The stack you already run. Nothing to replace on day one.
sleds sits between the people who request work and the harnesses that build it. It reads the seats, reads what came back from the repo, listens to the harness, and writes three things: the size and actual to the row, the dispatch decision to the seat, and the event to wherever the person is. Every layer above and below stays whose it was.
cmux
OrcaOpenRouterBedrockLiteLLMWorkOSOne connector. Every verb where the request was written.
mcp.sleds.ai/mcp with your workspace account. intake, estimate, dispatch, status, touches, attach, answer, hand off, pick up, help.
The same MCP server through its connector slot. Our instructions travel with it.
A plugin bundles the connector, the "send to engineering" skill and the spec template.
The prototype link travels with the dispatch.
SSO, SCIM, MFA and an admin portal in one connection. Okta, Entra and Google through it, never by hand.
Everything sleds does, Claude can do from where you are.
sleds is an MCP server. Every verb — file, size, route, take, answer, land, status — is a tool Claude can call, with instructions that travel with it. The same server sits behind the Claude.ai connector, the Cowork plugin, Claude Code on an engineer's seat, and a managed agent on the Claude Console. One identity model: a member can file and answer; a seat is for taking work. And one voice: the AI Guidance your org writes is what every seat says, on every surface.
intake draft the brief, ask what it cannot answer, file with attachments estimate a cost and time band, a confidence, a cap — never a point status where is it: stage, who has it, the band, the PR touches what a request reaches in the product map, and what collides route the nodes and waves; arm, answer a decision, re-draft take · land pick it up on a seat; close it with the measured actual handoff to a teammate, with the brief on top and the diff second help the verbs table and the setup guides, in the chat
mcp.sleds.ai/mcp. Sign in with your work email; Claude gets your seat, never a token in the browser. The intake card renders inside the chat./sled, and the loop held to by hooks: status, estimate before the first line, actual on merge. The engine runs route nodes through it, in tmux or any terminal.The status line, the prompt hook and the cockpit.
Status line, prompt hook, /sled skill, the exits before the next prompt. Runs headless for overnight and defer.
Through the proxy on the workspace's own OpenAI key; a provider entry sleds init writes and uninstall removes.
Cursor's agent runs dispatched work like any other seat; its usage is read as a bar.
An executor for dispatched work; credits as a rail.
Seat quota, seat credits, platform key. For every platform a seat uses.
| platform | rail | status | what sleds reads |
|---|---|---|---|
| Claude | seat | live | 5-hour and weekly bars per model, credits, the reset. Read as you, from your own usage page. |
| OpenAI | seat | live | The second seat platform. Codex sessions metered through the proxy; bars and credits read. |
| Anthropic API | key | live | The platform key. Continue within an allowance; batch tier at half price overnight. |
| OpenAI API | key | live | The workspace's own key for Codex; batch tier overnight. |
| OpenRouter | key | live | A first-class key rail for the long tail. Its prompt-level model choice is a fallback for task types with no actuals. |
| Bedrock · Vertex | key | next | Enterprise keys in the customer's cloud account; the self-hosted proxy holds them. |
| LiteLLM | key | next | An internal gateway as a rail. We route between seats; it routes requests. |
The GitHub App carries the code and the receipt.
Every pull request is sized on open, mirrored onto its row as it is (who opened it, who merged it, what built it, the diff), and receipted on merge from the SLEDS-* lines, in every shape they have ever taken. A GitHub App reads each repo's tree to build the product map that intake reads. GitLab is planned.
Jira and Linear stay in step with the board both ways, for as long as you run them. Confluence spaces import into Docs in one run. If you ever want to run on sleds alone, the old tool becomes read-only history; most teams keep both.
SLEDS-Source: ENG-482 · brief · the board Worth: high · Acme renewal SLEDS-Estimate: $4–11 · 20–55 min · 81% · cap $20 SLEDS-Seat: overnight · batch · fable-5.1 SLEDS-Actual: 31 min · $5.20 SLEDS-Attest: ed25519 · 2026-09-14T06:41Z · verified
The tracker, the errors, the design. Connected, not replaced.
Each connector is the same shape: the thing you already use files a request or receives the receipt, and the row on the board is the one record. All six are live.
Issues mirror both ways while you move: one filed in Linear lands on the board sized, and a row that lands writes its state back. Until the day you pick.
An error group becomes a request with the stack, the release and how many people hit it, sized on the board. The fix lands with a receipt and the issue resolves.
The GitHub App: every pull request sized on open and receipted on merge, the product map read from each repo's tree. How the receipt works.
The SLEDS app: file from a thread, answer a parked question with a letter, read the roll-up in the channel. Where the event reaches people.
A frame on a request travels with it: the seat that builds reads the design beside the spec, and the record keeps the version it built from.
A board on a request carries the map. A sticky that turns into work becomes a request in one step, sized like any other.
Where the event reaches the person.
cmuxliveThe seat line per pane, a ring at the wall, the sidebar block. sleds doctor says cockpit: cmux.
The engine opens each session in its own tmux session and worktree; tmux attach is the watch handle, and the menu bar app opens it for you.
The plain runner: the engine starts the agent command in a terminal you name, or headless, and follows the PR. Nothing to install beyond the CLI.
The same seat line and ring inside Orca's terminal; the notification path decides the build.
The SLEDS app for Slack. Hand-offs, walls, a node parked with a question, a route landed: each a message with the item key and a link to the row. Reply take, pass or not me, or answer the question, and it lands on the item. The Friday roll-up in the channel.
Your meter in the browser; your sled on the Mac. The menu bar app shows what is running, what is at PR and what needs you, arms items for the engine, and sets its profile (start items on approval or auto, how many at once, which model builds, decides and drafts).
Nobody opens sleds. It shows up where they already are.
Connect what you already run.
Request an invite. Setup is the connector, the GitHub App and one line on each engineer's laptop.
curl -fsSL sleds.ai/install | sh