Pod

Available as Markdown and JSON. Pod is also available over MCP.

Reported issues for Agent Swarm

Pod holds 6 of 6 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to Agent Swarm.

Most discussed

feat: support LiteLLM gateways and Microsoft Foundry for the claude harness

Intent

Self-hosted operators who cannot use a direct Anthropic key cannot run the claude harness cleanly. Many teams must send model traffic through a gateway (LiteLLM or another Anthropic-compatible proxy) or through their cloud provider (Microsoft Foundry) for key management, cost tracking, or data-residency reasons. Claude Code already supports both. The swarm blocks them.

What I hit while running the Full Swarm compose against LiteLLM:

Read the thread · 2026-10-01 · closed · external user · 9 comments

feat(memory): Azure / Microsoft Foundry embeddings in setup (404 on Custom endpoint)

Intent

Operators on Microsoft Foundry / Azure OpenAI cannot finish the memory step of /setup. The step offers presets for OpenAI, OpenRouter, and Vercel AI Gateway, plus "Custom". A Foundry resource URL in "Custom" fails with a message that does not say what is wrong:

Embeddings
The model was not found at this endpoint.
The endpoint could not find the embedding model (404).

Setup: Custom, model text-embedding-3-small, base URL https://<resource>.services.ai.azure.com, with…

Read the thread · 2026-10-02 · closed · external user · 5 comments

ACP credential check always reports ready:true, even with zero target configured

What's wrong

src/commands/provider-credentials.ts has a constant, always-ready checker for the acp provider:

// src/commands/provider-credentials.ts:106-107
// The ACP target process owns its own auth, so the swarm requires nothing.
acp: [],
// src/commands/provider-credentials.ts:126
acp: () => ({ ready: true, missing: [], satisfiedBy: "sdk-delegated" }),

Every other provider's entry in CREDENTIAL_PROVIDER_CHECKERS inspects real environment state…

Read the thread · 2026-09-14 · open · outside contributor · 4 comments

ACP provider hands the full swarm API key to an arbitrary target process (scope it to an ephemeral token)

The ACP adapter passes the worker's full swarm API key to an arbitrary target process.

In src/providers/acp-adapter.ts, createSession configures the swarm MCP server for the ACP agent with:

headers: [{ name: "Authorization", value: `Bearer ${config.apiKey}` }],

The ACP target is an arbitrary executable chosen by ACP_TARGET_COMMAND. It can read, retain, or forward that bearer, which is the full-privilege swarm API key — not a scoped one. This also contradicts the provider's…

Read the thread · 2026-09-03 · closed · outside contributor · 4 comments

Extension install: MCP tool denies workers while REST with the bare shared API key accepts

Found during the swarm extensions v1 live QA (PR #1441, report thoughts/taras/qa/2026-09-15-swarm-extensions.md).

Observed

Read the thread · 2026-09-15 · closed · outside contributor · 3 comments

ACP custom-target docs examples (Gemini CLI, claude-code-acp) omit ACP_TARGET_ENV_KEYS, so the credential the doc tells you to set never reaches the spawned process

What's wrong

The custom ACP target only forwards environment variables that are explicitly allowlisted via ACP_TARGET_ENV_KEYS (plus a fixed, credential-free base set):

// src/providers/acp-targets.ts:138-152
const customTargetProfile: AcpTargetProfile = {
  target: "custom",
  command(config) { ... },
  env(config) {
    const env = baseTargetEnv(config); // PATH, HOME, TMPDIR, TMP, TEMP, SHELL, USER, LOGNAME, LANG, LC_ALL, BUN_INSTALL, NODE_PATH — no credentials…

[Read the thread](https://github.com/desplega-ai/agent-swarm/issues/1560) · 2026-09-21 · closed · outside contributor · 1 comment

## Most recent

The remaining reports are on [the project's issue tracker](https://github.com/desplega-ai/agent-swarm/issues).