Summer

Check issue permission and repository visibility, or read one original GitHub issue by number. Restricted to company-selected repositories; no credentials or writes. Issue closure does not resolve the customer case.

Read the company GitHub installation, selected repositories/folders and cached commit status. Makes no GitHub request and returns no credentials.

Get the Summer browser URL for a company administrator to install and authorize the Summer GitHub App. Code access is read-only; issue handoffs use separate GitHub Issues permission and exact case approval. Does not grant access or connect an installation by itself.

List repositories available to this company GitHub installation. Provider metadata read; no file content.

Browse repository folders at a branch or commit before selecting access.

Choose repository branches/commits and allowed folders. A dot selects the whole repository. Replaces selections with an expected version and invalidates old cached access. Does not sync code.

Refresh the selected repository code cache once from GitHub. Reuses the cache for an unchanged commit; sandbox startup never pulls GitHub.

List cached source paths, or read one cached file with path. Only currently selected company folders are available. No live GitHub call or credentials.

Disconnect this company/environment GitHub source access and remove its cached archives.

Read one saved recurring-issue report and its source snapshot, including reports older than the latest list. Requires current operational company access.

Read the latest private recurring-issue reports, original case references, source coverage, generation status and human review. Generated themes are suggestions, not verified causes.

Generate a private recurring-issue report from the admitted case-summary cohort. One immutable request ID prevents duplicate inference. No customer, knowledge or publication effect.

Record a human disposition on a saved recurring-issue theme at its expected revision. Retains the original analysis and source references; does not change cases or company knowledge.

Read built-in tasks and company/environment-owned workflow templates. Selecting one never runs work.

Read the original revisions and authenticated edits for a saved workflow.

Save, retire or restore a task template with its expected revision. Changes no case or effect authority.

Prepare editable operator guidance from an exact active workflow revision and feedback. Starts no investigation and executes no action.

Read business hours, holidays, response commitments and scheduled policy revisions.

Save business hours and response commitments with the expected revision and an effective date.

Read company saved replies, optionally including retired replies.

Create, edit, retire or restore a saved reply with its expected revision and reason. Sends nothing. API reference: https://api.gosummer.ai/v1/openapi.json

Read original revisions of a saved reply.

Read this person's saved inbox views and shared company views.

Save a personal or shared inbox view. Known case filters are normalized by the same service as the UI.

Remove an owned inbox view, retaining the database's ownership checks.

Read quality scenarios, assessments and current outcomes.

Read one scenario and its saved assessments.

Create, revise or retire a quality scenario, with expected revision and reason.

Record an evidence-backed assessment or correction. Does not approve a plan or customer action.

Read evidence-backed corrections to previous quality assessments.

Read privacy export/deletion requests, status and original event history. Does not export or delete data.

File an export or deletion request for human review. Cannot approve, perform or attest completion of the work.

Read eligible teammates and automatic routing for new issues, plan reviews and human tasks.

Save automatic routing for new issues, plan reviews and human tasks with the expected version. Does not grant roles or approval authority.

Read the rules Summer applies before any agent runs, in plain words, with this week's decisions and how often each rule matched. Routing costs no model call.

Save the router rules with the expected version and a reason. Keeps the original revision and updates the Routing rules knowledge document. Run router_rules.test first. Needs a company admin.

Check proposed router rules without saving: lists rule problems, replays recorded ticket activity under the saved and the proposed rules, and routes optional sample facts. Changes nothing.

Read every saved revision of the router rules, with who changed them and why.

Read what the router decided and why: totals by outcome, day and rule for the last 7, 30 or 90 days, and pages of decisions with a short plain reason each and the case. Changes nothing.

Read private engineering issue associations, optionally checking original GitHub state. Does not close cases or send replies.

Read bounded original case-summary text matches, with source observation and matching terms. Does not establish customer identity, link cases, call a model or start work.

Read connected Slack posting channels, or inspect the current internal channel name/privacy, posting revision and authenticated sender. Enables no grant and posts nothing.

Read original Slack replies for an exact recorded case discussion. A root supplied for an unknown send must match the saved original sender, approved text and operation metadata before its source can be retained. Never posts or approves work.

Read case-linked Slack discussion receipts and original thread links. Unknown posting stays unknown; no message is sent.

Read a case's issue owner, plan reviewer and eligible teammates.

Assign an issue owner or plan reviewer with the previous owner as a conflict check. Does not grant roles or approve a plan.

Read all-time and month-to-date credits, pending reservations and monthly usage in the company timezone.

Read paginated original issue/message references behind service counts and timing samples, with date/channel/hour filters in the company timezone. No messages are sent or marked read.

Read response and resolution metrics, 90-day daily issue/message and timing trends, and weekday/hour incoming-message heatmaps in the company timezone. Missing timings stay unmeasured.

Read how long approved agent runs take: approval to start, each startup and work phase, and each step by action and category, with medians, 90th percentiles and recent runs. Phases that were not recorded stay unmeasured. Starts no work.

Read cancellation requests and how each ended: paused, renewal moved later, kept, cancelled, or a pause offered with no answer yet. The investigating agent marks each request; the approved plan option sets the outcome. revenue:true adds each saved plan's renewal amount and the paid renewals collected since the save (before refunds), read from live billing, plus gross totals. Changes nothing.

Read delivery attempts, failures and recorded results. Does not retry a delivery.

Act on several rows from reports.delivery at once: request a new plan, approve each case's current saved plan and run it, or start an approved plan that is waiting. Returns done, skipped or failed with a reason for each row. Rows with no confirmed result are never run again. Approval binds to each case's exact saved plan and prepared changes; plans that changed or need a choice between options are

Read outstanding customer promises and unfinished work.

Read work that Summer agents or blocked approvals asked a person to do, with owners, due times, recorded outcomes and change history for one selected task.

Complete, dismiss, reopen or assign a human task with its expected version. Records the person and outcome. Does not approve a plan or change customer records.

Read the company's sender exclusions and original edit history. Changes retain human approval.

Search and filter the complete Messenger inbox, including visitor conversations without Ops cases.

Read original Messenger messages with saved human sender_display_name, contact form details, read receipts, attachments and timestamped page_context (current page and UI controls) without marking the chat read.

Save a confirmed administrative shipping-address correction for the selected order on a live outbound pharmacy call. Requires an Ops admin and customer-effect access. Does not update carriers or clinical fields; cross-state requests go to clinical coverage review.

Read Summer answering lines, Quo number and keypad mappings, recent calls, source digits, case links, received-event times and settings history. Does not mark calls read or change Quo routing.

Save Summer answering and call-type mappings for an existing line with its expected version. Requires current company and Ops admin access. Does not publish a Quo menu, change ringing users or forward a number. New Quo numbers also need webhook subscription in Quo.

Read demo pharmacy-call voice service, voices, vocabulary and change history. Starts no call.

Save the demo pharmacy-call voice service, voice and vocabulary with the expected revision. provider is elevenlabs (default) or openai_live (OpenAI GPT-Live); omitted provider or openai_voice keeps the saved value. Starts no call.

Read Intercom migration previews, saved progress, counts, errors and Summer inbox destinations. Does not start an import.

Check the saved Intercom key against a workspace and preview contact/conversation counts. Save the selected Summer inbox and explicit account-matching choice. Import nothing; send nothing.

Start, pause, resume or rescan an Intercom migration with its current version. Optional incremental sync reconciles the same original IDs; no replies, notifications or agent runs. Names/email never verify an account.

Search imported Summer profiles and their original conversation references. Contact details do not establish account ownership.

List suggested female and gender-neutral teammate names whose gosummer.ai address is still free. Each has a suggested photo_url to show; create with use_suggested_photo: true to add that name, address and photo together. Any typed first and last name also works.

List the company's teammates with their address, photo, status, unread mail count and Slack connection.

Add a teammate. A managed teammate gets first.last@gosummer.ai, unique across Summer. A company-managed teammate uses the company's own address. Invites no one, joins no Slack workspace and sends nothing.

Change a teammate's shown name, photo or status, with the expected revision and a reason. Retiring disconnects Slack and keeps the address reserved.

Disconnect a teammate's Slack account from Summer and delete its saved token. The Slack account itself stays in the workspace. Connecting needs the Slack sign-in in Summer settings.

Live status of each step to get a teammate posting approved replies in vendor Slack channels: Slack connected, teammate added, Slack invitation received, teammate connected, chosen as sender, in each vendor channel, and agent replies allowed there. Unreadable sources are reported as unknown.

Join a public Slack channel as the connected teammate. Sends no message. A private or shared channel needs someone in it to /invite the teammate in Slack.

Make a one-time link that connects the teammate's Slack account. Open it in the browser signed in to Slack as the teammate and select Allow; no Summer sign-in is needed there. It works once, for 15 minutes, and the connection is recorded as yours. Treat the link like a password. Needs a company owner or administrator.

List the people who can DM this AI employee in Slack, with the list revision, plus `people`: names and member IDs Summer has seen in the selected channels, to choose from. Each DM from one of them opens an Ops case; a person approves the exact reply before the AI employee posts it. Needs a company owner or administrator.

Add or remove one person who can DM this AI employee, with the expected list revision and a reason. Summer reads only DMs from people on the list. Removing someone stops Summer reading and posting in that DM. Sends nothing.

List mail received by a Summer-managed teammate, newest first, with sender, subject and a short preview. Company-managed teammates have no mailbox in Summer.

Read one message in a Summer-managed teammate's mailbox, including its text. Marks it read for the team.

Check live Summer MCP/API versions, old client warnings and the seven-day discovery refresh interval. Optional cached client versions and the last successful check timestamp distinguish known and unknown status.

Show the signed-in person, Summer access status, profile photo path and this connection's grant.

Remove your own profile photo. Replies you send afterward show your initials; earlier replies keep the photo they were sent with.

List companies the person can access and whether this connection includes each one.

Read the person's Summer access request. Approval is a separate Summer team decision.

Submit or update a Summer access request. This cannot approve access.

List this person's agent connections, grants, last use and expiry.

Revoke one of this person's agent connections immediately.

Read the company's editable guidebook, draft articles, assets and publication state. Customer guides share the existing company knowledge source.

Create or edit guidebook details at the expected version. Saves a private draft; does not publish.

Read an editable guide article and its current knowledge revision.

Create or edit a Markdown customer article with source references and managed asset IDs. Saves a draft in the shared knowledge library; later edits do not change a published release.

Read the exact current draft, article revisions and authenticated preview URL before publishing. No public access token is issued.

Read the company's reviewed guide-authoring brief, draft article references and publication state. This read starts no model run and changes no customer data.

Publish a complete release of the company's customer help center. The release replaces every published article: list each article that should stay public (guides.preview publication.articles) and name any published article you mean to remove in unpublish. Requires explicit publication authority; does not approve customer business actions.

Remove the company's guide and release media from Summer's public serving path at the expected version. Downloaded copies cannot be recalled.

Read the company's guide edits, upload and publication history.

Read the optional custom guide domain, DNS instructions and verified HTTPS status. Summer-hosted guide URLs need no customer DNS setup.

Save or remove a custom guide hostname at the expected version. Saving a name does not verify ownership or provision HTTPS.

Check the saved domain's DNS ownership, attach it through Summer's hosting connection and verify HTTPS. Returns pending until each check succeeds.

List the company's private guide media, metadata and upload status. Uploading does not make a file public.

Read one guide asset, checksum and resumable chunk receipts. No secret upload capability is returned.

Start or recover a private resumable Markdown, image, video, PDF or caption upload from metadata. File bytes are transferred separately; this hosted endpoint has no local filesystem access.

Verify the original size, content signature and SHA-256 against saved private bytes, then mark a guide upload ready. A ready file remains private until published in a guide.

Edit guide media captions, alt text, provenance, transcript or poster at the expected version, or retire unused media. Published releases keep their pinned metadata.

What this connection can do here, with a reason for everything unavailable.

Saved, verified, pending, unavailable and enabled state for each setup item.

Read the same operational capability checklist, results, discovery paths and review history shown in Summer. Every usable API, browser or knowledge path can count.

Read one capability, its concrete test, evidence and latest 20 review records.

Edit a capability expectation or add a business-specific custom_ capability with a reason. Resets its result and retains history. Does not create a provider tool or grant execution access.

Record the setup agent's capability assessment with original evidence and the path Summer can discover and use. A pass needs evidence and a discovery path. Reports are labeled as agent-reported, never human approval or independently verified execution.

Request a read-only Summer capability test on an authorized sample case. Also needs investigations.request and current runtime access. Does not approve or execute customer changes.

Read company-wide settings, including the email subject customers see on new support emails. Changes affect every environment.

Update company-wide settings with the expected version. Affects every environment.

List provider integrations, their categories, required fields, alternative credential sets and saved metadata. Filter by search, category or connection. Connected means all required credentials and one complete alternative set are available, not a live provider check. read_only integrations are read with the company's own key and never written through. Never returns values.

Save a custom integration's description, setup instructions and credential field names. This saves metadata only, without credential values or a live connection check.

Prepare an integration-specific MCP setup prompt and Codex or Claude Code app links. Includes the selected company, environment and setup brief, never credential values. Does not launch an app or change configuration.

List saved credential names, sources and versions. Never returns values.

Save one selected credential value or Doppler reference. Replacing an active credential needs credentials.replace. Returns metadata only. Returns a local CLI command when a secret is involved.

List current company knowledge documents. Knowledge is company-wide, not per environment. Pass tag to list documents with that tag, such as an integration name (slack, beluga, mdi), which is how Summer shows knowledge related to an integration.

Upload a new company knowledge document with original revision history. The shared library is used by all roles and environments. Existing document IDs are refused. API reference: https://api.gosummer.ai/v1/openapi.json

Read company knowledge changes from people, delegated API/MCP connections and Summer agent plans, with before/after revisions and original provenance.

Read one company knowledge document and its current revision.

Read the edit history of a company knowledge document.

Edit an existing company knowledge document with the expected revision and a reason. Affects every environment and later runs.

List Messenger installations in this environment. Signing keys are never returned.

Create an installation with a client-chosen ID. The one-time signing key goes to the trusted local helper, never to model-visible output. Returns a local CLI command when a secret is involved.

Update installation settings with the expected version. Keeps the signing key.

Issue a new signing key. Existing signed sessions end. The key goes to the trusted local helper only. Returns a local CLI command when a secret is involved.

Read assistant settings per installation and whether the managed runtime is enabled.

Read consented website leads for this company and environment. Includes volunteered contact details and original submission references.

Update assistant instructions and escalation permission with the expected revision. Company facts live in shared knowledge. Cannot raise the spending limit.

Read voice chat settings per installation, the voice services and their voices, published help article count and the voice assistant's update status.

Update voice chat settings with the expected revision. Summer updates the voice assistant from these settings and the company's published help articles.

Read support email addresses: the private forwarding address, the one DNS record to add, whether mail has arrived, sending status, Gmail forwarding codes and emails that need a look.

Add a support email address. Summer creates its private forwarding address and signing key and returns the DNS record to add. The address starts off; nothing is sent.

Change the sender name or turn a support email address on or off, with the expected version. Turning on needs one received email. Turning off keeps every email but stops new conversations and replies.

Check forwarding, the DNS record and sending status for one support email address now. Records the latest sending status; sends nothing.

Read the company execution policy and its publishing history, plus autonomy for website chat, texting and email: whether the assistant gives fast replies and how sure it must be to answer alone, and the auto-run plan settings, including who turned auto-run on and website chat installation readiness (chat_assistant).

Change one channel's autonomy with the expected version. With fast reply on, the assistant handles that channel's messages and starts a plan only when it escalates. A lower threshold gives it more autonomy.

Preview the complete replacement execution policy. Conversation reply grants remain separate; explicit communication review exceptions apply to replies too. Returns an exact preview hash; changes nothing.

Publish the complete previewed company execution policy under the authenticated administrator's authority. Requires a person's delegated messaging.activate grant, expected version and preview_hash. Replaces legacy inbound-channel plan permission. Does not run old held plans.

Count latest pending plans per open ticket across the connected operations company, including partner/system/operator origins. Other tenants cover only their Messenger tickets. Returns aggregate scores, holds, coverage and an exact refresh preview; starts nothing. auto_run:true previews only plans that could run on their own under current auto-run settings.

Queue fresh research for the exact pending-plan preview, up to 100 tickets. Each fresh plan follows the ordinary auto-run eligibility rule. auto_run:true (matching the preview) picks only plans that could run on their own. Never replays old effects. Requires a new preview if work changed; successful repeats return the original receipt.

Legacy channel policy only; unavailable after an explicit company-policy replacement. Preview existing Messenger work under saved autonomy settings: counts, a plan confidence histogram and a preview_hash. Returns no messages or customer records. Starts nothing. Up to 100 items per apply batch; totals cover all matching work.

Apply an exact backfill preview using saved settings. Queue unanswered conversations and fresh investigations of qualifying pending plans, up to 100 items. Enabled assistants may reply; new plans follow saved auto-run settings. Skips paused, active, failed or uncertain work. Does not replay old executions. Requires the exact preview_hash; changed settings or work require a new preview. Repeating a

Read texting lines: each line's persona name, how it hands work to the team, reply hours and whether texts join conversations, plus recent send problems and the lines this deployment can text from. Whether the assistant replies is a channel autonomy setting.

Add a texting line. Received texts on it can join Summer conversations. The line starts off; nothing is sent. Fast replies follow the texting channel's autonomy setting.

Change a texting line with the expected version: persona name, handoff wording, reply hours, other Quo lines it takes texts in from and whether texts join conversations. Turn fast replies on or off in messenger.autonomy.update.

Read the support name and avatar shown on Summer's approved Slack replies, whether the Slack install allows them, and the change history.

Set the support name and avatar on Summer's approved Slack replies, with the expected revision and a reason. Null values use the plain Summer Ellis identity. Sends no message.

Allow or stop approved agent replies in one vendor Slack channel, with the expected version from teammates.setup.status and a reason. Each reply still needs a person to approve its exact text. Sends no message.

Choose who posts approved vendor Slack replies: the Summer Ellis app, a Summer Teammate (AI employee) from its own Slack account, or the person who approves each reply, from their own connected Slack account. Uses the revision from slack.identity.read and a reason. Sends no message.

Read who Summer reads Slack as: the connection itself (the Summer app or the installer's personal read access), an AI employee, or an admin's own Slack account, whether that account is connected, the revision and the change history.

Choose who Summer reads Slack as, with the revision from slack.reader.read and a reason. 'me' uses the caller's own connected Slack account; nobody can choose another person's account. Summer reads only the selected channels that account can see. Sends no message.

Read whether the caller's own Slack account is connected to Summer in this company, and the settings link to connect it. Connecting uses Slack sign-in in a browser.

Disconnect the caller's own Slack account from Summer in this company and delete its saved token. Replies set to post as the approver are then blocked for the caller, and Summer stops reading as the caller.

Read which Slack account Summer links to the caller in this company, how it was linked, the Slack people the caller can choose from, and whether Summer sends the caller Slack messages about assigned human tasks.

Choose the caller's own Slack account in this company, say they have none, or turn Slack messages about their human tasks on or off. Changes only the caller's settings.

Admin: match every company member to their Slack account again, by email and name. Accounts people signed in with or chose stay as they are. Runs within a few minutes; sends no message.

List the Slack channels the company selected for Summer, with read and post access and each channel's vendor and purpose. Agents see the vendor and purpose with every vendor Slack read.

Set which vendor a selected Slack channel is for and what it is used for, with a reason. Null clears a field. Changes guidance for agents and people; sends no message and changes no channel access.

List company members and pending invitations. Role changes and invitations stay in Summer.

Pause the assistant in one conversation so a person can take over, or ask it to resume. Pausing stops automatic replies immediately.

Run supported harmless checks. Unsupported checks are reported as unsupported, never as passed.

List company-owned cases or conversations with emotion filters (sentiment: happy, neutral, upset, angry, sad, mixed or unknown), a customer filter (customer: paid for a captured or authorized live payment, no_purchase for a matched contact who has not paid, unmatched for an unknown sender), ordering (sort, including customer_message for the latest customer message first), page size, coverage and a

Find cases by customer name, email, phone, date of birth, support code, order or tracking number, vendor or Stripe ID, or message words, with status and date filters. Returns one current workspace per verified customer (explicit splits stay separate), with the actual ticket code, all issue references, plan and run progress, orders and message excerpts. Message case_id opens the workspace; source_c

Read one case or conversation with its original messages and related records. Operational records include workspace with the current customer ticket and all owned issue references, plus the shared saved timeline, including original inbound phone transcripts, caller-supplied pharmacy intake and private tool receipts with attempts and results. Generated speech is not proof the caller heard it; saved

Get a short-lived download for an attachment that belongs to this case.

Read the case customer's subscriptions and upcoming shipments, the same account facts used by the chat agent and Details panel: renewal and next-charge dates, recurring price/currency, pauses and scheduled ends, refill due dates and prepaid fills, fulfillment and tracking. Planned dates are not confirmed ship dates. Read coverage distinguishes unknown from absent; no customer changes are made.

Read the case customer's profile: name, date of birth, contact details, default shipping and billing addresses, each order's addresses and earlier addresses. A saved address does not prove where an existing order ships.

Read shared investigation runs, saved plans, approval state and execution receipts across the verified customer’s cases, including historical issues. Each run retains its original support_case_id; use that ID for write operations and original run evidence.

List investigation and execution runs across cases, newest first, with status, plan approval and date filters.

List published plans, newest first: who approved each (a person or an agent), whether it ran and was checked, the agent's self-review, its own call on running the plan without approval, and auto_run: what the channel's auto-run setting did or would do with it (outcome ran/would_run/would_wait, the deciding reason, channel, on/off, confidence bar and any saved auto-run decision), and runs_as: code

Ask Summer to investigate the case, or revise a saved plan with feedback. The new plan follows the company's auto-run rules; to make every plan wait for a person, turn auto-run off in the execution policy.

Read the case chat with Summer: each question people asked about the case, oldest first, with its status and Summer's saved answer (the latest 10). An answer is Summer's explanation of its own work, not new evidence, approval or a customer outcome.

Ask Summer a question about the case, such as why it chose a plan or what it found. Summer reopens its last run on the case, with what it already read, and answers in the case chat. Read-only: it publishes no plan and makes no changes. One question at a time per case. Read cases.chat.list for the answer, which usually takes a minute or two. To change the plan, use investigations.request.

Stop an exact investigation owned by this case, preserving its evidence and history. Cannot stop execution runs. A stop request is not a finished cancellation: read investigations.list until the run is cancelled or completed, then use investigations.request with current guidance to rerun.

Edit wording in the current saved plan at its expected revision, including proposed company knowledge text. Preserves targets, amounts, feedback sources and selected option. Invalidates prepared approval; does not apply knowledge or customer changes.

Prepare the exact current plan and return its execution ID, approval hash and reviewed work. Does not approve or execute. A delegated actions.approve grant can approve authorized work through API, CLI or MCP.

Approve the exact prepared execution under the authenticated operator's delegated authority. Use only for work the operator authorized, read the prepared review first, and pass its exact hash. Records the authorization and queues ordinary execution; approval does not prove a customer outcome. No caller-supplied approver identity or identity-verification attestation.

Try a saved plan approval again when it is waiting to start (its id is in investigations.list planApprovals). Returns the approval's current state and the reason when it still cannot start. start_anyway ends the case's stopped runs that have no checked result, recording this person, so the approved plan starts; an earlier run may already have made its changes. A run still working keeps blocking.

Read configuration and activity events after a cursor. Pagination uses next_cursor.

List the company's service accounts.

Create, update, rotate or disable a service account. A new secret goes to the trusted local helper only. Returns a local CLI command when a secret is involved.

List webhook endpoints and delivery state.

Create, update, rotate or disable a signed webhook endpoint. A new signing secret goes to the trusted local helper only. Returns a local CLI command when a secret is involved.

Queue past events after a cursor for delivery again to one endpoint. Receivers deduplicate by event ID.