feat(agent-applications): manage MCP connections in the agent builder by benjackwhite · Pull Request #2929 · PostHog/code

benjackwhite · 2026-06-25T10:48:42Z

Builder UI for the agent-platform MCP layer — shared connections, the explicit agent/principal credential model, and per-agent tool permissions. Pairs with the runtime in PostHog/posthog#66007. An owner connects a remote MCP server once via PostHog Code's native mcp_store flow, attaches it to an agent, and every asker uses that one shared credential (or, for principal, each asker connects as themselves).

What you can do

Add an MCP to an agent from a connected server, or connect a new one via a popout modal (AddCustomServerDialog, shared with the agent-builder connect_mcp punch-out — native OAuth/DCR or api-key flow).
Choose the credential model for an MCP entry via an explicit dropdown — Agent-level (a shared connection) or Principal-level (a per-asker linked identity, surfaced as a fingerprint preview card with a jump to the identity detail). Writes spec.mcps[].kind and clears the opposite credential.
Edit per-tool permissions against the connection's live tool catalog (ToolPermissionList): a connection-wide default (allow / approve / deny) plus per-tool overrides, mapped to the spec's default_tool_approval / level.
Remove an MCP entry (drops the mcps[] ref; the shared installation is untouched).
All edits persist via draft-branch-then-PATCH (a fresh draft is branched off a non-draft revision, then selected).

Changes

mcp-server-manager/ (new shared module): useMcpConnect (installations query + connect/reauthorize over the host OAuth callback) and the moved AddCustomServerForm. The standalone MCP-servers scene now consumes the shared mcpKeys/createOAuthCallback (no behaviour change — core mcp-servers tests still green).
agent-applications: editable McpsOverview (add-from-connection + connect-new) and McpBody — a compact header (auth-mode dropdown + remove-server), then the mode-specific config (agent → connection picker; principal → identity selector + preview card) and the ToolPermissionList catalog.
Legacy integrations removed (deliberate). This PR drops the integrations config section (IntegrationsOverview/IntegrationBody, the cfg:integrations node, SECTION_INFO entry) and the AgentSpec.integrations?: string[] field in packages/shared — a total removal, not a soft UI-only regression. Nothing reads the field and the agent-platform schema already dropped it; team-OAuth use cases are served by mcps[].connection / identity_providers[]. (Closes @dmarticus's review note.)
Editing foundation, built on main: api-client.updateAgentRevisionSpec + useApplyAgentSpec (+ test). Intentionally ported compatible with the in-flight agent-model-policy-ui branch so the two reconcile cleanly on merge.
Drive-by: cast spec.entrypoint in ModelBody — fixes a pre-existing @posthog/ui typecheck error (AgentSpec's [key: string]: unknown index signature), unrelated to this change but blocking a clean typecheck.

Validation

@posthog/api-client + @posthog/ui typecheck clean (full pnpm typecheck green via pre-commit); biome clean; core mcp-servers suite (41) + new useApplyAgentSpec suite (5) pass.

Notes / follow-ups

Base branch: targets main (the ben/agent-supported-client-tools base PR was merged; this branch absorbed master and reconciled the supported_client_tools producer, keeping the connect_mcp client tool).
Overlaps the agent-model-policy-ui editing work (ported compatibly).
Component-level tests for the new MCP panels are a follow-up (hooks/logic are covered).

⚠️ Outstanding / known gaps before this is production-good

UI add/edit/remove was driven against the live app and works. Remaining concerns:

Tool catalog needs a verified connection. ToolPermissionList now renders the connection's live tool catalog with per-tool allow/approve/deny (deny = hidden), so choosing which tools to expose is solved for connection-backed MCPs — but the catalog only populates once the connection is verified/has listed its tools; until then it's empty. Non-connection (principal / BYO) entries still fall back to the legacy already-selected list.
needs_reauth not surfaced. The connectionMissing hint only covers "connection isn't in this project". There's no UI for an expired/revoked shared connection (the runtime returns needs_reauth; see feat(agent-platform): agent-level MCP connections, per-agent tool permissions, and capability-based client tools posthog#66007 gap chore: set up biome and ci #3–4) — the owner gets no prompt to reconnect.
Stale-spec agents can't be edited. Example agents seeded with an older models (plural) spec shape fail the write schema ('model' is a required property), so any MCP edit 400s. Re-seed to refresh. Pre-existing data issue, not introduced here, but users will hit it on old agents.
Runtime is unproven end-to-end — see the matching section in feat(agent-platform): agent-level MCP connections, per-agent tool permissions, and capability-based client tools posthog#66007. The UI can attach a connection, but it hasn't been observed driving a real agent→MCP tool call.

Agent context

Authored from a Claude Code session (Opus 4.8).

🤖 Generated with Claude Code

Lets an owner attach a shared MCP connection to an agent and edit its per-tool permissions, reusing PostHog Code's native mcp_store connect flow. Pairs with the runtime in posthog#66007 (spec.mcps[].connection). - mcp-server-manager: new shared module — useMcpConnect (installations query + connect/reauthorize over the host OAuth callback) + the moved AddCustomServerForm. mcp-servers' useMcpServers now consumes the shared mcpKeys/createOAuthCallback. - agent-applications: editable McpsOverview (add an MCP from a connection, or connect a new server inline) + McpBody (connection picker, per-tool requires_approval toggles), persisted via useApplyAgentSpec. - foundation (built on main): api-client updateAgentRevisionSpec + useApplyAgentSpec (+ test) — draft-branch-on-save then PATCH. Compatible with the in-flight agent-model-policy-ui branch so they merge cleanly. - drive-by: cast spec.entrypoint in ModelBody to fix a pre-existing @posthog/ui typecheck error (AgentSpec index signature) unrelated to this change. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

github-actions · 2026-06-25T10:49:12Z

React Doctor found 3 issues in 3 files · 3 warnings.

3 warnings

src/features/agent-applications/agent-builder/AgentBuilderDock.tsx

⚠️ src/features/agent-applications/agent-builder/AgentBuilderDock.tsx:104 Large component is hard to read and change no-giant-component

src/features/agent-applications/components/AgentConfigurationPane.tsx

⚠️ src/features/agent-applications/components/AgentConfigurationPane.tsx:1479 Large component is hard to read and change no-giant-component

src/features/mcp-server-manager/AddCustomServerForm.tsx

⚠️ src/features/mcp-server-manager/AddCustomServerForm.tsx:48 Many related useState calls prefer-useReducer

_{Reviewed by React Doctor for commit 5450500.}

greptile-apps · 2026-06-25T10:53:08Z

_{Reviews (1): Last reviewed commit: "feat(agent-applications): manage MCP con..." | Re-trigger Greptile}

- useApplyAgentSpec: invalidate the revision detail via a new agentApplicationsKeys.revisionPrefix() instead of a hardcoded key array, so the centralised key registry stays the single source of truth. - McpBody: destructure the ctx fields the apply() callback reads (revisionId, revisionState, onSelectRevision) into its dep array; ctx is a fresh object literal each render, so depending on it defeated the useCallback memoization. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…dock The Edit-with-AI meta-agent's connect_mcp client tool parks the turn and renders a PREFILLED AddCustomServerForm in the dock; the user completes the OAuth/api-key connect themselves (auth never reaches the agent). On success the new mcp_store connection is attached to the target agent's draft spec (mcps[].connection) and the session is woken with { connected, connection_id, mcp_id }. Mirrors the set_secret punch-out. - agentBuilderStore: pendingMcpConnect state + setter - useAgentBuilderClientTools: connect_mcp handler (parks + defers) - AgentBuilderDock: connect form rendering + connect -> attach-to-spec -> resolve - AddCustomServerForm: initialValues prop for prefill - useMcpConnect: connectCustomAsync + refetchInstallations Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… legacy integrations - buildTree renders the authorable folder sections (triggers, secrets, skills, tools, mcps, identities) even when empty, so the add/connect affordance is reachable on a fresh agent (you couldn't add an MCP to an agent that had none). - Remove the deprecated integrations UI from the config pane (tree section, overview/body, routing, description, the legacy mcps[].auth.integration row, now-unused LinkIcon). Removing integrations from the agent-platform spec + runtime is a follow-up. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The agent-builder dock now sends the kind:'client' tool ids it can execute (AGENT_BUILDER_CLIENT_TOOLS — set_secret, connect_mcp, focus_*, toast, get_context) in the /run body so the runner knows it can punch out the interactive connect_mcp form here instead of relaying a URL. Threaded AGENT_BUILDER_CLIENT_TOOLS → useAgentChat (supportedClientTools option) → AgentChatSession → agentChatService → runAgentSession (api-client) → POST /run body. Replaces relying on the never-sent x-posthog-client header. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The inline punch-out crammed the full AddCustomServerForm into the dock's narrow above-composer strip — ugly. Move connect_mcp to a proper Dialog (AgentBuilderMcpConnectDialog); the set_secret one-line punch-out stays inline. Adds a hideHeader prop to AddCustomServerForm so the dialog owns the title/close chrome (no in-form Back button / duplicate heading). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Producer counterpart to the agent-platform server change that gates client-tool exposure on a per-run capability list. The agent-builder dock now declares the `kind:'client'` tool ids it can fulfil and sends them in the /run body as `supported_client_tools`, so the runner exposes only those to the model (spec ∩ supported). - `runAgentSession` (api-client): optional `supportedClientTools`, added to the /run body when non-empty. - `AgentChatSession` (core): carries `supportedClientTools`; `agentChatService` forwards it to `runAgentSession`. - `useAgentChat` (ui): new `supportedClientTools` option threaded into the session config. - `AGENT_BUILDER_CLIENT_TOOLS` declares the dock's fulfillable ids (set_secret, focus_*, toast, get_context) — verified to exactly match the agent-builder spec's `kind:'client'` tools — and is passed from the dock. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The agent-config MCP sections expanded the custom-server form inline under a Connect-new/Cancel toggle. Pop it out in a dialog instead (matching the agent builder's connect_mcp punch-out). Extract AddCustomServerDialog as the shared modal chrome and refactor AgentBuilderMcpConnectDialog to delegate to it, so both the manual UI and the punch-out share one wrapper. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…' into ben/agent-mcp-connections # Conflicts: # packages/api-client/src/posthog-client.ts # packages/core/src/agent-chat/identifiers.ts # packages/ui/src/features/agent-applications/agent-builder/useAgentBuilderClientTools.ts # packages/ui/src/features/agent-applications/hooks/useAgentChat.ts

The MCP config could add (Connect new / Add from a connection) and edit the connection, but had no way to drop an mcps[] entry — only clear its connection. Add a Remove server action to the MCP detail that filters the entry out of the spec and returns to the list. The shared mcp_store installation is untouched; only the agent's reference goes away. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

dmarticus

Just one question on an unstated removal — no blockers on the draft itself, the gaps you list cover it.

…ions # Conflicts: # packages/ui/src/features/agent-applications/agent-builder/useAgentBuilderClientTools.ts # packages/ui/src/features/agent-applications/hooks/useAgentChat.ts

greptile-apps · 2026-06-26T08:49:24Z

Comments Outside Diff (1)

packages/ui/src/features/agent-applications/agent-builder/AgentBuilderDock.tsx, line 133-165 (link)

connect_mcp punch-out skips draft-branching

submitMcpConnect calls client.updateAgentRevisionSpec directly with pending.revisionId, but the server only accepts a PATCH on a revision that is already in draft state. pending.revisionId falls back to pageRevision (the currently-viewed revision in the agent-config pane), which can be live, ready, or archived. When it is, the server rejects the PATCH with a 4xx and the operation fails — the agent gets the error token and the user sees nothing useful.

Every other edit path in this PR goes through useApplyAgentSpec, which branches to a new draft when the source isn't one. submitMcpConnect needs the same treatment: check rev.state, and if it's not "draft", call client.createAgentDraftRevisionFrom first before patching.

_{Reviews (2): Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." | Re-trigger Greptile}

…ntSpec The UI removal of the integrations section (IntegrationsOverview/IntegrationBody, cfg:integrations node, SECTION_INFO entry) left `AgentSpec.integrations?: string[]` behind — a soft regression where a shipped spec's integrations would be silently un-editable. Nothing in the app reads the field (the agent-platform schema already dropped it), so remove it for a total removal. Per review: this is a deliberate "nobody uses integrations" call, not collateral from the MCP refactor. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ions # Conflicts: # packages/ui/src/features/agent-applications/components/AgentConfigurationPane.tsx # packages/ui/src/features/agent-applications/hooks/agentApplicationsKeys.ts # packages/ui/src/features/agent-applications/hooks/useApplyAgentSpec.test.ts # packages/ui/src/features/agent-applications/hooks/useApplyAgentSpec.ts

…+ per-tool overrides) The connection MCP detail now shows the server's live tool catalog (useMcpInstallationTools) with a connection-wide default permission and per-tool overrides, reusing the standalone manager's ToolRow / ToolPolicyToggle. Levels map allow/approve/deny <-> approved/needs_approval/do_not_use at the boundary; edits persist to spec.mcps[].default_tool_approval + tools[].level via the existing draft-branch-then-PATCH apply(). Add-from-connection now stamps a safe default_tool_approval: 'approve' so the runtime and UI agree from first save. Also clarifies the agent-level (shared connection — owner connects once, every asker reuses) vs principal-level (per-asker identity provider) distinction in the connection description. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…issions Extract the searchable/expandable tool list (counts, per-tool policy toggle, expandable descriptions) into a presentational ToolPermissionList beside the mcp-servers parts, and wire McpBody to it. The parent owns config + persistence (spec default_tool_approval + per-tool level overrides); the component is callback-driven so it can also back the global MCP-servers manager unchanged. ServerDetailView is left as-is. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ions

… on MCP detail Replace the implicit "set a connection vs wire auth.provider" model with an explicit credential-model dropdown reading/writing the new required spec.mcps[].kind: - Agent-level → connection selector + per-tool permissions. - Principal-level → identity-provider selector (+ jump, or a prompt to declare one). Switching modes clears the opposite credential and stamps `kind`. - Tidy the layout: the mode dropdown and "Remove server" share one compact header row at the top with a single contextual helper line, instead of a verbose blurb up top and a remove button buried at the bottom. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Make the principal-level identity UI nicer: the provider dropdown shows a fingerprint icon per option, and the selected identity renders as a JumpRow preview card (icon + provider + "per-asker linked identity" status, amber "undeclared" badge when auth.provider isn't in identity_providers[]) that jumps to the identity detail — replacing the plain text link. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

main dropped the `sonner` dependency (toast moved to @posthog/ui/primitives/ toast, #2965), but these two files still imported from it — passing local typecheck (sonner still in node_modules) yet breaking CI's clean build/ typecheck ("Rollup failed to resolve import 'sonner'"). Point both at the ui primitive like the rest of the app. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The runtime no longer enforces the shared connection owner's per-tool approval marks — the agent's own allow/approve/deny config is authoritative. Surface that in the MCP "Tool permissions" section so authors know the owner's marks are informational here. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

greptile-apps Bot reviewed Jun 25, 2026

View reviewed changes

Comment thread packages/ui/src/features/agent-applications/hooks/useApplyAgentSpec.ts

Comment thread packages/ui/src/features/agent-applications/components/AgentConfigurationPane.tsx

benjackwhite and others added 8 commits June 25, 2026 14:16

benjackwhite changed the base branch from main to ben/agent-supported-client-tools June 25, 2026 18:38

benjackwhite mentioned this pull request Jun 25, 2026

feat(agent-platform): agent-level MCP connections, per-agent tool permissions, and capability-based client tools PostHog/posthog#66007

Merged

dmarticus reviewed Jun 25, 2026

View reviewed changes

Comment thread packages/ui/src/features/agent-applications/components/AgentConfigurationPane.tsx

Base automatically changed from ben/agent-supported-client-tools to main June 26, 2026 08:17

Merge remote-tracking branch 'origin/main' into ben/agent-mcp-connect…

a8e8ba6

…ions # Conflicts: # packages/ui/src/features/agent-applications/agent-builder/useAgentBuilderClientTools.ts # packages/ui/src/features/agent-applications/hooks/useAgentChat.ts

benjackwhite marked this pull request as ready for review June 26, 2026 08:39

benjackwhite and others added 8 commits June 26, 2026 10:55

Merge remote-tracking branch 'origin/main' into ben/agent-mcp-connect…

16af469

…ions

dmarticus approved these changes Jun 29, 2026

View reviewed changes

benjackwhite mentioned this pull request Jun 30, 2026

feat(agent-applications): auth & identity section with admission editing #3021

Draft

dmarticus merged commit 6922518 into main Jun 30, 2026
23 checks passed

dmarticus deleted the ben/agent-mcp-connections branch June 30, 2026 22:27

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(agent-applications): manage MCP connections in the agent builder#2929

feat(agent-applications): manage MCP connections in the agent builder#2929
dmarticus merged 20 commits into
mainfrom
ben/agent-mcp-connections

benjackwhite commented Jun 25, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jun 25, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Jun 25, 2026

Uh oh!

Uh oh!

Uh oh!

dmarticus left a comment

Uh oh!

Uh oh!

greptile-apps Bot commented Jun 26, 2026 •

edited

Loading

Comments Outside Diff (1)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

benjackwhite commented Jun 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What you can do

Changes

Validation

Notes / follow-ups

⚠️ Outstanding / known gaps before this is production-good

Agent context

Uh oh!

github-actions Bot commented Jun 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

greptile-apps Bot commented Jun 25, 2026

Uh oh!

Uh oh!

Uh oh!

dmarticus left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

greptile-apps Bot commented Jun 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Comments Outside Diff (1)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

benjackwhite commented Jun 25, 2026 •

edited

Loading

github-actions Bot commented Jun 25, 2026 •

edited

Loading

greptile-apps Bot commented Jun 26, 2026 •

edited

Loading