Skip to content

AgentRun v0.1.0-beta.47

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Nov 21:13
· 36 commits to main since this release
27abab8

This release publishes the ACP harness package, so a server installed from npm offers Goose and Cursor once their pinned release is installed (agentrun harness install goose or cursor), and any ACP agent can be brought in as a recipe. AgentRun Server gets a Helm chart published with each release and Coolify and Dokploy templates. An ended run reads its booked cost instead of null, a compacted pi run reads its priced cost, and a run's cost carries its tokens, how its figure was reached and what its paid tool calls cost. A Composio server can act as each run's own account and price its calls, a model gateway is saved from the console, run notices post to Slack, agentrun doctor checks a setup, and a server whose runners all refuse to claim says so instead of queueing runs for good. The store moves on Postgres (migration 0020) and changes a table on SQLite: back up before you upgrade, since going back to beta.46 is that backup.

Upgrading and going back:

  • The store moves: Postgres migration 0020_runners.sql renames the runners' relay table runner_relays to runners in place, keeping its rows. beta.46 refuses a Postgres schema at migration 20 ("set up by a later release than this one"), so going back on Postgres is a backup of the schema taken before the upgrade. SQLite stays at version 13: the server's start drops runner_relays and makes runners, whose rows are heartbeats each runner writes again within 10 s. beta.46 opens such a data folder, but its runner in a process of its own with a public address (the image's default command with AGENTRUN_PUBLIC_URL set, or on Render) does not start, since the table it writes its relay to is gone. Back up the data folder or the Postgres schema before you upgrade; going back to beta.46 is that backup and the beta.46 image or package. From beta.44 or older, beta.45's notes hold first.
  • Start the server before any runner of its own. A runner started on a store this release's server has not opened yet stops with exit 3 on SQLite and Postgres alike: "The store in is behind this runner: start the server first, which upgrades the store in place, then the runner." The image's two processes start in that order. A beta.46 runner with a relay does not start against a store this release's server opened, so upgrade the runners with the server.
  • At each start the server heals, in the background, the ended runs whose stored cost is unknown while their booking is priced: it writes the booked figure once, in place of the unknown, and logs one cost_heal line with counts and no ids. A stored figure is never written over. The healed figures stay if you go back.
  • The server now sends anonymous usage counts, as docs/telemetry.md lists: a random id made for the data folder (kept in its .telemetry.json), the version, the platform and how many runs ended there as a range, never prompts, files, keys, names, paths or addresses. The first start that can send prints a notice and sends nothing until its first heartbeat a day later. AGENTRUN_TELEMETRY=0 or DO_NOT_TRACK=1 where the server starts turns it off; it is already off in CI, under a test runner and under --faux (#3191).
  • A run-list row a beta.46 or earlier server cached is not rebuilt for the new path and workflow fields, so GET /v2/runs and an agent's runs list can show such a run without them until the run changes. GET /v2/runs/{run_id} has both (#3258).

Breaking:

  • A run's read and each runs-list row carry one path and workflow: { name, digest } (the learned workflow that served it, null on any other run). The learning cost read's bill drops baseline and promoted: read summary.before and summary.takeover instead (#3229).
  • Daytona no longer offers a desktop on any connect (none ever started: its sandboxes block the tool user's namespaces): { kind: "desktop", provider: "daytona" } is refused with the reason (#3245).
  • @parcha/agentrun-harness-acp no longer records a binary's first-seen digest itself: fileDigestRecord, verifyFirstSeen, DigestRecord and ACP_BINARY_DRIFT are gone. Pin the binary's sha256 and call verifyPinnedBinary, or install it with agentrun harness install --recipe (#3238).
  • agentrun native takes only the flags it names: --base-url (use --url) and --json are refused, exit 2 (#3225).
  • A run's cost basis reads reported, estimate or unpriced, no longer always estimate; a run with no model request reads unpriced (#3145).
  • A refusal's words changed where a client may have matched them: a secret write refused to the API token ends with a console link (match the sentence it starts with), and a refusal a start-time switch causes names the switch (match the error code, which is unchanged) (#3137, #3132).

Known issues:

  • Under the image's default command, no run can use a paired computer, a paired Linux sandbox included. Start the image with the server's options alone for one.
  • A learned workflow refused for a host setting (a tool gated, denied or missing) still reads serving, while every run it is given hands back to the agent. A workflow this release's DSL refuses is held.
  • For up to about 10 minutes after the release is published, an install can fail while npm shows the new packages one by one. Try again a few minutes later.
  • An API token's GET /v2/settings/calendar is refused 403 in the words for saving a secret. Reading the calendar's settings takes the owner's signed-in session in the console, as saving them does.
  • Fixed: a runner of its own started on a Postgres store before its server had upgraded it said to start on a new, empty data folder. It now says to start the server first, on SQLite and Postgres alike (#3205, #3227).

Changes:

Harnesses

  • @parcha/agentrun-harness-acp, the adapter that runs an Agent Client Protocol agent as a harness, is published with this release, after the kernel and before the server, and the server depends on it. A server installed from npm offers Goose once its owner installs Goose's pinned release (agentrun harness install goose, or agentrun-serve install-harness goose --data <dir>) and restarts it. Goose 1.59.2 is pinned for Linux x86_64 and Apple silicon Macs, and its model requests go through AgentRun's model proxy under the run's cap (#3195).
  • Cursor runs as a harness (cursor) over ACP, on its owner's own Cursor API key: install its pinned release with agentrun harness install cursor, set CURSOR_API_KEY where AgentRun starts, and restart. Each run is billed to your Cursor plan, with no spend cap of AgentRun's: it stops after 20 turns or 100 tool calls (AGENTRUN_CURSOR_MAX_TURNS, AGENTRUN_CURSOR_MAX_TOOL_CALLS). Cursor makes no file edits (every Write is denied in its CLI config), and its own shell commands and calls are asked of a person who can approve them and refused where no one can. The key goes to Cursor's runs alone and the credential Cursor derives from it is never written to disk. A model is cursor/<model-id>, checked against the models Cursor lists for the key. The release is pinned by its archive's, launcher's and every unpacked file's digest, for Linux x86_64 alone: on any other host, a Mac included, Cursor is refused ACP_BINARY_UNPINNED (#3238).
  • Bring your own harness, door 1: a recipe. A JSON file names an ACP agent's registry release (the archive's URL and sha256, the binary inside it), its billing (through AgentRun's model proxy, or the owner's plan with turn and tool-call limits), its literal env, files and session folders in a private HOME. agentrun harness install --recipe FILE installs it, and agentrun harness-check --recipe FILE --model PROVIDER/MODEL runs it for real: its own tools off, its environment an allowlist, the cap or plan limits, the pinned binary at every start, a kill and resume, a denied call, one of AgentRun's tools over MCP and a second runner, then writes a JSON report and prints the resume, tool_gate and billing facts. The loader refuses a recipe that breaks one of five rules by its code (RECIPE_KEY_NOT_OWN, RECIPE_BINARY_UNPINNED, RECIPE_PATH_OUTSIDE, RECIPE_INSTALLS_ITSELF, RECIPE_ENV_NOT_ALLOWLIST), and a binary that differs from the one its install recorded is refused HARNESS_BINARY_MISMATCH. See docs/harness-recipe.md and HARNESS-SCORECARD.md, which rates 25 agents (#3231).
  • The ACP adapter ties Goose's tool title (agentrun: <tool with spaces>) to AgentRun's tool, so the host's gate is asked once per call (#3249). harness-check's gate refuses every call that is the agent's own, so a call of AgentRun's tool the adapter can't tie fails mcp_tool instead of passing, and its tools_off stage passes once the agent's first model request reached the provider; the adapter's and the model proxy's environment variables are one list (#3257, #3251).

Runs and the server

  • A server whose runners all refuse to claim (their run isolation refused, such as an AGENTRUN_RUN_USER the runner does not know, or the image's helper unable to switch users on the platform) now says so instead of queueing runs forever. A start is 503 runner_refused with the runner's words and their fix (set AGENTRUN_RUN_USER=same), a run already queued ends failed with runner_refused, GET /v2/capabilities carries the words in runner_refused, and GET /v2/health answers 200 degraded with its runner check refused, so a platform's health check does not restart the service in a loop. A message to a run and a backend SDK run's start are refused the same way, and it all clears once a runner that can claim starts. Each runner keeps a row in the store saying it is there, judged on the store's clock, so a runner on another host whose clock drifts is no longer refused as not running. A server with --no-runner and no runner yet still queues runs (#3239).
  • A run stopped while it waited on a person reads input_required, idle and a stopped pending from the stop's answer to its end, where it briefly read working again (#3236). A follow-up to a Codex or Claude Code run stopped while a tool call ran is the run's next turn, not a turn after the stopped one is finished (#3228, #3242).
  • A command running when its sandbox is lost is no longer run a second time on the new one: the call fails with code gone, saying the computer was lost and the call not run again, and the next call starts the new sandbox (#3202). A synced folder's replacement sandbox no longer puts the whole folder back on the run's disk; a run whose folder moves to an Archil mount loses the modes and links its synced claim recorded, and a kept synced computer whose keep lapses is deleted without a read-back (#3224).
  • Only a schedule trigger's run is handed its trigger's last result: a Composio, GitHub or webhook trigger's run no longer gets another event's answer (#3255, #3265). A schedule or any other trigger can start a connected agent's runs, uncapped, and refuses a max_cost_usd on one (#3159).
  • GET /v2/capabilities lists the switches only the operator sets at the server's start (AGENTRUN_MCP_ALLOW_LOOPBACK, AGENTRUN_WEBHOOK_PRIVATE_TARGETS, AGENTRUN_BOX_ALLOW, AGENTRUN_BOX_CPUS, _MEMORY, _PIDS) under switches, and a refusal one of them causes names the variable and the restart it takes. agentrun service install --env takes the first two and keeps them (#3132).
  • The server remembers its public address and warns when a start changes it (event=public_url_changed); a browser that opens the console at the old name is sent on to the new one with a 308, and the API's 403 names it (#3134). A server listening on every address prints its setup link at this machine's own IPv4 addresses too (#3152).
  • agentrun import --help, native discover --help, reset-password --help and runner --help print that command's own usage and exit 0, and agentrun-serve <command> --help does too (#3225, #3240).
  • The OpenAPI document requires what the server always sends (an agent's kind, revision, revisions and archived_at, capabilities' harnesses), lists the 403 on the four secret writes that left it out, and has computer_unavailable in RunErrorCode; the TypeScript and Python clients' types follow (Python 0.3.15) (#3217, #3222, #3239).

Costs

  • An ended run whose end stored its cost unknown, because its read of the booking failed or ran past its bound, reads the booked total on GET /v2/runs/{id} and the runs list, no longer cost.usd: null, and the store heals to it (see Upgrading). When the booking has nothing to read either, the cost says why in a new field, unknown: "booking_unread", and the console says "Unknown · cost not read" (#3259).
  • A pi run that compacted its context no longer reads its cost as unknown: the compaction's summary request is answered in the run's record, so the run, the runs list's live spend, /usage and the run page read its priced total. The cap's own reading of a run's journal no longer counts that request twice (#3259).
  • GET /v2/runs/{run_id} gains cost.tools_usd: the part of cost.usd its paid tool calls cost (Monid, priced MCP tools, Parallel, browser time), never added to it (#3234).
  • A run's cost carries its model requests' tokens (input, output, cache_read, cache_write, requests, unmetered) and a basis of reported (every priced request had the provider's or gateway's own figure), estimate or unpriced; GET /stats and GET /usage say the same over the runs they sum, and the console shows the tokens beside the cost (#3145).
  • A pi run's model requests that answer at about the same time are each priced at their own cost, where one subagent's request took every sibling's booking; the run's total was right throughout. The model proxy takes an x-agentrun-conversation header naming a request's conversation, never sent to the provider (#3207).
  • An imported session its source did not price, such as a Codex session, reads at the model catalog's estimate the import kept, wherever a run's cost is shown (#3203).

Learning

  • With no learning.agree_on, a learned workflow's check and its shadow audit compare two answers by their decision through the agent's fast model (model_tiers.fast, else AGENT_MODEL_FAST): one request at the replay's end, under the run's cap and booked as its spend. A reply that is not a verdict, a failed request or a cap refusal counts as a disagreement and says why; the check record keeps grader and each disagreement's differs_on. With no fast model the whole answer is compared as before (#3196).
  • A run a learned workflow served says so where runs are read: workflow on the run and its runs-list row, "Learned workflow" in the console's Kind column and on its run page (#3199); a list row an earlier server cached waits for the run to change (see Upgrading). workflow.repick (found, missed, null) says whether a served run's answer re-picked after a quote miss, a composed workflow's own re-pick included, and its run page shows it (#3273, #3282).
  • The SDK compile prompt calls a learned workflow compiled and unchecked and no longer tells the agent to prefer it; the client's workflow context opens "Compiled workflows for this task (not checked):" (#3264).
  • An existing eve 1.0.0 agent can connect through an additive extension that keeps its model, tools, permissions and durable execution, takes owner corrections and served workflows on later turns, and reuses checked workflows over approved read-only callbacks through a private local MCP bridge. @parcha/agentrun-eve is held from publication while its integration is reviewed, so it is not on npm (#3121, #3254, #3261, #3269).

Tools and integrations

  • A Composio preset server can act as each run's own Composio account: save it with user_id: "per_run", and each run names its account in mcp_auth.<server> as {composio_user_id, connected_account_id?} on its start or message. The binding is sealed into the run's launch and erased at its end, never on the server's row, the run's view or a log line; a run that names none has its calls of that server refused unsent. A binding to the preset's default user agentrun is refused 422 (#3246, #3283).
  • The Composio save takes price_usd, which prices every granted action at that amount a call, held under the run's cap and booked as its tool cost (before, every Composio call read $0), and auth_config_id, which uses an auth config you already have at Composio. A preset re-save keeps the prices set on the server (#3232). An action slug is at most 128 characters, and a guess is offered only for a case-only miss (#3220).
  • A model gateway can be saved in the console: Settings → Model keys → Use a model gateway takes a LiteLLM or other OpenAI-compatible proxy's address, its key, how it takes model ids and each tier's default, and Test lists its models. It takes effect with no restart; AGENT_MODEL_BASE_URL set where AgentRun starts still wins. The API is GET, PUT, DELETE /v2/settings/model-gateway and POST /v2/settings/model-gateway/test, secret writes for the owner's signed-in session (#3144).
  • The server posts run notices to Slack itself: paste an incoming-webhook URL under Settings → Slack (PUT /v2/settings/slack), and each run's end, wait on a person and failed trigger is posted with the agent's name, the status, the run's title and a link, never a run's input, output or a tool's arguments. Delivery is the outbound webhook's outbox, with deliveries, redelivery and a test send (#3143).
  • The outbound webhook's signing secret can be rotated for the same URL: POST /v2/settings/webhook/secret, and Rotate secret in Settings → Webhook. The old secret is retired at once (#3136).
  • A secret write refused to the API token says where to finish it: the same 403 forbidden ends "Finish it in the console: ", with the link in error.details.console_url, the console form filled with everything but the secret. The link never holds a token, a header or a key (#3137).
  • The README and the setup skill say how to run claude-agent through an Anthropic-compatible gateway with ANTHROPIC_BASE_URL (#3141), and docs/connect-tools.md says how to give an agent a service's tools without a token in its box (#3142).

Computers

  • A run on a computer that mounts its folder attaches again when the run is a process of its own, as it is under every runner since beta.46: such a run was refused about 25 s in, saying the computer did not hold the run's folder (#3288).
  • A smol cloud provider: providerFor("smolcloud", { apiKey }) makes each computer a smol cloud microVM booted from an OCI image, the box image by default, as a sandbox; Settings takes SMOLCLOUD_API_KEY. No mount and no desktop yet, and a synced folder's refusal says why (#3163, #3218).
  • GET /v2/computers lists each kind's settings, which the console's Computer step reads (#3250). A dial desktop whose display is down fails connect and watch as drift, where it answered with no screen (#3250).
  • A Linux desktop waits 100 ms after a key action before the batch's next one, so an app acts on the shortcut first (#3215). A paired Linux box stays held until each remove and KILL its create sent has answered (#3214).
  • The Fly, Cloudflare and Modal conformance modules take a kind (COMPUTER_FLY_KIND, COMPUTER_CLOUDFLARE_KIND, COMPUTER_MODAL_KIND), dial conformance gives a desktop run the computerd fixture, and MACHINE_TOKEN_SOURCE gives a machine token's shape for a scrub (#3216, #3263).

Console

  • The agent page says before Run what would refuse a run: GET /v2/agents/{agent_id} carries run_blockers, and each shows as a warning (#3156). An agent's description is editable through PUT /v2/agents/{agent_id}/definition and Edit agent; its name stays fixed (#3219).
  • A trigger whose agent asks before a tool call says a run will wait for you, and Run unattended fixes it in one click (#3131).
  • A finished run's Screen tab plays its own recording (#3204), and an open live viewer hears a Take over and a Hand back made from the run page's header (#3200).
  • The Computer step keeps image, persistent, GPU, surface, exec and record across a switch where the new pick takes them, and names each one it drops (#3213). A workflow run's Graph names a map item by its label's head when the label is text (#3223). The Dev tools tab no longer scrolls sideways beside an open panel, and its cache finding compares a request only with its own conversation's (#3206).

Install, deploy and the command

  • The first visit opens setup with its code filled in: --open, the installer and agentrun service install|start|restart open the setup page while the server has no owner, agentrun open does it at any time, and Settings → Webhook copies its secret in one press (#3119).
  • agentrun doctor checks how this AgentRun is set up and prints each problem with what to change: Node.js, the server and its version, the service, remote access, the console's password, models, harnesses, start-time switches and Docker. It is read-only and prints no token, key or code; --json for agents (#3135).
  • --public-url https://<name>/ on npx @parcha/agentrun and agentrun service install serves AgentRun at an address of your own behind a proxy you run; started on loopback with a signed-in Tailscale, the launcher says once how to serve it on the tailnet (#3140).
  • agentrun stop names a server running in the foreground and says to press Ctrl-C in its terminal (#3272).
  • A Helm chart, published with each release: helm install agentrun oci://ghcr.io/parcha-ai/charts/agentrun --devel runs the release's image as one pod on a 10 GiB volume, its token and seal key made as a Secret kept on upgrade and uninstall, with an optional Ingress whose address the server gets as its public address. The chart's version is the release's (#3252, #3253, #3262, #3267).
  • Coolify and Dokploy templates run the release's image as one service with its data on a volume (deploy/coolify/agentrun.yaml, deploy/dokploy/agentrun/) (#3268).
  • install.sh --vps pulls its image on every run, so a moving tag runs the registry's image (#3197), and AGENTRUN_TELEMETRY=0 on a run again replaces the line its .env has (#3230, #3244).
  • import shows progress, stops cleanly on Ctrl+C after the session being saved and says what it took, and says Stopped only when sessions were left (#3151, #3278).

Docs and examples

  • The README and docs/getting-started.md lead their local setup with the prompt to paste into a coding agent, the same text the website's copy button gives; the three manual steps follow it (#3286).
  • An example of Browser Use through Monid in a run (#3147).
  • The server README records Particle-1's estimate matching Particle's own bill (#3233), a Render deploy walked from creation to first run in 85 s (#3226), and how to count runs with an unrecorded link before unsetting DAYTONA_ARCHIL_EGRESS (#3241); docs/AUTOQA.md's expected results match the product (#3235).

CI and tests

  • Every way in is walked on main each day (npx @parcha/agentrun as published, the launcher, Docker Compose, Railway, Helm, Coolify, Dokploy, and Railway's and Render's own deploys when their secrets exist), and a red run opens an issue (#3271, #3276). The walks share one first run, the setup prompt is walked to a finished first run against 5 minutes, the Coolify and Dokploy walk keeps the first run's page as a screenshot, and the Render walk rides out the edge's 502 during a deploy (#3274, #3221, #3248, #3275, #3284). The Render walk runs the Railway walk's steps, and a walk Actions cancels finishes deleting what it made before it exits (#3287).
  • Internal simplifications with no change in behaviour, and test fixes (#3212, #3243, #3256, #3260, #3266, #3270, #3277).

Install

  • Server: npx @parcha/agentrun-serve@0.1.0-beta.47; every @parcha/agentrun-* package of this release is 0.1.0-beta.47 on npm
  • Launcher and CLI: npx @parcha/agentrun@0.4.15
  • Python client: pip install parcha-agentrun==0.3.15
  • Server image: docker pull ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.47 (linux/amd64, linux/arm64)
  • Helm chart: helm install agentrun oci://ghcr.io/parcha-ai/charts/agentrun --version 0.1.0-beta.47