Skip to content
Download
Cusp 0.1 preview is hereRead the post

Any model.
One agent.

One agent that works with Claude, GPT, Gemini, GLM, Kimi, Qwen, DeepSeek or your own gateway, and adapts its tools to each model. On your desktop and in your terminal, with the web app coming soon.

Download Cusp

Free account: 2,500 agent actions a month on your own keys · macOS, Windows and Linux

The point where thought becomes work. Your request and the model you chose meet here. The agent fits its tools to that model, then does the work.

A Cusp desktop window. A user asks Cusp to fix a flaky checkout test; the agent reads the test, searches for fixed sleeps, edits the file and reruns the tests until they pass. The session repeats with Claude, GPT, GLM, Gemini and Qwen models.

provider and gateway presets, including local models
20+
provider and gateway presets, including local models
API protocols: Anthropic, OpenAI Chat and Responses, Gemini
4
API protocols: Anthropic, OpenAI Chat and Responses, Gemini
edit formats, chosen per model
3
edit formats, chosen per model
runtime behind the desktop app, CLI and web app
1
runtime behind the desktop app, CLI and web app

Bring any model

Paste a key. Pick any model.

Cusp detects the protocol, lists your models and picks a sensible default. Switch models mid-session and your context comes with you.

  • Claude Opus 5.5
  • GPT-6.1 Sol
  • Gemini 4 Argon
  • GLM-5.3
  • Kimi K3
  • Qwen3.8 Max
  • DeepSeek V4 Pro
  • Claude Sonnet 5.5
  • GPT-6 Astra
  • Grok 4.7
  • MiniMax M3
  • Mistral Medium 3.5
  • Gemini 3.8 Flash
  • Claude Haiku 5.5
  • OpenRouter
  • Vercel AI Gateway
  • LiteLLM
  • Amazon Bedrock
  • Google Vertex AI
  • Azure OpenAI
  • Ollama
  • LM Studio
  • Together
  • Fireworks
  • Groq
  • Any OpenAI-compatible URL

The product

The full agent toolkit, for any model.

The same tools, permissions, checkpoints and extensions, whichever model you choose.

A harness tuned per model

On Terminal-Bench, the same model scores up to ~30 points apart in different agents. Cusp gives each model the edit format chosen for it, passes reasoning back between turns, and checks diagnostics on the files it touched before it stops.

Model familyGets
ClaudeEdit
GPTApplyPatch
GeminiEdit
Qwen · MiniMax FlashEditLines
DeepSeek · MiniMaxReasoning

Zero configuration

Paste a key or a gateway URL. Cusp detects the protocol and lists your models.

sk-or-v1-8f3a••••••••••••2c91
  • Detected OpenRouter · Anthropic-compatible
  • Listed the models that can call tools
  • Default: Claude Sonnet 5.5

Brings your existing setup

Your instructions, skills, subagents, hooks and MCP servers work as they are. AGENTS.md, CLAUDE.md and GEMINI.md are read in place; the rest imports in one step.

CLAUDE.md.claude/skills.claude/agentshooks.mcp.jsonpermissions

Read 23 items in place. Nothing to rewrite.

Permissions you can read

Permission rules in the familiar Tool(specifier) syntax, an OS sandbox for shell commands on macOS and Linux, and an audit log of every decision.

  • allowBash(git status)
  • allowEdit(src/**)
  • askBash(git push *)
  • denyRead(./.env)

Rewind to any turn

Each turn is checkpointed. Fork a session, or roll code and conversation back to an earlier turn.

  1. Read 4 files
  2. Edit api/limits.ts
  3. Bash bun testRestore here
  4. Edit README.md

Parallel sessions, isolated worktrees

Run several agents at once, each on its own branch and worktree. Review the diff, merge when it’s ready.

Rate-limit the public API

cusp/rate-limit

+214 −38Running

Fix flaky checkout test

cusp/flaky-checkout

+3 −1Ready to merge

Upgrade to React 19.3

cusp/react-19-3

+96 −112Needs you

Synced everywhereComing soon

Sessions follow you from desktop to browser to phone. Steer a local run from anywhere, or hand it to a cloud sandbox.

  • MacBook ProRunning locally
  • app.cusp.runFollowing along
  • iPhoneApprove git push?

MCP, skills and subagents

Connect any MCP server, package know-how as skills, and delegate to subagents that run in parallel with their own context.

githublinearpostgresplaywrightsentryfigmaslack+ any server

Diffs worth reading

Word-level diffs, inline review comments and LSP diagnostics after each edit, so problems show up while you review.

− const limit = 100;

+ const limit = config.rateLimit ?? 100;

export function guard(req: Request) {

One system

Every agent, in one place.

Sessions, subagents, races, reviews and crew members aren’t separate features. They’re runs of one system, so they share one set of rules, one ledger and one inbox, and you can steer any of them while they work.

Mission controlComing soon6 runs · $1.56 today
  • Rate-limit the public APISession · local$0.42Working
  • Find every public routeSubagent · read-only · local$0.03Done
  • Write the limiter testsSubagent · worktree$0.11Working
  • Fix the flaky checkout testRace · 3 models · worktree$0.57Winner picked
  • Review PR #482Crew · Kai, Engineer · local$0.19Needs you
  • Deep review of the diffReview · 4 findings · worktree$0.24Done
An example run tree: a session with two subagents, a three-model race, a crew member waiting for approval and a review agent, each with its place, status and cost.

  • One gate

    Every tool call from every agent goes through the same permission rules and audit log. Shell commands run in the OS sandbox on macOS and Linux.

  • One inbox

    Permission requests, questions and finished work from every session, in one queue.

  • One ledger

    A cost receipt for every turn, and usage by day, model and workspace.

  • One supervisorComing soon

    Shared provider rate limits and priorities across runs, so parallel agents wait their turn instead of hitting a provider’s limits.

  • One boardComing soon

    Tasks, findings and file leases shared between agents, so two agents don’t write the same file at once.

  • Agent-operable by design. Most of what you can do in Cusp, an agent can do from chat, under the same rules. Anything that spends money or acts outside the workspace waits for your approval.

Race · preview

Race any prompt across models.

Run one prompt on several models and keep the result that passes the tests.

“The checkout test is flaky on CI. Fix it.”3 lanes · $0.57

  1. Claude Opus 5.5Winner
    Tests
    12/12
    Diff
    +18 −6
    Cost
    $0.31
    Time
    2m 04s
  2. GPT-6.1 Sol
    Tests
    12/12
    Diff
    +41 −9
    Cost
    $0.22
    Time
    1m 38s
  3. GLM-5.3
    Tests
    11/12
    Diff
    +23 −7
    Cost
    $0.04
    Time
    2m 51s

  • Same prompt, isolated worktrees

    Each model works on its own branch and worktree, so lanes don’t touch each other’s files.

  • Compare side by side

    Each lane’s transcript, diff and cost receipt next to the others, so you pick on evidence.

  • Judge and merge · coming soon

    A judge that doesn’t see model names, and one-click adopt or merge.

Why the harness matters

Same model. Up to 30 points apart.

On Terminal-Bench, the same model scores up to ~30 points apart in different agents, per the published results below. The harness is the part Cusp works on, and these are the techniques it uses.

GPT-5.4 on Terminal-Bench 2.0 · pass rate by agent

+29.8 points

between the lowest and highest scoring agent, same model.

Agent A
47.2%
Terminus-2
62.9%
Agent B
71.9%
Agent C
77.0%

Source: arXiv 2604.25850. Same model and task set; only the agent differs. Scores as published by the source. Agents are labelled with letters, and the same letter is the same agent across models; the source names them. Terminus-2 is Terminal-Bench’s reference agent. Cusp is not in this chart: we haven’t published a Terminal-Bench score.

Native edit formats

The edit format alone moves scores on public editing benchmarks. Each model gets the format chosen for it: str_replace, V4A patches or hash-anchored lines.

Reasoning that round-trips

Dropping reasoning traces costs agents up to 30% on published evals. Cusp passes each provider’s reasoning back on the next turn, in the form that provider expects.

A shell that works like a terminal

Non-interactive defaults, a real PTY for interactive programs, background jobs instead of timeouts, and bounded output.

Checks before it stops

When the agent thinks it’s done, Cusp checks diagnostics on the files it touched and sends failures back into the loop. A Stop hook can run your tests; a built-in test gate is coming soon.

Cache-stable context

Stable prefixes, observation masking before summaries, and big outputs spilled to files, so more of each turn hits the provider’s cache.

Tolerance for open models

JSON repair, schema coercion and text tool-call parsing, so a malformed tool call from GLM, Kimi, Qwen or DeepSeek is repaired instead of ending the turn.

Show three more

Checks before it stops

When the agent thinks it’s done, Cusp checks diagnostics on the files it touched and sends failures back into the loop. A Stop hook can run your tests; a built-in test gate is coming soon.

Cache-stable context

Stable prefixes, observation masking before summaries, and big outputs spilled to files, so more of each turn hits the provider’s cache.

Tolerance for open models

JSON repair, schema coercion and text tool-call parsing, so a malformed tool call from GLM, Kimi, Qwen or DeepSeek is repaired instead of ending the turn.

Crew · preview

A crew that works on a schedule.

Crew members are agents with a role, goals and a schedule. They run on your machine while the app is open; a cloud mode is coming. Spending is capped per member, and anything public waits for you.

Ava · CEO

Wrote the Q4 plan

Kai · Engineer

Reviewing PR #482

Sage · SEO

4 pages ready to publish

Milo · Marketer

Drafting the launch post

Iris · Designer

Pricing mockups ready

Theo · Analyst

Weekly retention report

Noor · Support

5 replies to review

Remy · Researcher

Competitor teardown

Schedules and triggers

Daily, weekly or on a GitHub mention, while Cusp is running.

Approvals where it matters

Anything public or irreversible waits for you.

Budgets per member

Limits on spend and tokens, per run, per day and per month.

Reports in your inbox

Standups and results arrive in your Inbox. Slack and email are coming.

One runtime

Desktop, web and terminal. Same agent.

One runtime powers every surface, so sessions, settings and history are the same wherever you pick them up.

Desktop

A native app for macOS, Windows and Linux. Parallel sessions, diffs, terminal and keychain-stored keys.

Download the app

Web and mobile

The same workspace in your browser, plus sync, phones and cloud sandboxes. Coming soon.

How sync works

CLI

A terminal UI, and a headless mode that speaks JSON for scripts and CI.

CLI reference

And where you already work

Pricing

Start free. Upgrade when you need more.

Free includes 2,500 agent actions a month on your own keys and a small monthly allowance of Cusp models. Paid plans lift the limit on your keys and add more Cusp models, sync, cloud runners and Crew automation.

Billing period

Free

Any model on your own keys, plus a little of ours. Needs a Cusp account.

$0with an account

Download
  • Desktop app and CLI
  • Your own keys: 2,500 agent actions a month
  • $1 a month of efficient Cusp models + $3 welcome credit
  • 2 synced devices, 30 days of cloud history
  • 3 crew members, manual runs
  • Self-hosted runners, free

ProPopular

Unlimited use of your own keys, with Cusp models included.

$16/ month

Billed yearly

Get Pro
  • Everything in Free
  • Cusp models every week, any model: 100% of a month = $10 at our cost
  • Your own keys: unlimited
  • Unlimited devices and history
  • 20 cloud runner hours a month
  • 10 crew members, 200 automated runs a day

Max

For people who run several agents all day.

$80/ month

Billed yearly

Get Max
  • Everything in Pro
  • Cusp models every week: 100% of a month = $50 at our cost
  • 100 cloud runner hours, 8 at once
  • Unlimited crew members, 1,000 automated runs a day
  • Priority support

Team

Shared agents, keys and guardrails for the whole team.

$25/ month per seat

2-seat minimum

Start with Team
  • Everything in Pro, per seat
  • Cusp models every week, pooled: $10 per seat a month at our cost
  • Shared org keys and shared sessions
  • 10 crew members per seat, pooled; 2,000 automated runs a day
  • Sign in with GitHub or Apple, passkeys
  • Google and OIDC SSO: coming Q1 2027
  • Priority support

Enterprise

Enterprise is open to design partners. Talk to us.

From $60 / seat / month · Annual, 25-seat minimum

  • Provider allowlists
  • Unlimited crew
  • Priority support
  • On the roadmap: SAML and SCIM (2027)
  • On the roadmap: audit log export
  • On the roadmap: runners in your VPC
Talk to us

How we price

  • Your plan's usage at provider prices

    Cusp models are counted at the model provider's published price for the tokens you use, cache discounts passed through. Usage beyond your plan is charged at list price × 1.05, which includes a disclosed 5% fee.

  • Your keys, no meter

    Bring your own keys on every paid plan, unlimited. Free includes 2,500 agent actions a month on them.

  • A percentage in the plan, dollars for money

    Your plan shows how much of this week you've used and when it resets. Credits, estimates on your own keys and invoices are always in dollars.

  • Receipts per task

    Every Cusp-model request is metered by user and model, and you can see it.

  • No 5-hour windows

    One weekly allowance: no daily token quotas or rolling lockouts. Rate limits are only there to stop abuse.

Prices are in US dollars and include any VAT we charge consumers. For business customers, VAT is added at checkout where it applies.

Compare every feature

Questions

Answers, plainly.

Something else on your mind? Read the docs or write to hello@cusp.run.

What does Cusp do?

Cusp is one coding agent that runs on any model or gateway and adapts its tools to each model family. Your CLAUDE.md, skills, hooks and MCP servers work as they are. It comes as a desktop app and a CLI; Crew is in preview, and the web app and cloud sync are coming soon.

Will my existing setup work?

Mostly, yes. Cusp reads AGENTS.md, CLAUDE.md and GEMINI.md, .claude/skills, .claude/agents, hooks, MCP servers and permission rules in place, and imports settings from the agents you already use. The import guide in the docs lists what isn’t supported yet.

Which models can I use?

Anything that speaks the Anthropic, OpenAI (Chat or Responses) or Gemini APIs: Claude, GPT, Gemini, GLM, Kimi, Qwen, DeepSeek, Grok, MiniMax, Mistral, plus OpenRouter, Vercel AI Gateway, LiteLLM, Bedrock, Vertex, Azure and local models through Ollama or LM Studio. Bedrock and Vertex work with API keys and bearer tokens for now.

What does the Free plan include?

With a free Cusp account you get the desktop app and CLI, 2,500 agent actions a month on your own keys (you pay your provider directly), and $1 a month of efficient Cusp models plus a $3 welcome credit. Paid plans add more Cusp models every month, unlimited use of your own keys, sync, cloud runners and Crew automation. Usage inside a plan counts at the model provider's published price; usage beyond a plan is charged at list price × 1.05, which includes a disclosed 5% fee.

Where do my code and keys go?

Sessions run on your machine and your API keys stay in the system keychain. With your own keys, requests go from your computer straight to your provider; Cusp models go through our gateway. For the Free allowance we count agent actions, without their content. The privacy policy has the details.

How much does the harness matter?

On Terminal-Bench, the same model scores up to ~30 points apart in different agents, per the published results linked in the chart above. We haven’t published a Terminal-Bench score for Cusp yet.

Try it on a real task.

Download Cusp, sign in, paste a key, and ask for something you’ve been putting off.

Download Cusp