Introducing CoreSpeed
An operating system for agents.
We're excited to introduce CoreSpeed to the world. For the vision behind it, read the letter from our founders.
Software was built for people. The next user is an agent. We're building the foundation it needs to work inside your organization — whichever agent you choose to run.
Today, agents run on bare metal.
Credentials get pasted straight into the process. Permissions get patched together for each agent. There is no shared filesystem — spin up another agent, and it starts without the files, memory, or configuration it needs. There are no resource quotas — a runaway loop can burn $300 over a weekend before anyone notices.
Decades ago programmers managed hardware directly too. Then the operating system moved in: one layer that owns the resources, mediates every privileged operation, and doesn't care which program you run on it.
That's what we've been building for agents. Model vendors give your agent its intelligence. CoreSpeed gives it everything else it needs to act safely in the real world.
One endpoint owned by your org
CoreSpeed gives any agent — local, cloud, or still running after your laptop closes — its full working kit through a single MCP endpoint.
https://api.corespeed.io/mcpThe unit of ownership is your organization. Grants, memory, budgets, and the audit trail live in the org. Bring in a different agent, and the shared context and connections are already there.
Agents are first-class citizens in CoreSpeed.
We believe agents should have their own place in an organization, with the same clarity about identity, access, and responsibility that we expect for people. That belief shapes how we build CoreSpeed.
Each Agent Identity has its own credentials, private memory, and a record of its actions, with a human owner responsible for it. It can use the connections shared with your organization, while your private connections stay private.
Your organization
AppsYour apps without the OAuth homework
Every service you connect to an agent today costs you the same ritual: register a developer account, create an OAuth app, wait through an approval queue, then babysit refresh tokens forever. You repeat it per service, and the result is welded to a single agent on a single machine.
CoreSpeed collapses the ritual into one action. Authorize an app once in the dashboard, and every agent in your org gets it through the same endpoint. Each member holds their own grants — the activity log shows who did what — and the same app can carry multiple accounts side by side. For services we don't host yet, register your own remote MCP server and it joins the same surface, under the same identity and the same log.
One property matters more than convenience: your agent never touches a raw credential. Tokens live in the platform's custody. The agent asks for an action; the platform performs it with the right account and writes it down. In August, Cloudflare open-sourced its own take on this pattern: credential-proxying "Gatekeepers", agents never holding tokens. Credential custody is trust infrastructure, not an integration convenience.
Memory that belongs to the team
An agent that forgets everything between sessions is a contractor you have to re-onboard every morning. CoreSpeed memory is durable and lives at two scopes: shared, where your team's conventions, decisions, and context compound; and private, for what only you should see. Any agent you sign in can recall it and add to it — the memory belongs to the org, and every agent you run works from the same one.
Batteries included
These tools are built into CoreSpeed. No extra provider accounts or API keys to manage; they draw from the same credits and appear in the same activity log.
Media
Generate and understand images, audio, and video. Upload media for your agent to work with, and get generated files back as links you can open and share.
Image, audio, and video. Same endpoint, same credits, same log.
Pick the model per job; nothing to sign up for.
Nano Banana ProImage
GPT Image 2Image
FLUX.2 [pro]Image
Seedream 5.0 ProImageNano Banana 2ImageFLUX.1 Kontext [pro]Image
Grok ImagineImage
Z-Image TurboImageGemini 3 Pro ImageImageFLUX1.1 [pro] ultraImageSeedream 4.5ImageFLUX.2 [klein] 9BImageSeedVR2 UpscaleImageVeo 3.1 FastVideo
Kling 3.0 ProVideoSeedance 2.5Video
MiniMax H3 MaxVideoKling 2.6 ProVideoSeedance 2.0VideoMiniMax H3VideoKling 2.5 Turbo ProVideo
ElevenLabs Eleven v3SpeechMiniMax Speech 2.8 HDSpeech
Qwen3-TTSSpeechSeed SpeechSpeechxAI TTSSpeechWeb
Search the web for source material and turn individual pages into clean Markdown your agent can read.
notes.
pricing and limits.
Social research
Explore public posts, profiles, comments, and trends across YouTube, Instagram, TikTok, X (formerly Twitter), Reddit, and RedNote. No social account needed.
YouTubeFrom meeting notes to follow-up tasks.
RedditWhat happens next?
X (formerly Twitter)Let me find the decision without replaying the whole call.
Budgets and a ledger
Agents are loops, and loops sometimes run away. Set a spending limit in CoreSpeed, and a runaway loop hits that limit before it empties your wallet. Every action and every charge appears in the activity log, so you can see what your agents did and what it cost. Pricing combines a subscription with usage-based credits.
Any agent can log in
Like any good OS, CoreSpeed doesn't care whose program you run. Claude Code, Codex, Cursor, a harness you wrote yourself — anything that speaks MCP signs in and gets the same kit. This matters most the moment your agent leaves the laptop: when it keeps working after the lid closes, when it's one of five running in parallel, when it works for a team. Wherever it runs and whoever made it, the accounts, the memory, the budget, and the log are already there — because they never lived in the agent to begin with.
Turn this week's bug reports in the support inbox into Linear issues.
GmailLinearWhat's next
An agent needs an identity, a place to keep its state, somewhere to run, and ways to act in the world. An Agent OS brings these together and manages what the agent may do, what it may spend, and the record of its work. That's the full CoreSpeed we're building.
Today's release puts the first parts of that system in your hands. Here's the bigger picture, including what's still ahead.
Agent OSSmart approvals will be the next piece: describe in plain English what your agents may do, what needs your approval, and what must never happen. CoreSpeed will turn those instructions into rules you can review and enforce them before an action runs.
Post our launch update to X.
We’re introducing CoreSpeed, an Agent OS. Bring your agent, connect your apps, and put it to work.
Get started
set up https://corespeed.io/SKILL.mdIt adds CoreSpeed and guides you through sign-in.
Connect the apps you need in the dashboard, then ask:
Summarize this week’s Linear blockers and post them to #standup.
The docs are at corespeed.io/docs, and, fittingly, every page — this one included — is also served as plain markdown for agents: /llms.txt.
We're building this in public, one component at a time. Come tell us what your agents need next — the team is on Discord.
Send the action items straight to my task tracker.