Repository navigation
Releases
v0.1.7
Run durable service workflows, review service-defined actions, and separate template publication from deployment.
New
- Author typed workflow graphs with generated tool call descriptors, branching, mapping and pure control functions
- Inspect workflow graphs, schemas and quotes, start durable runs, and read status, continue or cancel them through crab workflow and the SDK
- Retain successful checkpoints across explicit continuation, enforce cumulative owner spend caps, and settle customer task fees once on success on supporting hosts
- Declare UTC workflow schedules with stable occurrence identities; scheduled runs do not retry failures or backfill missed occurrences
- Discover and control accessible workflows through MCP, and inspect services through the local host MCP endpoint
- Request browser review of service-defined actions with the Outbox email example, and follow their outcomes with crab request --wait
- Register Relay agent handles and exchange messages between agents
- Package the CrabTalk plugin with its shared service-use skill and OAuth connection configuration
- Add a runnable Page Watch workflow example that reads webpages, compares snapshots and saves changes
Changed
- Use crab push to build and publish or withdraw templates without changing running instances; crab publish is removed and crab deploy remains the deployment command
- Rename crab upgrade to crab update
- Embed public visibility in service images for deployments and template copies; remove the crab visibility command and SDK visibility setter
- Move installation, documentation and hosted service URLs to crabtalk.ai
Fixed
- Accept SDK-issued secret resource IDs in guest requests
- Wake Relay delivery on demand instead of polling by cron, and use account access for personal identities
- Stop newly scaffolded services from claiming memory catalog jobs for their shared remember and recall state
- Refresh reviewed provider source pins for clean service builds
- Clarify database transaction visibility and commit behavior in service development documentation
v0.1.6
Release idle service code automatically, report database allocation, and preserve template storage defaults.
New
- Expose database allocation, capacity and measurement time in live usage, history and SDK snapshots
Changed
- Bound cached executable images to 64 MiB of accounted bytes and expire them after five idle minutes by default; reload cold services on demand
- Restore deployment metadata without retaining executable code and enable the host disk compilation cache by default
- Show database allocation in crab usage, with missing measurements displayed as an em dash
Fixed
- Preserve the declared database size in built images so published copies use their template storage default
- Keep cold services discoverable and scheduled work runnable after restarts, and retain active memory accounting through redeployment and removal
- Reclaim idle cached code before memory refusals and preserve E0019 when cold code exceeds the account allowance
- Share byte and price formatting between the CLI and SDK
- Clarify Relay template deployment, agent access and Grok host permissions
v0.1.5
Publish and deploy services by image address, call running services across accounts, and ship smaller release binaries.
New
- Address service calls as owner/service with berm_lang::call, preserving qualified service caller identities
- Retain the target owner in generated bindings from crab idl owner/service
Changed
- Replace name in crab.toml with image = "owner/service"; source-only services use their Cargo package name
- Build and publish to the image address with crab publish; crab deploy consumes the published version under the selected account, while --build selects local source
- Resolve directly selected published images without Cargo or Rust; standalone manifests require an exact version suffix
- Remove calls declarations from service manifests
- Strip symbols from release binaries
Fixed
- Use main for service source links and released tags for Rust dependency examples; correct moved provider source paths and source-deployment commands
- Restore clean builds of TinyFish and Litescrape against their reviewed OpenAPI documents
v0.1.4
Build services without manual Rust setup, start with durable state, and inspect pending deployment approvals. Device login requests dedicated agent credentials; webpage tools distinguish HTML titles from SVG titles.
New
- Show pending approval links alongside recorded attempts with crab deploys, including before a service is first deployed
- Expose agent login authorizations and pending deployment approvals through the SDK
Changed
- Request a dedicated agent API key for device login, with its account, permissions and daily budget reviewed in the browser
- Include durable remember and recall tools and a 64 MiB storage reservation in crab new scaffolds
Fixed
- Install missing Rust build tools and the WebAssembly target automatically, making installed tools available to the same build invocation
- Check local host status through its health endpoint and explain that crab plans requires a Cloud login
- Exclude SVG titles from webpage titles and SEO title counts
v0.1.3
Measure service budgets with reproducible Wasm benchmarks, select template versions and control automatic patch updates. The SDK exposes release history, upgrade policies, yanking and template deletion; Markdown and SEO are public services.
New
- Measure CPU, rows, peak memory, elapsed time and estimated costs with crab bench, using repeatable JSON Lines cases and arranged HTTP responses
- Select an exact template release with crab deploy --from owner/name --version, and opt into stable patch updates with --auto-upgrade
- Read retained template releases and manage upgrade policies, release yanking and template deletion through the SDK
Changed
- Configure web/markdown and web/seo as public running services without publishing templates
- Document measured budget sizing, required headroom and destination host ceilings before deployment or publication
Fixed
- Bound webpage parsing by nodes, nesting, tags and attributes, and size Markdown and SEO tool budgets from WebAssembly workloads
- Keep OSV tool CPU budgets within the default 100 ms host ceiling after measuring large responses and maximum-size batches
v0.1.2
New agent, web and ops services join Connect integrations. Deploy and publish catalogs by directory, configure service logos, and manage API keys with granular permissions. Schedules and secret destinations now belong to the service image.
New
- Share revision-checked, namespaced agent notes with the memory service
- Inspect SEO metadata, extract webpage Markdown and collect website analytics with the web services
- Monitor endpoints and receive signed, deduplicated events with the uptime and webhooks services
- Declare an HTTPS service logo in crab.toml, retained by published templates and private copies
- Show purchased and trial credit separately in crab billing, including trial expiry, reserved funds and the hosting recovery deadline
- Expose plan credit allowances, eligible account kinds, trial terms and organization creation status through the SDK
- Replace an API key's priced-service grants through SDK key updates
- Add host invocation scopes for nested and scheduled calls, preserve admission refusal statuses and identify outbound HTTP requests by chain and span
- Check on every pull request that changelog.json contains release notes for the current crab-agent version
Changed
- Organize services under connect, agents, web and ops catalogs, with reusable libraries under services/kits
- Scope crab deploy and crab publish to services beneath the supplied directory; -p still selects across the workspace, and publish includes services marked deploy = false
- Replace deploy and auto API-key scopes with resource read/write permissions and remove crab key roll; write permission includes read access to the same resource
- Embed cron schedules in the image and apply them on every deploy; change crab.toml and redeploy instead of editing schedules on the host or using deploy --pull
- Fix credential destinations in the manifest and require fresh provisioning when they change; remove secret host editing and never return saved secret values
- Fetch provider OpenAPI specifications into OUT_DIR at build time, with pinned revisions or reviewed digests
- Use Cargo default members for routine CI tests and linting, and prepare shared Wasm fixture images with nextest setup scripts
Fixed
- Reject provisioning undeclared secrets and refuse outbound secret substitution to destinations absent from the manifest
- Check the Context.dev OpenAPI source against a reviewed SHA-256 digest
v0.1.1
Kits replace guests: auth, crawl, lake, neon, stripe and supabase link into a service's own program. Calls are metered and bounded by CPU time, databases by the size berm.toml gives them, and crab installs and upgrades itself from GitHub releases.
New
- Link a kit into a service by depending on it: a kit declares its secrets, hosts and migrations in a berm.toml of its own, the build finds each kit among the crates compiled into the image, and what the kit declares joins the service's manifest and its tables live in the service's database
- Link kits through crabkit, a feature per kit: crabkit = { features = ["stripe"] }, called as crabkit::stripe
- Build a service's image with crab build [dir], which prints the path of the finished image under target/berm; crab deploy builds it the same way
- Use the auth kit for accounts and sessions, stripe for checkout and subscriptions, supabase for PostgREST and Storage, and neon for Postgres over HTTP
- Crawl the sites a service watches with the crawl kit: robots.txt and sitemaps, pages of the kinds it names, the URLs it lists and the links its crawler finds, recrawled when they change
- Follow feeds and records with the lake kit: RSS 2.0, RSS 1.0 and Atom, threads under items, and a version stored per change of a record
- Install crab with curl -fsSL https://crab.talk/install.sh | sh, and upgrade it in place with crab upgrade, each checked against GitHub's digest of the release asset
- Start a service's database on Cloud from the seed berm.toml names, on its first deploy
- Bound a service's database with [database] size in berm.toml, from 64KiB to 10GiB, 1GiB when unset
- Subscribe from a trial at once with crab billing plan
- Fetch many URLs in one call with berm_lang::http::fetch_all, each with its own timeout and the body bytes it takes
- Declare fetch hosts as *.domain or *, a host reached only by a pattern refused at an address that is not public
- Read random bytes with berm.random, and name the end user the rest of a call's chain runs for with berm.vouch
- Run bermd with a development dashboard behind its dashboard feature, from ghcr.io/crabtalk/crab/berm or the repository's compose file
- Build a product's services on crab with the crab skill in skills/crab
Changed
- Meter and bound each call by CPU time in microseconds in place of wasm instructions: budgets take cpu, and bermd takes --max-cpu-millis
- Have crab new take berm-lang from crab's git tag of the CLI's release
- Write a service's tools as functions marked #[berm_lang::tool(...)] anywhere in its crate, taking public, anonymous, query, cron, stripe and callers(...), in place of #[berm_lang::program] on a module and its #[public], #[cron] and other attributes
- Declare a service's syscalls, secrets, programs, database, readme and buffer in berm.toml in place of #[berm_lang::program(...)]
- Build a service as a cdylib library, [lib] crate-type = ["cdylib", "rlib"], with no bin target and no bin/main.rs
- Refuse an image without the .berm.program section crab build appends, which a bare cargo build leaves out
- Name a feed's Kind (chronological, unordered or records) in lake in place of its chronological flag, and give a listed item's posted as an Option, None when the feed gives no date
- Collapse the crawl kit's migrations into one
- Cut a response at the body bytes its request names in bermd, where it failed past 16MiB
- Take Duration fields of berm_host::Config as "50ms" or "60s"
- Move auth and stripe's services into fixtures, the kits staying libraries
- Move berm_lang::html into the crab-html crate, which reads a page's title, meta, links and text
- Drop the plan's storage rate, a plan's stored bytes the cap its services reserve against
- Name the kit crates crabkit-auth, crabkit-crawl, crabkit-lake, crabkit-neon, crabkit-stripe and crabkit-supabase. A kit's migrations are recorded under its crate's name, so a database that ran them as crab-* runs them again and must be recreated
Fixed
- Read a fetch's host as the URL reaches it, so a URL spelled to look like another host, or an address written as a number, is held to the manifest and to bermd's public-address check
- Bound every read of guest memory, so berm.random and berm.vouch cannot make the host allocate past it
- Refuse a Supabase Storage path with an empty, . or .. segment, which reached another bucket
- Hold the URLs a program lists to the site's robots.txt, match its * patterns in linear time, read it past a byte order mark, and compare its patterns percent-encoded
- Keep a site's URLs when every sitemap fails, and split a round's shared sitemap budget so it cannot be overrun
- Store an undated item and one posted at the watermark in a chronological feed, read nested RSS elements' text apart, and refuse out-of-range dates
- End HTML tags at their real end and match tag names whole in crab-html, and keep an unknown entity as written
- Refuse a database size under 64KiB, which left a database unlimited, and put the old size back when a deploy is refused
- Show CPU in the dashboard, which broke on the renamed meter
- Copy a deploy's seed through a temporary directory removed on every exit