proofable/sdk, proofable/mcp, and proofable/docs.
Published APIs and verifier schemas may change between minor versions. Breaking changes are listed in each entry and in the repo changelogs.
MCP 0.1.11: 18 tools on every connection
- One list everywhere. Every connection serves the same 18 tools, signed in or out, in every client. Verify users, give AI agents their own ID and limits, and hand them work you can see and stop. ChatGPT serves 8 of them.
- Runtime work moves into jobs. Web search, page fetch, the browser, connected apps, remote MCP servers and saved keys run inside a job, within the agent’s permission. A client that calls one directly gets unknown tool. Hand the work to an agent with
proofable_job_createinstead. - Only the owner approves. A run waiting for approval is decided by its owner. The job’s own agent cannot approve it.
- Upgrade: reconnect or restart your client so it reloads the tool list.
Benchmark page: PrecisionMemBench
- 77 of 77 cases, precision 0.98, recall 1.00 on PrecisionMemBench, the public test of whether a memory returns the current fact and nothing stale, run with the Agent Memory Benchmark toolkit. The row sits beside the published Hindsight run (66 of 77) and the benchmark’s own table, with the per-case results and the code that ran.
- Docs read in Geist, the same face as the product. The serif body face is gone.
Rate limits now scale with your plan
- New proofs scale with the plan. Creating proofs was capped at 50 per 15 minutes on every plan. It is now a per-minute budget: 30 a minute on Free, 120 on Pro, 300 on Enterprise. API calls stay at 60, 240 and 600 a minute.
- The budget follows the account that pays. A signed-in account or access key is counted on its own wallet, including when it writes into one of its agents’ vaults. A wallet-signed request with no session is counted on the wallet that signed it.
- One count across the service. Limits are counted once for the whole service rather than per server, so the number in
RateLimit-Policyis the number you get.
Talk to a hosted agent from any app, and read your spend
@proofable/sdk@0.1.11 adds two client calls.client.agentChat()talks to a hosted agent. Name the agent and send the conversation. The turn runs on Proofable as that agent, under the delegation you signed for it, and every tool call is checked before it runs. Events stream toonEvent, and the call resolves with the answer, the records the agent read ascitations, and the credits it spent. See Talk to your agent from any app.client.getUsage()reads your credits and spend: the balance, recent spend events, and spend by line and by UTC day, for any period withfromandto.
A ChatGPT endpoint, and keys stay on your profile
@proofable/mcp@0.1.8 adds https://mcp.proofable.me/mcp/openai, which serves the eight read and guided-verification tools to the ChatGPT plugin with the same OAuth sign-in. Cursor, Claude Code, Codex and every other client keep the full tool list on https://mcp.proofable.me/mcp.Breaking: /mcp no longer accepts a key as a tool argument. The tool that saved one is gone from tools/list, and a call to it is rejected as an unknown tool. Save keys on your profile; listing, removing and calling a provider with a saved key still work, and never return the key.The release also adds examples/injection-test, which hides the five AgentDojo attack templates in a note your agent reads and grades every run from its signed receipt.Verify any CAIP-380 envelope, import an agent in one step
@proofable/sdk@0.1.10 covers 0.1.7 through 0.1.10.- Any CAIP-380 envelope verifies (0.1.9).
verifyPortableProofEnvelopetakes the signer from the envelope’sdid:pkhidentifier, so an envelope withoutwalletAddress, including the specification’s test vectors, verifies instead of throwing. Use 0.1.9 or later to verify envelopes from other implementations. proofable importcreates the agent (0.1.10). A signed-in terminal creates the agent on your profile with its full instructions and mounts it into the current project.--expires-atsets an end to its permissions,--no-mountskips the mount, and--dry-runshows what would happen.proofable auth --oauthsigns in the CLI only (0.1.10). It no longer writes its short-lived token into your editors; pass--clientto write it to one.fulfillGate()completes AI service orders (0.1.7). It sendsserviceParametersfor the listing’sserviceInputs, and the run’s answer appears on the buyer’s receipt atmeta.marketplace.serviceJob.result.- The CLI exits with code 1 after an error on Windows (0.1.8), and a dry-run mount lists files as “would write”.
A named agent uses only its own delegation
@proofable/sdk@0.1.6 will not apply a delegation to a named agent unless that delegation names the agent. A mounted agent no longer carries an unused runtime field, so the host that runs it picks its own runtime. The packaged skills now match @proofable/mcp@0.1.7, including the external-signer prepare-and-submit flow.One listing description everywhere
@proofable/mcp@0.1.7 uses one description across OpenAI, the MCP Registry, npm, Claude, Cursor, and Codex. It opens with “Context and control for any AI,” followed by the identity, access, tools, and proof details. Earlier releases carried different copy on each platform.Proofable naming, current skills, typed portable proof
@proofable/sdk@0.1.5 carries the fixes since 0.1.4. The personal_sign wallet error now names Proofable instead of NEUS, the CLI --live help no longer assumes an IDE, and the packaged skills match the proofable/mcp 0.1.6 set. A mounted agent no longer carries an unused skills field, and types.d.ts declares portableProof and options.disclosePortableProof.One plugin listing, one version line
@proofable/mcp@0.1.6 aligns the ChatGPT and Codex plugin manifests on one shared display name, description set, policy URLs, and starter prompts.The SDK and MCP also version independently now. @proofable/sdk@0.1.5 and @proofable/mcp@0.1.6 are both current; a change to one does not force a version bump on the other.A delegation with no actions grants nothing
@proofable/sdk@0.1.4 makes an empty or missing allowedActions list grant nothing to a delegated agent, matching the hosted contract. A delegated host action is denied with ACTION_NOT_ALLOWED; mounts you own keep their existing behavior. See Run your first guarded action.Checks are free
Checking an existing proof now costs 0 credits. You pay only when you create a new proof.GETandPOST /api/v1/proofs/checkare free. Rate limits per plan are the only control.- Per-request surcharges for verifier count, query complexity, time window, and result limit are removed.
/api/v1/proofs/checkis no longer an x402 pay-per-call resource. Proof creation and share grants still are.- Access grants are a flat 2 credits. Duration tiers (
perHour,maxDurationBonus) were removed; the grant’s own duration limit is enforced where the grant is created.
Agent authority, listed explicitly
An agent’s authority is what its delegation lists, not a scope string.allowedActions minus deniedActions is the grant, and a delegation with no listed actions grants nothing.- Delegations require
allowedActionsand only accept canonical action names. - Spend authority is bounded:
make_paymentneedsallowedPaymentTypesandmaxSpendtogether. - App agents list the web origins they may act from in
allowedOrigins, or"*"for servers. - Revoked, superseded, and expired delegations grant nothing.
Builder path without a listing
Define the checks in your app. Check the visitor by account. Reuse a current proof. Verify only what is missing.defineGate+gateCheck({ gate, subject })does not create a Proofable profile.POST /api/v1/proofs/checktakes the same x402 pay-per-call rail asGET.- MCP sign-in stays OAuth. Machine API calls stay x402, sponsor, or credits.
gateCheck({ gateId })still works for published listings.
One description per public surface
GitHub, npm, the MCP Registry, and product pages no longer repeat the product line. Each surface now carries its own description, so search and directories show what that surface is instead of one slogan everywhere.- MCP package and registry: Give AI agents identity, scoped access, trusted context, and verifiable actions through MCP.
- SDK package: SDK and CLI for verification gates, reusable proof, and agent permissions.
- Docs: Docs for verification, gates, proofs, agents, and MCP.
- The category line stays the remembered brand: The trust harness for AI.
- Release notes now summarize the predecessor
@neusreleases in one readable history section.
First Proofable release
Connect your AI client once. Carry agent identity, limits, and current proof into the next client or app.- Connect context, tools, and authority to your profile.
- Reuse a current proof instead of repeating the check.
- Give each agent an identity, a spend limit, and permissions you can revoke.
- Require current proof before access or a protected action.
@neus packages, see the migration guide.Breaking changes
@neus/sdkis now@proofable/sdk.@neus/mcp-serveris now@proofable/mcp.- The CLI binary is
proofable(wasneus). Runnpx -y @proofable/sdk setup. - MCP tools are named
proofable_*.tools/listadvertises only these names. - The MCP config key is
proofable. The token store moved from~/.neus/to~/.proofable/, so sign in once.
History
Before the Proofable packages, the same product shipped as@neus/sdk and @neus/mcp-server. The notes below are the highlights of each predecessor release. The complete record stays with those packages on npm and in the archived source at proofable/network under the neus-final tag. To move an @neus install to Proofable, see the migration guide.