Skyflo keeps the objective, plan, agents, tools, and memory together across long-running engineering work.
Skyflo
linking · 0%
Mission lifecycle
standbyrl-311 · objective
6%Add per-tenant rate limiting to the public API
Four registered checkouts, one pinned header contract, and no breaking change for existing clients.
Execution path
1 / 6 active
Live operation
discover‣ registered checkouts loaded
Checks
0 / 12
Actions
0 / 2
Skyflo is a local-first AI engineering meta-harness. It coordinates coding agents across repositories, browser, terminal, and automations while the approved mission remains the durable record.
AI engineering meta-harness
Skyflo keeps the approved objective and evidence in one engineering control plane. Each runtime works in an assigned worktree.
Skyflo Harness orchestration
NativeApp Server
Agent SDK
ACP
Supervised CLI
ACP
ACP + server
RPC
ACP-compatible agents register by manifest.
Provider direct + local
Provider-direct requests use the accounts and keys you configure.
Compatible endpoints bring local and open-weight models into the same mission.
Kimi K3
Moonshot · BYOK
GLM-5.3
Z.ai Coding Plan · BYOK
GLM-5.2
Z.ai · BYOK
GLM-4.7-FlashX
Z.ai · BYOK
GLM-5V-Turbo
Z.ai vision · BYOK
Skyflo Managed
Choose a managed model without adding another provider key.
Execution and tools stay local. Model requests use the separately metered gateway.
GPT-5.6
Luna · Terra · Sol
Claude Sonnet 5
Managed catalogue
Gemini 3.6 Flash
Managed catalogue
Claude Fable 5
Max catalogue
Kimi K3
Managed catalogue
Open + local
Use BYOK or a compatible local endpoint.
Mistral Medium 3.5
BYOK / local
Mistral Small 4
BYOK / local
DeepSeek V4
BYOK
gpt-oss-20b
Local
The mission loop
It advances through six states, persists them to disk, and blocks in plan mode until you approve the implementation boundary. When review comes back short, the mission can return to planning.
Where a single agent run stops
Cross-repo scope
A task can touch several repositories. The shared contract and landing order still need one owner.
Continuity
The objective, decisions, and evidence should survive the session that produced them.
Execution surfaces
Code, terminal output, browser checks, and automations need to stay attached to the same objective.
Independent review
The process that made the change should not be the only process that judges it.
Skyflo keeps the objective when the session ends. The plan, specialists, execution surfaces, review, and accepted memory stay attached to the mission.
See how it is put togetherThe harness
Surfaces on top, memory at the bottom, and the mission loop that moves work between them.
Surfaces. Desktop, browser, terminal; all on one mission.
SF-05 · one mission, several surfacesIndependent review report
mission local-042 · reviewer profile
Change evidence
Reviewer boundary
Independent review
It runs as a separate profile holding file.read, file.search, and agent.report, with read-only capability operations and mutationAllowed set to false. Its findings attach to the mission; the edits stay with the specialists that made them.
See how independent code review worksCross-mission memory
A record is written only after you accept it, keeps a link to the mission and file that produced it, and is retrieved against the current workspace, so unrelated records stay out of the context.
The Skyflo native harness stays Free with supported provider keys and compatible endpoints. External agent runtimes use their existing authentication, while paid Skyflo Managed remains a separate inference lane. Execution stays on your linked Mac. BYOK credentials stay in the macOS Keychain.
Skyflo is a local-first AI engineering harness. A persistent mission holds one objective, its approved plan, isolated specialist work, browser and terminal activity, automations, monitors, and useful memory.
Coding agents execute development tasks. Skyflo is the mission-level harness that keeps the approved objective, delegated work, execution evidence, independent review, and accepted memory together across repositories and sessions.
Yes. You register the checkouts a change touches, and the mission plans across all of them at once. Each coder specialist then works in its own git worktree, so parallel changes cannot collide.
Skyflo Desktop runs the mission on your linked Mac using your repositories and local credentials. Model calls use the provider or compatible endpoint you configure.
Skyflo runs Codex, Claude Agent SDK, Cursor Agent, Antigravity, Grok CLI, OpenCode, Pi, and conforming ACP agents inside one mission. Provider-direct Kimi, GLM, Mistral, DeepSeek, and open-weight models use your configured access; Skyflo Managed remains a separately metered lane.
Personal, source-linked memory can be recalled by a later mission and reviewed, accepted, dismissed, or forgotten. Organization-wide shared memory is Planned and is labeled that way wherever it appears.
Looking for the Kubernetes agent? It continues as a separate Apache-2.0 project, with its own pages and docs.
Skyflo Desktop
Pick a change that already spans more than one repository.
The native harness and external adapters work with your configured access. Account required. Free requires no card.
