The pull
The right model at the front
Each step is done by the model it needs — small work on small models, hard work on strong ones — picked before an LLM token is spent.
No. 01 The AI engineering OS
One workspace where your people and agents work together, on the Claude Code and Codex seats you already pay for. A person approves every change.
Don't be left out when others are going at double speed.
Contact sales01 Built like a paceline
In a paceline, riders take turns at the front. When the leader tires, they peel off and the next rider pulls through, while the rest ride in the slipstream. That is how Paceline runs your work.
The pull
Each step is done by the model it needs — small work on small models, hard work on strong ones — picked before an LLM token is spent.
The rotation
When a model hits its usage or rate limit, the next allowed one pulls through — on another vendor if it has to. Nobody waits for a reset.
The draft
The conversation, the approved plan, the staged edits and a handoff note ride along, so whoever takes the front is already in the slipstream.
The pacer
Each channel keeps its goal and roadmap, and the Pacer agent notices when the work drifts from it. Jira stays in sync.
02 The AI engineering OS
One place where your people, your agents and every model you already pay for work as a team — under one policy, with a person approving every change.
Who saves — and what
MoneyTimeProductivity
Stage 1
A small team with specialist agents as teammates, on the subscriptions you already have — no second AI bill to start.
Stage 2
Keep Copilot, Cursor and Claude Code. Put one policy and a bill per task over them, and carry on on another seat instead of buying overage.
Stage 3
One governed layer over every vendor, with an audit trail your security team can export for review.
Stage 4
Finance, health and the public sector: keep work on your own Bedrock, Vertex or Azure, or on local models.
Stage 5
A channel per client, each with its own budget and a bill per task, so AI spend lands on the right invoice.
03 Race stats
CostFor the CFO
72%lower list price when each step runs on the model it needs — $0.695 became $0.193
One measured task — build, design review, code review. Its actual token counts, priced at list twice: Opus for every step, and the Sonnet build and Haiku reviews Paceline chose. It measures choosing the model per step, not a run against another tool.
ProductivityFor the VP Engineering
0dead ends at "limit reached"
Of the 23 AI coding products we researched, none document continuing a task on another vendor at a limit with its context.
SecurityFor the CISO
100%of writes approved by a person
Not yet certified: SOC 2 Type II is on the roadmap, with SAML and SCIM. Company sign-in with OpenID Connect works today.
04 The peloton
Channels where people and agents talk, plan and hand work to each other — and every change still waits for a person.
05 The course
A person asks. Paceline decides the path. Nothing is written until someone approves it.
People and agents in one thread, with the goal and roadmap beside it.
Reads the request and sizes each step: model, thinking depth, which agents ride.
Planner → Builder → Designer → Reviewer, on the models your policy allows.
Plan, change, design and review each stop for a person.
Written only when approved. Stale files are refused, never overwritten.
06 Time trial
Measured · 24 Sep 2026
72% lower list price for the same work. One build → design review → code review task, run live in the desktop app; its token counts are what the models reported, priced at list both ways. This isolates the saving from choosing the model per step. It is not a head-to-head run against another tool — that is what the sandbox is for. Subscription seats are not billed per token; this shows relative cost.
07 Standings
Capabilities as each vendor documents them. We score ourselves by the same rule, and mark what is still on our roadmap.
| Capability | Paceline | GitHub Copilot | Cursor | Claude Code | OpenAI Codex | Devin / Devin Desktop | Factory | Conductor | Cline | Warp (Oz) |
|---|---|---|---|---|---|---|---|---|---|---|
| Models from several vendors | Documented | Documented | Documented | Not documented or unverified | Partial or limited | Documented | Documented | Documented | Documented | Documented |
| Own cloud (Bedrock / Vertex / Azure) or local models | Documented | Documented | Documented | Documented | Documented | Partial or limited | Documented | Partial or limited | Documented | Partial or limited |
| Uses the team's existing Claude / ChatGPT subscriptions | Documented | Not documented or unverified | Not documented or unverified | Documented | Documented | Not documented or unverified | Not documented or unverified | Documented | Not documented or unverified | Partial or limited |
| Keeps working on another vendor when a limit is hit, context kept | Documented | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified |
| Admin policy: which models and vendors may see code | Documented | Documented | Documented | Documented | Partial or limited | Documented | Documented | Not documented or unverified | Documented | Not documented or unverified |
| Secret redaction before the model | Documented | Not documented or unverified | Not documented or unverified | Partial or limited | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified | Not documented or unverified |
| Tamper-evident audit log of model calls | Documented | Partial or limited | Documented | Documented | Documented | Not documented or unverified | Documented | Not documented or unverified | Partial or limited | Partial or limited |
| Spend caps / token ceilings | Documented | Documented | Documented | Documented | Partial or limited | Documented | Partial or limited | Not documented or unverified | Partial or limited | Partial or limited |
| Human approval before every write and command | Documented | Documented | Documented | Documented | Documented | Documented | Documented | Partial or limited | Documented | Not documented or unverified |
| Team channels with planner / builder / reviewer agents | Documented | Partial or limited | Partial or limited | Partial or limited | Partial or limited | Documented | Documented | Partial or limited | Not documented or unverified | Partial or limited |
| SSO / SCIM and SOC 2 | Partial or limited | Documented | Documented | Documented | Documented | Documented | Documented | Documented | Partial or limited | Partial or limited |
| Cloud / background agents | Roadmap | Documented | Documented | Documented | Documented | Documented | Documented | Partial or limited | Not documented or unverified | Documented |
Positioning · our assessment
Governance across vendors (one policy, continuity, audit, redaction) →
Every edit is staged on a change card a person approves.
A command card shows exactly what will run, and runs only on approval.
Keys and tokens in files, chat or output become placeholders first.
Company cloud, local, API keys and subscription CLIs, each allowed or blocked.
Each line carries the hash of the one before, so edits and deletions show.
A sandboxed window, keys in the OS keychain, prompt-injection fencing.
What we don't have yet
The leading tools already carry certifications we don't. They are first on our roadmap:
Until then, code and keys never leave a developer's machine except to the model vendors you have already approved, and every call is in the audit log.
08 The pilot
One team of five to fifteen, on your real backlog.
Week 1
Your existing Claude Code and Codex seats plus your company cloud. Set the data policy and the ceilings.
Weeks 2–3
In channels, with a weekly readout: tasks, first-time approvals, tokens per task, cache share.
Week 4
A security review of the exported audit log, and an ROI comparison against your current overage spend.
09 Pricing
Priced per active engineer. Agents are unlimited, guests are free, and there's no markup on tokens: you bring your Claude Code and Codex seats, your API keys or your company cloud.
Team
One team, up to 25 engineers
Talk to usMonthly · per active engineer
Business
25 to 200 engineers, several teams
Talk to usAnnual · 25-seat minimum
Everything in Team, plus
Enterprise
200+ engineers, or regulated teams
Talk to usAnnual · custom contract
Everything in Business, plus
On the roadmap: SAML and SCIM, SOC 2 Type II, cloud agents, and buying through the AWS, Azure and Google Cloud marketplaces.
Contact salesDesign partners: a free 30-day pilot for our first five teams, on your own backlog, with a week-4 readout of tokens per task and what you saved. In return we ask for a case study and a reference call.
Apply →10 Questions
Keep them. Paceline runs Claude Code and Codex as engines and adds what no single vendor can: continuity across vendors, one policy over all of them, and a team workflow with approval gates.
No Paceline server sees your code. Agents run on each developer's machine, with the models your policy allows. The routing classifier receives only the request text, and an admin can turn it off.
Not yet — it is first on the roadmap, alongside SAML and SCIM. Company sign-in with OpenID Connect (Okta, Entra, Google Workspace) and workspace roles already work on the team server. Meanwhile, code and keys never leave the developer's machine except to the model vendors you have approved, and every call is in the audit log.
Nothing is written or run without a person's approval, a file that changed since the agent read it is refused rather than overwritten, and a runaway task stops at its token ceiling.
Per active engineer: anyone who sends a message or approves a change in the month. Agents are unlimited and guests are free. Team is billed monthly; Business and Enterprise are annual.
No. You bring your own models: your Claude Code and Codex seats, your API keys, or Bedrock, Vertex and Azure under your company's own contracts. We charge for the platform, with no markup on usage.
For our first five design partners, nothing: 30 days on your own backlog, with a week-4 readout. After that, a paid pilot is credited against the annual contract.
Not yet. Marketplace listings are on the roadmap so the purchase can count against your cloud commitment. Until then, we invoice directly.
11 Contact sales
Tell us about your team. We'll set up a 30-day pilot, run the benchmark on your own repository, and send you the enterprise brief.