TBD.
• alfred serve API hardening. v0.5.1 ships the local control surface used by Alfred Desktop. Next work is a versioned API contract, compatibility tests, richer event-stream traces, and clearer error payloads for native clients. Effort: S. Issue: TBD.

Horizon (no committed quarter)

Candidly speculative. No IC, no quarter, no committed effort estimate.

• Content-fleet codename pack (Scribe, Herald, Curator) for blog, LinkedIn, and SEO drafts.
• Sales-fleet codename pack for prospect identification, event-page sourcing, and outreach drafts.
• Marketing and SEO-fleet codename pack for site-page generation, content-drift detection, and search-visibility monitoring.
• Ops-fleet codename pack for uptime, release notes, and customer-health signals.
• Personal-assistant codename pack for inbox triage, calendar, and daily digest.

How these reach OSS. Cross-department codename packs build in the private internal orchestrator first, validated against real production usage, then port to OSS once generalised. See the private-to-public boundary workflow for the rules each port has to clear (no local host paths, no internal infra, no customer data, scrubbed prompts).

Considered, not committed

Decisions considered and left out. Listed so contributors do not re-pitch them.

• Plugin or skill marketplace bundled into Alfred. Considered and decided against. Skills are user-installed Claude Code skills; a bundled marketplace would push maintenance onto the framework. The convention-only resolver stays.
• Hosted Alfred SaaS. Not on the roadmap. Alfred is self-hosted by design; multi-tenant is a different product.
• First-class GitHub App instead of local gh auth. Larger onboarding surface; deferred until there is demonstrated demand.
• Pluggable spend backends (filesystem, SQLite, Redis). Single-host is the design, so this stays speculative.
• pipx / PyPI install. Git clone is the supported path today; a packaged install would widen the audience but the install story is fine.

Design boundaries

These are the design, not missing features.

• Single-person install. One person, one host, one config. Not multi-tenant, not a hosted SaaS.
• The OS schedules; Alfred runs. No long-running orchestration loop.
• Local CLIs, not a model gateway. Alfred shells out to claude and codex through your local CLI auth. The default path uses subscription-backed CLI accounts and does not require provider API keys.
• Lean on the platform. Adopt Anthropic-native capabilities (Agent Teams, the Memory Tool) rather than re-implement them.
• Browser automation is per-codename: installed in the codename’s own bin script.

Influence

• Strong: a working PR for something already on the in-flight or next list.
• Medium: a well-scoped feature request with a real use case and a proposal.
• Low: “would be cool if” comments.

Want to take Alfred somewhere new? Open a discussion first.