Everything your AI agents need. One platform.
Connect your models, MCP servers, apps and built-in tools through one key. Route every request, share memory between agents, and keep permissions, approvals and spending in one place.
$1 for your first 7 days, then $250 a month.
ObolOne key
- 01Models
- 02Tools
- 03Built-insRolling out
- 04WorkersEarly access
- 05Memory
- 06Rules
- 07Records
- 08Copilot
This key · $38.20 of $200 / month
Support agentTicket #4821
2 of 4 handled
- 01
Read the customer’s message
Claude, on your Anthropic key
DoneDone - 05
Found last week’s notes on this order
DoneDone - 02
Refund $180 in Stripe
Waiting for your OKWaiting for your OK06Your rule: Every refund in production waits for your OK
- 02
Delete duplicate customer record
Up nextUp next06Your rule: Never delete customer records
Works with Claude Code, Codex, Cursor, VS Code and your own agents.
AI agents need more than a model. Obol gives them the rest in one place.
Connect what your agents use, route every request, and control it all from one dashboard — with the agents you already have.
Models & routing
One address for the AI models you use, on your own provider accounts, with optional backup models when one is down.
- Busy
Main
Your OpenAI account
- Running
Backup
Your Anthropic account
Tools & MCP
Your apps, MCP servers and APIs behind one connection, and agents only see the tools you allow.
- Search the web with ExaAllowed
- Read GitHub issuesAllowed
- Refund in StripeNot shown
Built-in tools
Rolling outWeb search, page reading and text from images, run by Obol, with no extra accounts to open.
- Web search1¢
- Read a web page0.5¢
- Read text in an image0.5¢
- Generate an imageComing
Included each month: $20.00 on Obol Cloud, $1.00 during the $1 trial.
Workers
Early accessAI workers that run real jobs in managed code and browser environments. Not yet on Obol Cloud.
Check our website for broken pages and list them.
- Managed browser
- Code sandbox
Shared memory
Notes your agents can search and add to, so the next session starts where the last one stopped.
- Refund steps for ticket #4821Shared
- Draft reply to the customerPrivate
Rules & approvals
Nothing runs unless your rules allow it, and risky actions in production wait for a person's OK.
- Refund $180 in Stripe, in productionWaiting for your OK
- Read issues in an allowed GitHub repositoryAllowed
- Delete a customer recordBlocked by your rule
Records & spend
A record of every app action, what your agents spent, and monthly limits that stop new calls.
$38.20of its $200 monthly limit
- Refund $180 in StripeSupport agent · 10:42Done
Copilot
An assistant in your dashboard that proposes setup changes for you to review, test and apply.
Let Support agent read messages in Slack.
- Review itDone
- Test against last monthDone
- ApplyWaiting for your OK
One way to reach every model you use. On your own accounts.
Point your apps at one address and keep the OpenAI format they already use.
- Use your own accounts. OpenAI, Anthropic, OpenRouter, or any service that speaks the OpenAI format. Your provider bills you directly, and Obol adds no markup.
- Choose which models each agent and each key may use.
- Optional: set rules for which model handles which request, with backup models when one is down or busy.
- See tokens used and estimated cost by provider, model and key. Anything Obol can't price is marked, not hidden.
Takes over when the primary is down or busy.
Anthropic is busy, so requests for claude-sonnet-5 go to the backup, OpenRouter.
Every tool your agents need. Through one connection.
Connect GitHub, Slack, Stripe and HubSpot today, plus web search (Exa, Brave Search), code sandboxes (E2B) and browsers (Browserbase).
Apps
- GitHub, hidden from Support agent
- Slack, Support agent can use it
- Stripe, Support agent can use it
- HubSpot, Support agent can use it
Web search
- Exa, Support agent can use it
- Brave Search, hidden from Support agent
Code sandbox
- E2B, hidden from Support agent
Browser
- Browserbase, hidden from Support agent
Yours
- Your MCP server, hidden from Support agent
- Your API, Support agent can use it
Support agent can see 5 tools: Slack, Stripe, HubSpot, Exa, Your API.
- Bring your own MCP server. Obol applies the same rules and keeps a record of each call.
- Add your own internal API from its API description file, and review each action before agents can use it.
- Your agents only see the tools they're allowed to use.
- Agents can search for the tool they need instead of loading every tool up front — an option you turn on per key.
- Tool Radar shows which routed tools are failing or slow, with typical and worst-case response times.
Two routers. Two different jobs.
Model routing and tool routing sound alike. They answer different questions.
Model routing
Optional“Which of my providers should answer this model name?”
You set rules per model name; backups take over when a provider is down or busy.
- AnthropicBusy
- OpenRouterAnswering
- Another modelNot allowed for this key
Tool routing
“Which of my connected apps should do this job?”
When more than one of your connected apps can do the same job — like web search with Exa or Brave Search — Obol picks one by your rules and each app's track record.
- ExaPicked
- Brave SearchStanding by
- Another search appNot allowed by your rules
Routing never overrides your rules. A tool your rules don’t allow can’t be picked.
Built-in tools. No extra accounts to open.
Web search, page reading and reading text from images, run by Obol and called with your Obol key.
Obol built-in tools
Rolling out- Run by
- Obol
- Called with
- Your Obol key
- Priced in
- USD
Tool
Price per action
Web search
1¢per search
SearXNG · Open source, operated by Obol
or Brave Search API, a hosted API on Obol's accountRead a web page
0.5¢per page
Scrapling · Open source, operated by Obol
Read text in an image
0.5¢per image
Tesseract OCR · Open source, operated by Obol
Generate an imageComing
2¢per image
FLUX.1 [schnell] on MuAPI · Hosted API: MuAPI, on Obol's account
When it arrives, it waits for your OK in production unless your rules allow it.
Included each month
- Obol Cloud
- $20.00
- During the $1 trial
- $1.00
Of built-in tool use, shared across your organization. After that, your prepaid balance pays, if you have one.
The same controls. Built-in tools follow the same rules, approvals and records as any other tool your agents use.
Your admin decides. A workspace admin switches them on or off. Workspaces created before built-in tools existed start with them off.
Not in Free Self-hosted. They run on Obol Cloud. A server you run yourself doesn't include them.
Rolling out. Built-in tools are rolling out to Obol Cloud. Image generation is coming.
Run agents, not just connect them. Obol Workers, in early access.
Describe a job and Obol runs it in a managed browser or code sandbox, under the same rules, approvals and records.
Obol Workers
Early accessNot yet on Obol Cloud
Job
Check example.com for broken links
- 1Open example.com in a managed browserDone
- 2Visit 42 pages and check every linkDone
- 3Find 3 broken links, on /pricing and /blogDone
- 4Open a GitHub issue listing themProduction: your rule holds new issues for your OK.Waiting for your OK
- 5Post the summary in Slack
Early access. Obol Workers is not yet on Obol Cloud.
Describe, review, publish. Tell Obol the job in plain words. It drafts the worker, you review the draft, and you publish it.
Start from a template. Starter templates cover website checks, coding tasks and research.
Your rules still apply. Each run uses its own short-lived key, so your rules, approvals and records cover everything it does.
Priced by the minute. 2¢ per active minute ($1.20 per hour), paid from your prepaid balance. Your plan’s included built-in tool use never pays for it.
Memory that outlasts the session. Shared on your terms.
A shared notebook for your agents, so the next session starts where the last one stopped.
Ticket #4821
- Support agent
- Read only
Private until you share it.
Acme Corp account
- Support agent
- Can write
- Billing agent
- Can write
- Sales agent
- Read only
Billing agent’s change was sent back: it started from version 2.
Workspace memory
search “acme renewal” · 1 card
Private unless shared. Notes stay private until you share them.
Writing is yours to allow. Agents write to shared notes only when you allow it.
No silent overwrites. Edits can't silently overwrite each other. A change made from an older version is sent back.
Search by keyword. Agents find notes by the words in them.
Its own access. Memory has its own short-lived access, separate from your agents' main key.
Real controls for real actions. Your agents can't overrule them.
- HubSpot ticketsRead tickets and post replies.Allowed
- GitHub issuesOnly these GitHub repositories: acme/web, acme/helpAllowed
- Customer recordsNever delete customer records.Blocked by your rule
- Stripe refundsEvery refund in production waits for your OK.Waiting for your OKWaiting for your OK
Refund $180
Support agent · ticket #4821 · Stripe
RuleEvery refund in production waits for your OK.
0 of 1 approvals
- Nothing runs unless your rules allow it. Set rules per agent, per tool and per environment — for example, only these GitHub repositories — and test them against past activity before you publish.
- Your agents only see the tools they're allowed to use. Everything else stays hidden from them.
- Risky actions in production wait for a person's OK — one or two approvers — and run once after approval.
- A retry or an approval never makes the same change twice.
support-agent
The Support agent's own key
- Environment
- prod
- Models
- gpt-4.1-miniclaude-haiku-4-5
- Tools
- Stripe refunds, Slack messages
- Monthly limit
- $200.00
- Expires
- 2027-03-31
- App logins
- Locked in the vault. This agent never sees them.
On
- Each agent gets its own key, limited to certain models and tools, with a monthly spending limit and an expiry date.
- Switch a key off and new calls stop within a minute. Or stop every agent in a workspace with one switch.
- Development, staging and production stay apart. A test key can't reach production.
- Your logins and API keys stay locked in an encrypted vault, unlocked only at the moment of the call. Your agents never see them.
Obol can't make an AI choose the right action. It makes sure an action your rules don't allow can't run.
- 09:41:03DoneLook up the customer in HubSpotCost: —Support agent · Rule: HubSpot tickets: read and reply
- 09:41:20DoneRefund $180 in StripeCost: —Support agent · Rule: Every refund in production waits for your OK
- Who asked
- Support agent, for ticket #4821
- Which rule
- Every refund in production waits for your OK. Approved by Dana.
- App answered
- Stripe: refund created, $180.00
- Obol saw
- The request and Stripe's answer, for itself
- 09:41:26Blocked by your ruleDelete a duplicate recordCost: —Support agent · Rule: Never delete customer records
- 09:52:11DoneOpen an issue in acme/webCost: —Triage agent · Rule: Only these GitHub repositories
- 10:04:47DoneRun code in a sandboxCost: Not pricedResearch agent · Rule: Research agent: code sandboxes
AI requestsmetered today: 14 requests · 18,240 tokens · est. $0.31
- A record of every app action your agents take: who asked, which rule allowed it, what the app answered, and how much of that Obol saw for itself.
- Every AI request is metered: tokens used and estimated cost.
- Export the record of every action.
- support-agent$38.20 of $200
- research-agent$96.40 of $150
- nightly-report$50.00 of $50Limit reached · new calls stop
Estimated costs from your providers. Obol's own fees are shown on their own line.
- See spend by provider, by key and by agent, with Obol's own fees shown apart from your vendors' estimated costs.
- Anything Obol can't price is marked, not hidden.
- Give each key a monthly spending limit. When it's reached, new requests and actions stop — nothing runs over.
- Your plan's monthly limits pause new calls instead of charging overage.
Set it up by asking. Copilot proposes. You approve.
Copilot lives in your dashboard. It helps you set up apps, rules and keys, and explains what your agents did and spent.
You asked: “Let the support agent refund in Stripe and reply in Slack, but nothing risky without me.”
- AddConnect SlackReplies go to #support
- AddKey support-agent, limited to two toolsStripe refunds, Slack messages
- ChangeRule: refunds in production wait for your OK
Tested against the last 30 days · 0 past calls would change
- Copilot proposes each change for you to review. Nothing changes until you apply it.
- Test it against the last 30 days of activity before it applies, and see which past calls would have gone differently.
- Apply it — and undo it if you change your mind.
- On Obol Cloud, Copilot uses a service-only Claude Sonnet 5 route. Obol funds a $20 weekly allowance per requesting workspace, shared with Obol’s background evaluations.
- That route never serves customer application traffic. Your agents still use your own provider accounts.
See it all work together. Try it in the Playground, included with the $1 trial.
Give an agent a task, then change a rule and watch the decision change. Approve or decline what it asks to do, lower its spending limit, and read the record each action leaves.
The apps in the Playground are simulated, so nothing real is refunded, emailed or changed.
- 01Give an agent a taskRunning
- 02Change a rule and run it againBlocked by your rule
- 03Approve or decline what it asksWaiting for your OK
- 04Read the record each action leavesDone
Works with the tools you already use. Keep your agent.
Your agent plugs in. As it is.
- Claude CodeApps over MCP
- CodexApps over MCP
- CursorApps over MCP
- VS CodeApps over MCP
- The OpenAI SDKAI models
- Your own agentModels and apps
Step-by-step setup for each lives in your dashboard.
Your apps connect. With your own accounts.
- GitHubCode and issues
- SlackTeam messages
- StripePayments
- HubSpotCustomer records
- ExaWeb search
- Brave SearchWeb search
- E2BCode sandboxes
- BrowserbaseBrowsers
- Your own MCP servers and internal APIsThe same rules, and a record of each call.
Start on Obol Cloud. Or run it yourself.
The same rules, approvals and records either way. The difference is who runs it.
Obol Cloud
We run it for you. Sign up, connect your apps and set your rules in the dashboard.
- Hosted and kept up to date by Obol
- The dashboard for your rules, approvals, records and spending
- Monthly plan limits that pause new calls instead of charging extra
- Built-in toolsRolling out
- Obol Workers, not yet on Obol CloudEarly access
Free Self-hosted
Run Obol on your own server. You operate it: upgrades, backups and the encryption keys, which Obol can’t recover for you.
- Runs on your own server with Docker Compose
- $0 in Obol fees, for commercial use too, with no card and no expiry
- The same rules, approvals and records
- No support contract or uptime promise
- Obol Workers (early access) and built-in tools are not included
Obol Workers and built-in tools are Obol Cloud services. They are not part of Free Self-hosted.
Start for $1. Then one monthly price for your whole team.
Every plan comes with your rules, approvals for risky actions and a record of every app action. A bigger plan raises your limits; it never changes your rules.
Obol Cloud
For teams that run AI every day.
$250per month
Per organization · no seat surcharge
$1 for your first 7 days, then $250 a month. Cancel before day 7 and you pay only the $1.
- AI requests
- 100,000
- Actions in your apps
- 100,000
- of which change data
- 25,000
- of which are risky (refunds, deletes)
- 2,500
- Built-in tool useRolling out
- $20.00
One price for your whole team, before tax. At a limit, new calls pause instead of costing extra: $0 overage.
Enterprise
For companies with security reviews and contracts.
$5,000per month
USD · annual agreement, signed with our team
- Limits for every kind of action
- Payment terms
- Support for running Obol on your own servers, if agreed with sales
Your AI and app providers still bill you directly.
Your model and app providers bill you directly. Workers compute and built-in tools have their own prices.
Obol Workers and built-in tools are Obol Cloud services. They are not part of Free Self-hosted.
Workers compute
Early access2¢per active minute ($1.20 per hour)
Obol Workers is in early access and not yet on Obol Cloud. It is paid from your prepaid balance; the monthly included built-in tool use never pays for it.
Built-in tools
Rolling out0.5¢ to 2¢per action
Obol Cloud includes $20.00 of use each month for the tools Obol runs itself, which are rolling out, and the $1 trial includes $1.00; after that, your prepaid balance pays, if you have one. See what each action costs
The $1 trial
The trial lasts 7 days and costs $1, charged at checkout to the card you add. On day 7, Obol Cloud starts at $250 a month on that card. Cancel from Billing before day 7 and you pay only the $1; the $1 is not credited toward the subscription.
During the trial you get Obol Cloud’s call limits and $1.00 of built-in tool use; the $20.00 starts with your first paid month. One trial per organization.
Organizations that already have Free keep it, and their upgrade starts with the same trial.
Obol Cloud
Included usage is shared across your organization’s workspaces and resets each paid monthly subscription period.
Only dispatched calls spend the allowance. At any allowance, new calls pause until the next paid period; dashboard and audit access remain. No automatic overage and no unused allowance rollover.
$250/month Obol spending cap before tax. $0 overage.
Enterprise
- Annual platform commitment
- Contracted annual limits for each of the four call classes
- Enforced limits for every call class
Annual volumes cover model requests, read-only tools, writes, and destructive tools. Unused included volume does not roll over.
The platform commitment is $60,000 per year, equivalent to $5,000 per month, before tax. Call limits and payment terms come from your signed agreement. Contracted support for self-host, Kubernetes, VPC, or air-gapped deployments is confirmed with sales before it enters a contract. Existing design-partner agreements keep their signed terms.
Every plan
On Obol Cloud, destructive calls also count as writes and tool calls. Actions your rules block, and replays of an action that already ran, use no allowance. Upgrading never changes your rules or which actions wait for your OK.
Spending caps cover the Obol plan before tax. Your vendor bills and optional Workers compute are separate; Workers is in early access.
The included built-in tool use pays for the tools Obol runs itself, each priced per action: see what actions cost. It is shared across your organization’s workspaces and resets on the first of each month (UTC). Workspaces that existed before built-in tools start with them off until an admin turns them on. Image generation needs an approval in prod unless your policy grants it, and is not available yet. Built-in tools are rolling out to Obol Cloud.
Questions, answered plainly.
The short answers are here. The technical ones are on the developers page.
Technical questionsObol is one platform for everything around your AI agents. Connect the models, MCP servers and apps they use through one key, and keep your rules, approvals, spending limits and a record of every app action in one dashboard.
No. Keep the agent you have, and point it at Obol with one Obol key. The dashboard has step-by-step setup for Claude Code, Codex, Cursor and VS Code, or any app built on the OpenAI SDK, and your apps keep the OpenAI format they already use.
You do, directly. Obol works through your own accounts: your OpenAI, Anthropic or OpenRouter account, or any service that speaks the OpenAI format, and your own app logins. Each provider bills you at your rate, and Obol adds no markup. Obol bills you separately, for its own plan.
You pay $1 for your first 7 days of Obol Cloud, with a card. On day 7, Obol Cloud continues at $250 a month on the same card. Cancel before day 7 and you pay only the $1. The trial has Obol Cloud's monthly call limits, $1.00 of built-in tool use and the Playground.
No. Nothing runs unless your rules allow it. Your agents only see the tools they’re allowed to use, anything outside your rules is blocked before it runs, and risky actions in production wait for a person’s OK unless your rules allow them. Obol can’t make an AI choose the right action. It makes sure an action your rules don’t allow can’t run.
New requests and actions stop, so nothing runs over. You can give each key its own monthly spending limit. Your plan’s monthly limits work the same way: at a limit, new calls pause instead of costing extra until the limit resets or you upgrade, and your dashboard and records stay open.
In an encrypted vault. Your logins and API keys are unlocked only at the moment of a call, and your agents never see them: they hold an Obol key instead.
Obol Workers, where you describe a job and Obol runs it for you in a managed browser or code sandbox under the same rules, is in early access and not yet on Obol Cloud. Built-in tools for web search, reading web pages and reading text in images are rolling out, and image generation is coming. The Playground, included with the $1 trial, lets you try an agent first. Its apps are simulated, so nothing real is refunded, emailed or changed. Open the Playground.
Yes. Free Self-hosted runs on your own server with Docker Compose for $0 in Obol fees, with the same rules, approvals and records. You operate it, with no support contract or uptime promise. Obol Workers and built-in tools are Obol Cloud services. They are not part of Free Self-hosted. Compare Cloud and self-hosted.
Built-in tools are the tools Obol runs itself, which are rolling out, and each action has a fixed price: a web search is 1¢. Obol Cloud includes $20.00 of built-in tool use each month and the $1 trial $1.00, shared across your organization, and after that your prepaid balance pays, if you have one. Your model and app providers still bill you directly. See what each action costs.
MCP, the Model Context Protocol, is a common standard that lets AI agents use outside tools and apps through MCP servers.